kensaurus/kenji
You say the job; the playbook runs. Agent skills that ask first, test first, and prove the result: build, fix, audit, test, ship. Tuned for React, Next.js, and Supabase; most of it works on any stack.
Changelog
All notable additions and changes to kenji (kensaurus/skills, named cursor-kenji before 2.0.0) are listed here.
[Unreleased]
Changed
- The animations now replay real output. The three mock session
animations are replaced by captures of actual command output (the
completion gate,
/kenji,--list-profiles), with GIF copies for sites that do not take SVG.
Added
- A diagram of the completion gate, in English and Japanese.
- An article on the gate, in English and Japanese
(
docs/articles/completion-gate.md,completion-gate.ja.md).
[2.8.1] — 2026-11-17
Fixes found after 2.8.0, and what the Agent Plugins marketplaces lint for.
Fixed
- A core-only install can finish every core workflow.
workflow-quality-gatechains totest-red-team,audit-bundle-sizeandaudit-performance, which were outsidecore. They are now in it (31 skills, a listing of about 7,200 characters), andvalidate-skillsfails when a core skill chains outside core. --prunealso removes orphaned commands: this pack's command files that only name skills outside the profile. A command you edited is left alone; removed ones are copied to.kenji-pruned/first.- All 162 skills pass
vally lint, the check GitHub's Copilot plugin marketplace runs. TwoSKILL.mdfiles were over 500 lines (debug-sentry-monitor,mobile-emulator-test; one section of each moved toreferences/), and three linked outside their own folder (enhance-skill-prompts,meta-skill-creator,mushi-integration), which breaks when a skill is installed on its own.validate-skillsnow enforces both rules. evals/README.mdnamed the wrong results folder.
Added
- docs/LAUNCH-KIT.md: measured reach, the rules of each announcement channel, the objections to answer, and an order of work.
- A VS Code and Copilot install link in the README.
check-docs-factschecks the README profile table againstprofiles.json.
[2.8.0] — 2026-11-17
One command to remember, a smaller install when you want one, and skill text that no longer assumes a model (ADR-0016, ADR-0017).
Added
/kenji, a router skill. Type/kenjiand the job in your own words. It names the one playbook that fits, says what you get, and runs it; it asks at most one question. It is user-invoked, so it adds nothing to the skill listing. Its full index (skills/kenji/references/index.md) is generated bynpm run gen:skill-indexand checked innpm test.- Install profiles.
--profile coreinstalls the 28 skills of the daily loop (a listing of about 6,500 characters against about 39,000 for the full pack).web,backend,mobile,growthandmaintainadd to it; combine with commas.--pruneremoves this pack's skills outside the profile by folder name, after copying each to.kenji-pruned/; folders the pack does not ship are never touched.--list-profilesprints them. The lists live inprofiles.json, andvalidate-skillsfails when a skill is in no profile.allstays the default. - Antigravity CLI target.
--antigravitywrites the skills to~/.gemini/config/skills;--autofinds it when~/.gemini/configexists. It is not part of--all. - Agent Plugins and Gemini extension manifests. Root
plugin.json(Agent Plugins 1.0.0, read by Codex, Copilot, Kiro and VS Code) andgemini-extension.json. Both carry the skills only. - Trigger evals.
evals/holds twelve cases forclaude plugin eval, one per core route, each passing when the right skill loads. They spend real tokens, so they are a maintainer step, not part ofnpm test. - Animated walkthroughs. Three animated SVG sessions in the README, the
getting-started guide and the site, written by
npm run brand:walkthroughwith no browser or dependency. Viewers who ask for reduced motion get the final frame. - GitHub Pages workflow for
site/.
Changed
- Skill text names no model. Five skill files, the manifests and
llms.txtno longer say Opus 5.5.docs/MODEL-AND-EFFORT.mdis now a per-model guide (Fable 5.2, Opus 5.5, Sonnet 5.5, Haiku 5.5) with the token budget and how to measure it. - The status line shows the effort level next to the model.
- The always-on verification rule tells an ordinary obstacle (work through it) from a deliberate blocker (leave it and report), and states what a subagent costs.
- Codex CLI no longer gets prompt files by default. Codex removed custom
prompts in 0.117.0.
--codex-promptswrites them for older releases; existing files are left alone. - A profile install leaves
skillListingBudgetFractionunset, because the smaller listing fits Claude Code's default budget. - README and getting-started guide open with a three-step quick start,
/kenji, and an animated session; the guide no longer assumes Cursor. check-docs-factsfails when any manifest version differs frompackage.json.- Listing docs corrected: the old npm package is fully deprecated (all 56
versions); the Claude plugin directory is submitted at
claude.ai/directory/manage; the VoltAgent rename PR merged. - More from KENSAURUS (README) now follows
https://kensaur.us/apps.json: current names and taglines, Skin Analysis and Hear, Hear added, one deep link per app, campaign-tagged hub link.llms.txtgains a "More from kensaurus" section, and the site page gets a "Made by kensaurus" credit, three related apps, and a JSON-LD author that points at the sharedkensaur.us/#orgnode.
[2.7.0] — 2026-10-19
Claude Code sessions tell you when they need you, every skill keeps its description in the listing, and the skills read the way the 2026-10 authoring guidance asks (ADR-0015).
Added
- Skill writing pass against Anthropic's skill-authoring guidance (read
2026-10-19; the guidance itself is linked from ADR-0015, no dates in skill
bodies):
- 71 skills gain a "This skill vs neighbors" table, only where routing
overlaps (the description names another skill, or another skill's
description names it). Each "Owns" cell is taken from that neighbor's own
description. 22 isolated skills stay as they are.
enhance-capacitor-ui's existing table gets the heading and its missing inbound neighbor. - Every
workflow-*skill carries a copyable checklist (the 11 that lacked one: environment-ready, git-commit, grilling, gtm, housekeep, launch-ready, merge-conflicts, mobile-native-uiux, onboard, parallel-agents, quality-gate). Steps index the skill's own phases in its own words; a gate is its own step. - Trigger clauses for the four descriptions that had none (
handoff,meta-skill-creator,mobile-rn-performance,protocol-browser-anti-stall), paid for inside the 39,000-char listing ratchet by trimmingdesign-apiandworkflow-ship-and-observeto four triggers each. Listing: 38,994 chars. argument-hinton the two skills that read an argument:researchandaudit-responsive.scripts/check-reference-toc.mjs, innpm test: a reference file over 100 lines opens with a## Contentslist; references stay one level deep (a sibling file a reference links to is also linked fromSKILL.md); every relative Markdown link resolves. Links inside fenced examples are ignored.npm run fix:reference-tocwrites a missing Contents list from the H2 headings. First run: Contents lists added to six files, three broken links inenhance-web-ux/references/details.mdrepaired (companion files are one directory up), andenhance-web-ux's three companion files linked fromSKILL.md.- Playbook techniques T9 (progressive-disclosure hygiene) and T10
(copyable checklist), and the explicit decisions not to adopt
when_to_use(same listing budget, ratchet at ceiling) or dated reads in skill bodies (time-sensitive text).
- 71 skills gain a "This skill vs neighbors" table, only where routing
overlaps (the description names another skill, or another skill's
description names it). Each "Owns" cell is taken from that neighbor's own
description. 22 isolated skills stay as they are.
- Claude Code gets a desktop Notification hook, a status line, and a
skill-listing budget.
--claude(and--all,--auto) now writes three scripts to~/.claude/kenji-hooks/and merges them into~/.claude/settings.jsonin one pass with the Stop hook, so the file is written once and backed up at most once per run:notify.mjsis registered as aNotificationhook forpermission_prompt,idle_prompt,agent_needs_inputandagent_completed, so a session that needs you raises a desktop alert (notify-sendon Linux,osascripton macOS, a WinRT toast on Windows). It spawns no shell and always exits 0; Claude Code ignores its exit code anyway (hooks reference, read 2026-10-19).statusline.mjsbecomesstatusLineonly when none is set (or the one set is ours). It shows model, folder, branch, context used, session cost, five-hour usage and session name, and spawns nothing: the branch comes from.git/HEAD, through the gitdir file in a worktree (status line docs, read 2026-10-19).--only statuslineinstalls just this; leavingstatuslineout of--onlyskips it.skillListingBudgetFractionis set to0.03only when unset. Claude Code shows the skill listing against 1% of the context window by default and drops descriptions from the least-used skills when over budget, which made most of this pack invisible to the model on a full install (skills docs, read 2026-10-19). The installer now prints the listing size against the budget in force.- The plugin manifest
hooks/hooks.jsonregisters the same Notification hook forclaude plugin installusers.
scripts/test-install.mjscovers each piece: registered once, idempotent on a re-run, a user's ownstatusLine, budget and Notification hooks survive,--only skills,commands,agents,ruleswrites no hooks or status line, andsettings.jsonhas at most one backup per run.
Changed
- Claude Code no longer receives a
commands/*.mdwhose name is also a skill. Claude Code gives the skill precedence over a same-name command file (skills docs, read 2026-10-19), so the command copy was unreachable and only padded the skill listing. The nine affected names (burndown-full,complete-everything,handoff,housekeep-backlog,housekeep-files,housekeep-gates,research,test-mutation,thirdparty-web-interface-guidelines) are skipped on install,--verifyexpects them absent, and an unmodified older copy is pruned on the next run. Cursor still gets both. A command file you edited yourself is left alone.
[2.6.0] — 2026-10-11
Codex CLI and Gemini CLI get the skills, and the README works for any stack and any agent.
Added
--codexand--geminiinstall the skills. Both tools now readSKILL.mdskills, and both document~/.agents/skillsas a user skills folder: Codex's only user scope (docs), and the Gemini alias that takes precedence over~/.gemini/skills(docs). Both pages were read 2026-10-11. The installer writes one shared copy of this pack's skills there, once per run. When Cursor is also a target, the Cursor mirror's copy is reused. That mirror copies all of~/.cursor/skills, your own skills included.--dry-runlists each skill,--verifyhash-checks them,--skill <name>installs one, and--only rulesskips them.--no-agents-mirrornow means nothing goes to~/.agents/skills, with a warning that Codex and Gemini get no skills.AGENTS.md,GEMINI.md, the portable commands, and their.bak-<stamp>backups are unchanged. The skill routing index is still left out of the merged file.scripts/test-install.mjscovers the change:- the skill count in
~/.agents/skills; - no
~/.gemini/skills; - no Cursor builtins;
- an idempotent re-run with no backups;
--verifypasses, then fails on a tampered file;--dry-runwrites nothing;--only rules,--skill, an unknown--skill(exit 1), and--no-agents-mirror;- a foreign skill survives a re-run and a Cursor
--clean.
- the skill count in
- Workflow chains in frontmatter (ADR-0014).
Mushi Mushi runs a workflow skill as a pipeline of the skills it calls, but
it found them by scanning the body for
skills/<slug>/SKILL.mdpaths, which the workflows no longer write, so every workflow ran as one step. Each workflow that runs other skills now declares them asmetadata.chain: "a b c", in run order, inside the Agent Skills spec (string values only). Optional steps, routers and "see also" mentions stay out.validate-skillsfails on an unknown or repeated slug.workflow-build-feature:workflow-spec-tdd test-unit test-playwright workflow-prworkflow-fix-and-ship:debug-error test-playwright workflow-prworkflow-quality-gate:test-red-team audit-security audit-bundle-size audit-performance test-unitworkflow-launch-ready:enhance-web-seo enhance-pwa audit-bundle-size audit-i18n workflow-quality-gate deploy-verify iterate-post-launchworkflow-gtm: the seven-step sequence,plan-gtmthroughiterate-gtm-weeklyworkflow-release-prep:audit-code-review workflow-git-commit workflow-pr
workflow-mobile-native-uiux: the end-to-end native UI/UX run on an existing Capacitor, Expo, or React Native app. Baseline screenshots and scorecard, thenenhance-capacitor-ui(hybrid only) →mobile-capacitor-platform(Capacitor only) →enhance-mobile-native-feel→audit-accessibility→audit-responsive→mobile-emulator-test→workflow-pr, on one branch, ending in one PR that says whether a store build is needed.
Changed
-
README rewritten for a coder on any stack and any agent. It now has:
- a three-command quickstart: dry run, install, verify;
- a support matrix for Claude Code, Cursor, Codex CLI, Gemini CLI, and any
SKILL.mdagent via the skills CLI, with a cited doc for every path; - a table of every path the installer writes and what happens when a file is already there;
- what the Stop hook does and how to skip or remove it;
- which rules assume React, Next.js, or Supabase, and how to leave them out;
- verify, update, rollback, and manual uninstall steps;
- troubleshooting for Windows locks and CRLF, existing files, Codex's skill
listing budget, Codex's removed custom prompts,
CODEX_HOME, and Gemini's consent prompt and Antigravity CLI notice.
Removed:
- the skills.sh install-count badge;
- unverified model and version claims;
- every unguarded count. The counts left are the phrases
check-skill-countrewrites.
-
Re-rendered the hero, OG, architecture, and pieces images. They said 160 skills, and the architecture image showed Codex and Gemini without skills.
-
Installer output:
- the
--verifyhint names the targets you installed; - the skill/command name-clash note prints only for Cursor and Claude Code;
- Gemini's hint uses
/memory reloadand/skills reload, the commands its docs name.
- the
-
GETTING-STARTED, DISTRIBUTION, and MODEL-AND-EFFORT describe the Codex and Gemini install as it is now.
-
Mobile skills checked against 2026 primary sources (Apple HIG, Android 15 and 16 behavior changes, Google Play target-SDK policy, Capacitor 8 docs and changelog, WCAG 2.2, MDN compat data, Expo SDK 54–58 changelogs; read 2026-10-08). Each skill now owns one job and points to the others:
mobile-capacitor-platformowns the shell plumbing: a Native shell plumbing table (SystemBars and safe-area CSS, deadStatusBarconfig, keyboard resize, splash, Android back, haptics defaults, large screens), each row with a device probe and a done line; Capacitor 8, targetSdk 36, and Xcode 26 floors with their dates.enhance-mobile-native-feel: Android back is part of the chrome step and a scorecard gate; Expo edge-to-edge is mandatory from SDK 54 (was "SDK 53+ default"); Capacitor'simpact()defaults toHeavy; Liquid Glass stays in the control layer; the Capacitor edge-to-edge block moved to its owner.design-mobile-first: the WCAG 2.2 rules it lacked — 2.5.8 target-size floor (24 px) under the 44 px goal, 2.5.7 tap alternatives for every drag, 2.4.11scroll-paddingunder fixed bars;dvhsheets with a close button; a probe table replaces the one-line validation.audit-accessibilityclaimed WCAG 2.2 AA but checked none of the AA criteria new in 2.2. It now probes 2.4.11, 2.5.7, 2.5.8, and 3.3.8 and reports them.plan-mobile-readiness: a thin web-view app is a Guideline 4.2 rejection and app-generator output is 4.2.6, not 2.5.2 (downloaded code, now inplan-capacitor-hardening's OTA pillar). The "25%" figure is now cited (Apple Newsroom, May 2025: over 1.9 M of 7.7 M rejected in 2024). Adds targetSdk 36 (Play, 2026-08-31), the Xcode 26 upload floor (2026-04-28), and age-rating answers.- "Capacitor 8.3.2+" was replaced everywhere: SystemBars exists since 8.0.0 and its inset fixes continued through 8.5.2, so the skills now say "latest 8.x".
audit-responsiveadds phone-landscape and tablet cells for native apps (Android 16 ignores orientation locks at ≥ 600dp);mobile-emulator-testgains adb probes for back, font scale, reduce motion, large screen, and nav mode.
Fixed
--cleanno longer wipes all of~/.agents/skills. The Cursor mirror deleted the whole folder, with no backup, before copying. That folder is shared with Codex, Gemini, and other installers such asnpx skills add -g. A--cleannow removes only the skill folders this pack ships.--codexand--geministamped "this tool has no skills loader" into the generatedAGENTS.mdandGEMINI.md. The header now points at~/.agents/skills. Existing installs get one.bak-<stamp>copy on their next run.--skill <name> --codexprinted "no skills loader" and did nothing. It now installs that skill, and an unknown name exits 1, as it does for Cursor and Claude Code.scripts/shorten-skill-descriptions.mjsclipped mid-quote. A clip inside a quoted trigger or a parenthetical left it open ("tidy.), and trailing/or→survived (Motion /.). It now drops the dangling opener and those separators;--self-testruns invalidate.yml.plan-rls-audit's description again carries theservice_role keyand security-advisor triggers its body and the cheatsheet promise, at no listing cost.- Review copy fixes:
.env.examplelost its orphaned fal.ai note, the DISTRIBUTION clone one-liner is copy-pasteable,mcp/README.mdnamesSUPABASE_PROJECT_REF, the trigger cheatsheet lists each skill once, andllms.txtcallsplan-*planning skills, not audit skills.
[2.5.0] — 2026-10-06
A lighter package and a README that shows more than it tells.
Fixed
- The npm tarball was 9.8 MB (12.2 MB unpacked, 484 files). Ten fal.ai
logo experiments under
assets/candidates/were tracked and shipped, along with every README screenshot. npm's package page loads README images from GitHub, not from the tarball, so none of it was needed.filesnow ships the pack,assets/logo.pngandassets/logo.svgonly: 1.2 MB packed, 3.4 MB unpacked, 459 files. Supply-chain scanners score what ships; this is the part of the Socket.dev score that was ours to fix. The rest of that score is download count and "recently published", which only time changes.
Added
- Architecture diagram under the install command: what gets installed
(skills, commands, subagents, rules, MCP templates), through
npx @kensaurus/skills --all, into~/.cursor,~/.claude,~/.codex, and~/.gemini. States the installer's footprint: no install scripts, no network, no shell. - Pieces strip replacing the what's-inside table: seven tiles with pixel glyphs and counts.
- Section icons: twelve 8×8 pixel glyphs rendered from
showcase.html(?view=icon&name=…) and placed in the README headings. - More from KENSAURUS rebuilt as cards with the real store listing names and links: glot.it (App Store: glot.it; Google Play: glot.it – Learn Thai), yen-yen – Expense Tracker, the wanting mind – Living Book, Help Her Take Photo (Google Play: help her take photo – Pose Cam), lets-talk – Conversation Coach, Solo Boss, Tsumagoi, Mushi Mushi, plus the Google Play developer page. Names and links come from the store listings and the portfolio data file, not from memory.
Changed
npm run brand:rendercovers the new views;docs/screenshots/README.mdlists them.
[2.4.0] — 2026-10-06
The README reads as a page again, and the biggest skills load in a fraction of the tokens.
Changed
- README rebuilt around two images. The hero, then a new ladder diagram, "How one request lifts the repo" (Orient, Assess, Change, Prove, Ship, each rung naming what the repo gains), then install, the "What should I say" table, and the tour. Everything after that collapses: install options, commands, subagents, MCP servers, rules, shell helpers, upgrading, FAQ. Visible length drops from 823 lines to about 150; the source is 420 lines.
- Skill index moved to
docs/SKILLS.md.generate-skill-index.mjsnow writes the full one-line-per-skill list there and only the family counts table into the README.--checkguards both files. The 314-line table that made the README a wall is gone from it. - 22 skills (the 8 largest by prose and the 14 heaviest in inline code) slimmed into reference files, following Anthropic's
authoring guide: SKILL.md keeps the reasoning skeleton, one short example per
concept, and the checks; long code, pattern catalogues, probes, and templates
move to
references/<topic>.mdwith a contents list. Frontmatter and rule meaning are unchanged. Their SKILL.md bodies drop 43% in bytes (364 KB to 206 KB, about 49,000 fewer tokens when all are loaded); the largest, enhance-web-landing, goes from 42 KB to 9.5 KB. Twenty-eight older reference files over 100 lines gain a contents list. - Plan skills share one execution line. Ten
plan-*skills carried the same two-sentence execution note; it is now one line each. /thirdparty-web-interface-guidelinesdelegates to its skill instead of carrying a 185-line copy of the rule set.docs/PLAN-LOOPS.mdlinks to the new README section.
Added
docs/screenshots/ladder-{dark,light}.pngrendered from theladderview indocs/screenshots/src/showcase.html;npm run brand:rendercovers it.
[2.3.0] — 2026-10-06
A new brand and a plain-language pass, plus the two methods behind them added to the skills so other repos can do the same.
Added
- Brand built as code.
docs/screenshots/src/showcase.htmlrenders every shipped image from one source: the README hero (dark and light), the four tour cards, the OG image, the logo, and the favicon. The mark is a hand-drawn pixel kensaurus (one character per pixel), set in Bricolage Grotesque, Press Start 2P, and JetBrains Mono, in lime, yellow, and coral on ink or paper, with 3px borders and hard offset shadows. No gradients, glass, or generated art.node scripts/render-brand-assets.mjs(alsonpm run brand:render) renders all eleven files with Playwright and copies the site set; counts come frompackage.json. enhance-readmebrand hero kit.references/brand-hero-kit.mdexplains the system, the npm and GitHub rendering limits (npm shows the hero at half size and strips<picture>), the build steps, and an anti-slop check.scripts/render-views.mjsrenders any HTML view to PNG.docs-writerplain-language reference.references/plain-language-ste.mdcarries the ASD-STE100 rules that transfer to developer docs (20 and 25-word limits, one instruction per sentence, active voice, one meaning per word), an AI-tell list, rules for alt text, captions, and package descriptions, and a Vale prose-lint recipe.docs-writer,enhance-readability, andenhance-readmepoint to it.researchgrounding contract. Step 5b: cite or abstain, date the query, pin the version, quote the primary source, separate observation from inference, measure numbers.
Changed
- README, npm, and marketplace copy in plain language. Hero alt text and
caption, the four tour captions, the npm description template, both plugin
manifests, the marketplace strings,
llms.txt, andsite/index.htmlnow use short active sentences. The npm description leads with the inventory and the behavior. Keywords grow from 15 to 24 search terms. site/index.htmluses the new system and reads its counts from the same template as the manifests (check-skill-countnow checks it).docs/screenshots/README.mddocuments the render pipeline.
Removed
scripts/render-brand-assets.py(Pillow geometric mark),scripts/gen-brand-assets.mjsandscripts/gen-brand-candidates.mjs(fal.ai experiments). The shipped brand no longer depends on an image API;FAL_KEYleaves.env.exampleandSECURITY.md.
[2.2.0] — 2026-10-06
Adds the app-wide "make this mobile app feel native" pass the pack lacked, and refreshes the mobile and UX skills against 2026 platform facts: iOS 26 Liquid Glass tab bars, Android 16 enforced edge-to-edge, Expo SDK 54–58 native tabs, Reanimated 4, FlashList v2, Capacitor 8.3 system bars. Every fact was pulled from the vendor docs or maintainer posts dated 2025–2026 and is cited in the new skill's references.
Added
enhance-mobile-native-feelskill: one coherent pass over an existing Expo / React Native or Capacitor app that reads as a website: system tab bar and sheets, edge-to-edge insets, virtualized lists, one spring vocabulary on the UI thread, haptics by meaning, platform symbols, Dynamic Type, two accent roles. Scores ten native-feel checks before and after (references/native-feel-scorecard.md), names the web-shaped smells it removes (references/anti-patterns.md), and carries version-gated recipes for both stacks (references/stack-recipes.md). Routes per-screen polish tomobile-rn-screenand surface splitting toenhance-capacitor-ui.
Changed
- Mobile and UX skills refreshed for 2026.
design-mobile-firstgains a native-shell-first section (NativeTabs, Android 16 edge-to-edge, Capacitor 8.3 safe-area variables, sheets, press feedback) and moves the web drawer example toreferences/web-nav-drawer.md.mobile-rn-screenadds the native-feedback rule and pain classes S19–S20.mobile-rn-performancestates the FlashList v2 (New Architecture, noestimatedItemSize) and LegendList rules and the Reanimated 4react-native-workletssplit.enhance-capacitor-uiandplan-capacitor-hardeningrecord thatStatusBar.setBackgroundColor/overlaysWebView: falseare no-ops on Android 16 and that Capacitor 8.3.2+ handles insets in core.plan-mobile-readinessadds pillar F (native feel as 2.5.2 evidence).design-motionandenhance-motionadd the spring vocabulary (Material 3 spatial vs effects) and hand native apps to the new skill.audit-ux-journeysadds IA10–IA12 (tab destinations, one question per screen, sheet vs page) with methods in its checklist;audit-uxandmobile-emulator-testscore the same signals. - Descriptions trimmed on the twelve skills above (418 chars reclaimed) so the new skill fits under the 39,000-char listing ratchet (ADR-0010); the ratchet did not move.
Fixed
- Phantom modified files on every checkout. Fourteen tracked files were
committed with CRLF while
.gitattributesdeclareseol=lf, sogit statuslisted them as modified with an empty diff. They are renormalized; content is unchanged. enhance-web-web3dnow links itsreferences/css-canvas-effects.md, which was shipped but unreachable from the skill.
[2.1.1] — 2026-10-01
Fixed
- Installs on Windows survive a briefly held file. An open editor
watching
~/.cursoror~/.claude, or an antivirus scan, can hold a file for a moment right after it is written. Windows then failed the next copy withEPERM, which aborted the install and left the target half-copied: Cursor failed three runs in a row on a differentSKILL.mdeach time. The installer now retries a copy, write or delete onEPERM,EBUSYorEACCESfor up to about 3 seconds per file. If the lock never clears it still fails, and says which file is held and that closing the editor fixes it. Other platforms are unchanged.
[2.1.0] — 2026-10-01
UX laws as a measured fix pass, and the Miller misreading removed from the pack.
Added
enhance-ux-laws. Measure-then-fix pass on one screen or flow against seven Laws of UX: Fitts, Hick, Miller, Jakob, Zeigarnik, Goal-Gradient and Von Restorff. A browser probe measures tap targets, competing accents, choice sets, unlabeled fields and progress cues at phone and desktop widths; an eight-check flow walk covers what the DOM cannot show (resume after reload, dead controls, pasted codes). Each law states its precondition and what it does not say. Progress and resume cues must read real state, and drafts never store card numbers or one-time codes.
Changed
- Miller and Hick read correctly.
audit-uxandenhance-web-uxno longer cap menus or navigation at seven items: Miller is about what the user must hold in mind, and Hick applies only when options must be weighed, not to known-item lookups.audit-uxgains Von Restorff and Goal-Gradient/Zeigarnik rows. - Releases are staged.
npm-publish.ymlrunsnpm stage publish --provenance; a version goes live only when the maintainer approves it with 2FA. CI actions moved to v5 (Node 24). - Six skill descriptions dropped duplicate trigger phrases so the listing stays under its 39,000-character ratchet with the new skill (ADR-0010).
[2.0.0] — 2026-10-01
cursor-kenji is now kenji: kensaurus/skills on GitHub, @kensaurus/skills
on npm, plugin kenji@kenji. Same pack, tool-neutral name. Breaking because
the plugin ID and the /cursor-kenji: slash namespace change (ADR-0013).
Upgrade steps: README → "Upgrading from cursor-kenji".
Changed
- Renamed. Repo
kensaurus/cursor-kenji→kensaurus/skills(old URLs redirect), npm@kensaurus/cursor-kenji→@kensaurus/skills, CLIcursor-kenji→kenji, plugin and marketplacecursor-kenji→kenji, Windows shimcursor-kenji.cmd→kenji.cmd. - Existing installs migrate. The installer replaces a pre-2.0.0
completion-gate entry (
cursor-kenji-hooks/) withkenji-hooks/, keeps every other hook and setting, and removes the orphaned folder.--restorealso reads the legacy.cursor-kenji-backups/. Claude plugin installs follow the marketplacerenamesmap.CURSOR_KENJI_GATE_STATE_DIRandCURSOR_KENJI_DIRstill work besideKENJI_GATE_STATE_DIRandKENJI_SKILLS_DIR. - Tool-neutral listing. Descriptions and keywords name Claude Code, Cursor, Codex CLI, and Gemini CLI; the homepage is the GitHub README until the new skills.sh page is indexed.
- Ten skill descriptions dropped duplicate trigger phrases to keep the skill listing under its 39,000-character ratchet (ADR-0010). Every routing pointer stays.
Added
audit-agent-speed. Measures, then fixes a slow coding-agent host: transcript-parsing status lines, per-call and blocking Stop hooks, AGENTS.md skipped under a parent CLAUDE.md, un-ignored worktree folders doubling search, global effort and Explore model, worktree pile-up, antivirus, indexer and vendor WMI pollers. Shipsscripts/stop-typecheck.mjs, anasyncRewakeStop hook that typechecks in the background, skips when nothing changed, wakes the agent only on a new error set, and summarizes codegen cascades.- Routing rule: agent or machine slow →
audit-agent-speed; product slow →audit-performance.
Fixed
docs/AGENTS.template.mdnow says Claude Code skipsAGENTS.mdwhen aCLAUDE.mdexists in the repo or any parent folder, and how to load both.
[1.40.1] — 2026-09-29
The completion gate stops looping. In Cursor it ran twice on every stop, and the copy loaded from the Claude Code config had no cap, so an open checklist sent the same follow-up for hours, including to sessions that did not own it.
Fixed
- The completion gate looped for hours in Cursor. Cursor also runs Claude
Code hooks from
~/.claude/settings.json, with noloop_limit, so every stop was gated twice and the Claude-config copy never stood aside. Each registration now carries--host=cursoror--host=claude, and inside Cursor the Claude-config copy stands aside when the native entry exists (ADR-0012). - Two agents in one repo restarted each other's loop. The follow-up budget
was one workspace counter. It is now per conversation, kept in
~/.cache/cursor-kenji/completion-gate/: three follow-ups while none of the items that conversation saw closes. The old.cursor/completion-gate.count.jsonis deleted on the next run. - The gate pushed work that waits on a person. Items marked
blocked by:,blocked on:, orwaiting on:no longer gate, and neither does a state file untouched for 24 hours. The follow-up now says how to stand down: mark an item that waits on a person as blocked, leave a checklist another agent owns unchanged, say so in one line, and stop.
[1.40.0] — 2026-09-29
Ships the housekeep-files document organizer, with the Windows script fixed
so it actually runs on the PowerShell it claims to support. The fix came from
running the skill on a real Downloads tree: the bundled script would not parse
under Windows PowerShell 5.1, and an earlier run under a Japanese code page
had copied an access-key bundle that its secret rule should have refused.
Added
housekeep-filesskill +/housekeep-filescommand: organize a drifted document tree (Downloads, a synced drive) into per-owner homes by copy, not move. Bundledscripts/housekeep-files.ps1(Windows PowerShell 5.1+) andscripts/housekeep-files.sh(macOS/Linux bash) read the sameplan.tsvand run inventory (SHA-256 of every source) → dry run → apply → verify. Keys, credentials, media over the size cap, and installers are cataloged in place, never copied. Vague filenames get{entity}_{area}_{date}_{doc-type}__{original}; each owner root gets a visible_ai-catalog.jsonland_SEARCH.md.verifyre-checks every source path and size, re-hashes every copy, recounts untouched media folders, and exits non-zero on the first miss. Layout options (root per owner, one root with owners inside, PARA-lite) are offered before the plan is written. Sources inreferences/naming-and-catalog.md.
Fixed
housekeep-files.ps1failed or misbehaved on Windows PowerShell 5.1. The file was UTF-8 without a BOM and held three non-ASCII characters (theアクセスキーsecret pattern and two em dashes). PowerShell 5.1 decodes a BOM-less script with the ANSI code page. Under cp1252 an em dash becomes a curly quote that ends the string, so the script does not parse. Under cp932 it parses, but the CJK pattern turns into mojibake, so a file named for an access key is copied instead of cataloged in place. The script is now pure ASCII, and the pattern is written as a\uXXXXregex escape. Verified by running the old and new script natively under 5.1 on a folder holding an…アクセスキー…zip: the old one throws a parse error; the new one marks the zipindexed-in-place:secret.- Guard:
validate-skillsfails on any shipped.ps1,.psm1, or.psd1that contains a non-ASCII byte without a UTF-8 BOM, and reports the file and line. Probe: a BOM-less em dash in a scratch.ps1fails; the same text with a BOM passes.
If you ran housekeep-files on Windows before this release, check each
_ai-catalog.jsonl for copied-source-kept rows whose path contains
アクセスキー, and rotate those keys.
[1.39.0] — 2026-09-23
Closes the last three open items from the Opus 5.5 pass: the always-on routing rule, facts that went stale, and broken markdown. Each fix was researched from primary sources (package source, vendor docs, npm) and checked by an independent verifier before it was applied.
Changed
skill-workflows.mdchalved (3,246 → 1,614 bytes), ADR-0011. Claude Code loads it in every session. It now lists only the mappings a skill name gets wrong, for exampleworkflow-release-prepnever merges andworkflow-onboardis not product onboarding. It also keeps the multi-phase → workflow principle and the six audit-and-fix exceptions. Everything else was already in skill descriptions or the verification rule.- Supabase MCP calls match current servers. Current servers (the pinned
0.12.0 and the hosted one) hide
get_logsand listquery_logs, which takes ClickHouse SQL filtered bysource. The skills now use it, with SQL taken from the server's own per-service queries. The pack's config is project-scoped (--project-ref), so calls no longer passproject_id, andlist_projectsis described as available only on an unscoped server. Affected:deploy-verify,iterate-post-launch,workflow-fix-and-ship,workflow-feature-flag,workflow-feedback-to-closure,mushi-*,test-*,mobile-emulator-test,audit-db-schema,audit-fe-api,audit-langfuse-llm,debug-fe-be-integration,design-prd, and thefull-stack-ship-disciplinerule. - Sentry MCP calls match the current server.
get_issue_breadcrumbs,get_issue_tag_values, andfind_releasesare catalog tools, reached throughexecute_sentry_tool;search_sentry_toolsis the fallback.update_issueneeds the Triage skill on the connection.- Relative issue-search dates take a sign (
firstSeen:-1h).
enhance-pwauses current facts:- Next.js caching goes through Serwist;
@ducanh2912/next-pwais webpack-only. - Lighthouse 12 removed its PWA category, so installability is checked in
DevTools.
workflow-launch-readyand the catalog no longer ask for a PWA score. navigateFallbackis the SPA shell, not an offline page.beforeinstallpromptfires only in Chromium.
- Next.js caching goes through Serwist;
enhance-web-instant-navgives the current Speculation Rules eagerness defaults and Next.js 16.3 Instant Navigations.plan-capacitor-hardeningnames the current Capsec package (@capgo/capgo-sec) and its rule families, and says Ionic Identity Vault stopped new sales in 2025-02.- No calendar-stamped advice.
- Search queries use
[current year]. - "(2026)" labels on current guidance are gone.
- A cost tier is named as a tier, not a model.
- Token-cap and streaming-usage fields are named per provider.
- The OWASP red-team checklist says it maps the 2021 edition.
- Search queries use
Added
scripts/check-md-structure.mjs, innpm testand CI, fails on a nested or unclosed code fence.check-skill-refsflagssupabase:get_logs.
Fixed
- 13 code-fence errors in 9 files. A nested or unclosed fence rendered
the rest of an output template as live markdown. Affected:
design-prd,plan-dead-code,audit-db-schema,audit-langfuse-llm,audit-uiux-design-system,backend-error-handling, anddocs/CONTRIBUTING.md. - Duplicate headings. The ones in
audit-code-qualityanddocs-coauthornow have distinct names. The catalog's duplicateaudit-skill-conflictsentry is now a pointer to its Audit entry. - The README said
audit-*skills may fix things. It now says they report and stop by default, names the six audit-and-fix skills, and notes the few others that end with an explicit apply step. CITATION.cffstill said version 1.31.0.
[1.38.0] — 2026-09-23
Follow-through on 1.37. A section-by-section check of the Opus 5.5 plan against the repo, run in fresh contexts, found edits that were only half applied. Reading the Claude Code client showed its skill-listing budget works differently from what ADR-0008 assumed.
Changed
- Skill descriptions trimmed. 137 descriptions were rewritten to about 230 chars. The rewrite kept the distinct trigger phrases and the handoffs that separate overlapping triggers, and dropped filler, synonyms, and lists of chained skills. The pack's Claude Code listing drops about 15%, from 45,573 to 38,979 chars; every request carries that listing on both hosts.
validate-skillsmeasures the listing the way the client does: one- name: descriptionline per model-invocable skill and command, with agents reported separately. It fails aboveLISTING_MAX_CHARS(39,000). ADR-0010 supersedes ADR-0008's measure and corrects its facts:- the budget is 1% of the context window at 3 chars/token for current models, 30,000 chars on a 1M Opus 5.5 session;
- built-in skills share that budget;
- over budget, the least-used skills show by name only, and nothing is dropped.
- Keep every description visible on Claude Code with
"skillListingBudgetFraction": 0.02in~/.claude/settings.json. The installer prints the installed listing size and this setting; it does not change your settings. - Authoring surfaces say T1–T8:
enhance-skill-prompts,meta-skill-creator,audit-skill-conflicts, CATALOG, and the playbook describe T2 as a classification contract and T4 as an evidence rubric. docs/DISTRIBUTION.md: the completion-hook column lists Claude Code and the plugin's Stop hook, and warns against running the plugin hook and the--claudehook together.
Fixed
- Leftovers from 1.37.0's mechanical edits. In about 30 files a passage
was replaced but its old lines stayed under the new ones:
audit-uxhad a second set of sections 0b–0e;audit-securitykept its old vulnerability examples;audit-code-reviewkept its DO/DON'T lists after the new review rule;enhance-web-seohad broken JSON in its research step;debug-erroranddebug-sentry-monitorrepeated checklists;tdd-patternshad an orphan comment;- several paragraphs had fallen into the list or blockquote above them.
backend-realtimeprinted secrets. It told the agent tocat .env*, which prints secret values; it now lists variable names only.rules/approved-plan-execution.mdchad invalid YAML frontmatter: an unquoted:in the description. A strict parser rejects it, which can drop the rule'spaths:scope.- Merge installs now prune leftovers from older releases on both hosts:
- the renamed
/plancommand, pruned only while it still carries the pack's description; - the
native-rn-monorepocommand bundle, which older installers copied into the global commands folder, registering tennative-rn-monorepo:*commands in every project.
- the renamed
migrate-to-skillsskipped creating the skills folders when no subagent tool was available.create-skillandmeta-skill-creatortold every skill to repeat the always-on update rule. They now say never to suppress it, and to add a turn-end line only for unattended skills.audit-auth-flowsand its annotated exemplar match again: same effort key, same finding-shape wording.- Closure runs could not record a verification rung that does not apply.
In a headless
/complete-everythingprobe on a repo with no typecheck or lint, the unchecked rungs made the Stop hook block, and the model invented a[-]marker that the hook then skipped with no reason required. The skill now defines- [-] <rung> — n/a: <evidence>. The completion gate counts a[-]line without ann/a:reason as open, andcompletion-judgechecks the reason. The skill also notes that Claude Code shows "Stop hook error occurred" for every Stop-hook block; that is the gate working, not a failure.
[1.37.1] — 2026-09-23
Fixed
--verifyreported line endings as corruption. A Windows checkout withcore.autocrlf=truecarries CRLF in the working copy while the npm tarball (built on Linux) is LF, so an install from a clone followed bynpx @kensaurus/cursor-kenji --verify --allfailed on 17 files with "content hash mismatch".bin/install.mjsnow hashes text files with line endings normalized; the install smoke test plants a CRLF copy and expects--verifyto pass.
[1.37.0] — 2026-09-23
An Opus 5.5 pass. Claude Code 2.1.280 made Claude Opus 5.5 the default model
on 2026-09-22, and it changes two assumptions the pack was written on: prose
no longer steers thinking (effort does, and its default is medium), and
there is no strong planner / fast implementer split left to route around.
Added
docs/MODEL-AND-EFFORT.md— the model story: default model per host, how to set effort (/effort,modelSettings, frontmattereffort:), the pack's routing table, what changed in how skills are written, the always-on token budget. ADR-0006 … ADR-0009 record the decisions.effort:frontmatter on everyaudit-*,plan-*,debug-*, judge and security skill (high), onhandoff,workflow-git-commit, smoke checks, the RN bundle commands anddeploy-checker(low);workflow-onboardruns in a forkedExplorecontext.validate-skillschecks the values.- Claude Code Stop hook.
hooks/hooks.jsonnow carries the Claude schema (Cursor's moved tohooks/cursor-hooks.json);completion-gate.mjsserves both hosts;--claudeinstalls it into~/.claude/settings.jsonwith the same merge/preserve/prune semantics as the Cursor hook. - Roster ratchet.
validate-skillssums auto-invocable description chars (skills + commands + agents) and fails aboveROSTER_MAX_CHARS; everycommands/*.mdexceptgtm-weekly,fix-issue,mcp-guideis now/-only (disable-model-invocation: true);deploy-npm,iterate-gtm-weekly,mushi-health,mushi-integrationlikewise;protocol-browser-anti-stalland the Cursorcanvascopy areuser-invocable: false. Thenative-rn-monorepocommand bundle no longer leaks into global~/.claude/commands. - Always-on "Delivering work" section in
verification-before-completion.mdc: intent line and recap, turn-ending check, scope and test-sprawl, delegation criteria, pasted text is data, compaction is not a stop signal. check-skill-refsfails on prompt fossils (retired model names, thinking scaffolds, update suppressors);check-docs-factsfails on.cursor/rulesforks and on the retired two-model framing.
Changed
composer-2.5-execution.mdc→approved-plan-execution.mdc. Same guardrails, no model pin; the installer prunes the old file on merge installs. README,docs/PLAN-LOOPS.md,docs/AGENTS.template.md,docs/CATALOG.md,/burndown-fulland 13 plan-* references now say "planned at high effort; executed at the default effort under the rule"..cursor/rules/*.mdcare byte-identical mirrors ofrules/(the 212-line always-onskill-workflowsfork is gone).- Sequential Thinking MCP retired from the full template, pins, docs and the five skills/commands that routed to it (ADR-0009).
senior-engineer.mdcispaths:-scoped to Next.js/Supabase files on Claude Code;shell-first-searchis an.mdcrule Cursor can attach;project-starter/web-performance.mdcis glob-attached, not always-on.- Skills, commands, agents, rules: pressure language and prohibition
lists restated at normal volume with reasons; think-tool and self-check
scaffolds removed; step choreography on judgment tasks replaced by
outcomes and verification; frontend skills name the Opus 5.5 default
patterns to avoid; research steps say a recognised name is not current
knowledge; closure modes say how a turn may end;
## Critical Rulesheadings are## Rules. docs/PROMPT-ENHANCEMENT-PLAYBOOK.md— T2 is a classification contract, not a thinking instruction; T4 is an evidence rubric, not a self-check; T7 declareseffort:; T8 sets volume, shape and communication.enhance-skill-prompts,meta-skill-creator,create-skill,create-subagent,docs/CONTRIBUTING.mdteach the same.llms.txt,.claude-plugin/*.json, README, GETTING-STARTED, CATALOG, PROMOTION — name the model and the effort routing.
Fixed
scripts/test-install.mjs— the Windows shim test rancursor-kenji.cmdfrom the cwd, which hardened shells block withNoDefaultCurrentDirectoryInExePath=1. The test now clears that variable for the child and also exercises the explicit.\cursor-kenji.cmdform.
[1.36.3] — 2026-09-22
A command-surface pass. /plan became a host built-in in three of the four
supported tools while we still shipped a file of that name, and four commands
still promised tools the pack stopped shipping. The collision guard now covers
every install target instead of Claude Code only.
Changed
/plan→/plan-mode, in both command sets. Claude Code, the Cursor CLI, and Gemini CLI each ship a built-in/planthat enters Plan mode, socommands/plan.mdduplicated it andcommands-portable/plan.mdshadowed it as~/.gemini/commands/plan.toml. Both are renamed; the file is the procedure, the built-in is the mode switch. Natural-language triggers are unchanged. Restart the host after updating — Claude Code and Cursor registercommands/at session start.commands/mcp-guide— describes the actual default template (Firecrawl, Context7, Supabase) and headedplaywright-cli. Sequential Thinking and the Playwright MCP are full-template only, so nothing routes to them by default.commands/fix-issue— dropped aSemanticSearchcall and the hardcodednpm run test:unit. Diagnosis routes todebug-error, the ship loop toworkflow-fix-and-ship, and verification runs the repo's own checks.commands/debug-issue— no longer claims Sequential Thinking is part ofdebug-error; notes that/debugis a built-in in both hosts.commands/test— tagline matches the body: route to the matching test skill rather than run one generic suite.commands-portable/research— resynced withskills/research: repo first, docs for the pinned version, no implementation until asked.docs/CATALOG.mdaudit-uxandskills/plan-uiux-unification— stopped listing Sequential Thinking as available; the skills never called it..claude-plugin/marketplace.json— listing said 145 skills and 57 commands; the real counts are 156 and 62.
Fixed
- The reserved-command guard only covered one of four hosts. It scanned
commands/against a Claude Code list, so acommands-portable/name could collide with a Gemini CLI built-in unnoticed — which is exactly how/plansurvived there.scripts/validate-skills.mjsnow scans both directories against the documented built-ins and aliases of Claude Code, the Cursor CLI, and Gemini CLI. Probe: acommands-portable/tools.mdfails the gate. - The count ratchet skipped both Claude Code manifests.
.claude-plugin/was never checked, which is how its listing drifted 11 skills and 5 commands behind.scripts/check-skill-count.mjsnow coversplugin.jsonandmarketplace.json.
Known issue
- Codex CLI deprecated custom prompts and stopped loading
~/.codex/prompts(0.117.0), which is where--codexwrites the portable playbooks. Its merged~/.codex/AGENTS.mdrules are unaffected. Choosing a replacement surface touches ADR 0001 and is deliberately not bundled into this release.
[1.36.2] — 2026-09-17
A plan-dead-code → housekeep-dead-code pass on the pack's own
repository, run with the pack's own skills. The JS surface was already
clean; the deliverable is the gate that keeps it that way.
Removed
scripts/migrate-browser-mcp-to-cli.mjs— the one-off Playwright MCP → playwright-cli migration script from 1.2x. Nothing ran it (no npm script, workflow, or doc); Knip in default and--productionmode and a repo-wide reference grep agreed. The CHANGELOG line that introduced it stays as history; git keeps the file.
Added
knip.jsoncandnpm run check:dead-code/check:dead-code:prod(npx -y knip@6.36.0 --max-issues 0, config hints as errors on the default run). Entries come frompackage.jsonbin/scripts and the workflowrun:lines via Knip's plugins, plus the shipped helper scripts and the three maintainer-run brand/description tools nothing invokes by reference. The two consumer-side imports ofskills/enhance-readme/scripts/record-readme-tour.mjs(playwright,@ffmpeg-installer/ffmpeg) are the single written-back false positive. Appended tonpm testand to the existing job invalidate.ymlandnpm-publish.yml— no new parallel check. Baseline 0 / 0; the number only ever goes down. Probes: an unreferencedscripts/zz-probe.mjsfails the gate; a fresh clone passesnpm test.docs/examples/plan-audits/cursor-kenji/plan-dead-code.md— the audit (stack, baseline, surfaces Knip cannot see, keep/kill list with evidence, ratchet) that approved the deletion and the config.
Changed
docs/PUBLISHING.md— thenpm testgate list names the dead-code ratchet.
[1.36.1] — 2026-09-17
Fixed
plan-gtm/references/benchmarks-2026.md— added the "warm invites 10–25%" row (nativeviralloop) thatenhance-growth-loopsalready quoted. The completion judge caught the number as untraceable against the reference file's own rule ("quote a number only with its row here"); the source was in the research all along, the row was not.
[1.36.0] — 2026-09-17
The six follow-ups named in the 1.35.0 report, built in one closure pass:
pricing, growth loops, comparison pages, registry listing, the weekly loop,
and lifecycle email. plan-gtm and workflow-gtm now hand off to them by
name, so the go-to-market sequence has an owner for every step from "what
to charge" to "what did last week's post do".
Added
plan-pricing— plan-only pricing and packaging audit: inventory of plans/gates/metering/price points/trial logic/cost, value-metric scoring (value connection, fairness, predictability, scalability, billability) simulated on historical accounts, ≤3 good-better-best tiers with a right-hand anchor and an enterprise "from" price, free boundary just before the activation event's repeat, price corridor via Van Westendorp → Gabor-Granger → MaxDiff → conjoint. Emitsplan-pricing.mdwith a grandfathering row; changes no price, plan enum, Stripe object, or copy. Command/pricing-plan.enhance-growth-loops— apply-now badge / shareable artifact / template / invite / referral loops, one K-factor per loop; rewards granted idempotently on the referred user's activation; invites user-initiated and single-send with recipient opt-out; badge removable on paid.docs-comparison-pages— ≤5 honest "X vs Y" / "alternatives to" / "migrate from" pages per pass from facts fetched and dated this session; fixed shape including "where the alternative wins";lastVerified+reviewEveryfreshness contract; refuses bulk generation.audit-registry-listing— read-only audit of README first ten lines, npm/PyPI metadata andnpm pack --dry-runcontents, GitHub description / topics / social preview, plugin and skills.sh manifests, and a timed clean-profile install. Findings with an owner skill per row.iterate-gtm-weekly— thirty-minute weekly loop: funnel by source this week vs last, launch Results tables filled, largest absolute loss named, last experiment attributed with a number, exactly one experiment and one post chosen from a levers-by-step table, scorecard row appended. Command/gtm-weekly.enhance-lifecycle-email— event-driven sequences (welcome, activation nudge, setup stall, habit, trial expiry split by activated vs stalled, limit reached, win-back, dunning) with triggers, exits, idempotent(user, sequence, step)sends, per-email transactional/marketing classification, unsubscribe and suppression, quiet hours, daily cap, andemail_*events. Timers only as the fallback for silence.
Changed
plan-gtm— phased burndown names the new owners: Phase 1 addsaudit-registry-listingfor repo-as-product; Phase 2 addsenhance-lifecycle-email; Phase 3 usesdocs-comparison-pages; Phase 4 addsenhance-growth-loopsand hands the loop toiterate-gtm-weekly; Phase 5 starts withplan-pricing. Neighbor table extended.workflow-gtm— sequence, step sections, scorecard, and worked example route through the same owners; Step 7 isiterate-gtm-weekly(iterate-post-launchkeeps production defects).plan-gtm/references/benchmarks-2026.md— new sections for pricing research / value metric and lifecycle email, each row with its URL.docs/CATALOG.md,docs/TRIGGER-CHEATSHEET.md,docs/PLAN-LOOPS.md(23 plan skills,/pricing-planalias, execute-after-approval rows),skills.sh.json, README ("What should I say?" and commands rows) — the six skills and two commands listed with their neighbors.
Fixed
- The five 1.35.0 GTM skills and the README skill index carried
double-encoded arrows (
→for→) in their descriptions — a shell encoding slip during the description trim. Re-encoded to UTF-8; the installed copies refresh on--auto.
[1.35.0] — 2026-09-17
Go-to-market for a repo that already works. The pack covered how to build, prove, and ship a product but had nothing for the part after: who it is for, how it is paid for, why signups never come back, where to be found, and how to announce a release without a vote ring. Five skills and three commands close that gap. Every benchmark quoted was checked against a 2025–2026 source this week (ChartMogul × Kyle Poyar × ProductLed free-to-paid report, Userpilot activation medians, Show HN / Product Hunt author reports, public OSS-vendor revenue mix).
Added
plan-gtm— plan-only go-to-market audit. Greps how the product is sold today (billing SDKs,isProgates, license family, hero/meta positioning, auth providers and steps to first value, analytics + consent, robots/sitemap/llms.txt, registry presence, trust signals), then interviews the founder withworkflow-grillingrules down a fixed ladder (90-day metric → ICP → numbers → monetization intent → budget → market → alternatives → constraints). Diagnoses in Dunford order, judges the monetization model against 2026 conversion rows (freemium 3–5% good / 8–12% great; card-required trial 25–35%; median 8%; OSS cloud carries 48–73% of vendor revenue), definesactivatedand one north-star, picks ≤2 channels + 1 loop with caveats. Emitsplan-gtm.md; changes nothing. Business decisions (model, price, license, ICP) stay with the founder; "unmeasured" is a legal cell value.workflow-gtm— execution arm.plan-gtm⏸ approve →audit-analytics→enhance-web-conversion(+enhance-readmefor repo-as-product) →enhance-onboarding→enhance-web-seo/plan-aeo-readiness→docs-launch-kit→iterate-post-launchweekly. Ends with a GTM scorecard (funnel before/after, skips named, next week's one fix).enhance-onboarding— apply-now first session to first value. Definesactivatedas an outcome (never signup or tour end), walks landing → activation as a stranger and counts steps, cuts fields before value (OAuth/magic link, inferred locale, try-before-signup), ships empty-state templates / sample data and a ≤5-item checklist ending at activation, instrumentssignup_started → signup_completed → activated → habit_reachedbehind the existing consent gate. Defers, never deletes.enhance-web-conversion— apply-now message / offer / proof / prompt pass on landing, pricing, and upgrade paths: five-second hero check with a model-matched CTA verb, ≤3 tiers with an anchor and annual toggle, explicit free-tier limits, proof ordered by strength, upgrade prompts only at value moments, checkout and prompt events. Removes dark patterns; adds none. Visual system stays onenhance-web-landing.docs-launch-kit— versioneddocs/launch/<version>.md: claims table every sentence must trace to, one problem-first angle, native copy per channel (Show HN, Product Hunt, subreddit-specific Reddit, X/Bluesky/ LinkedIn, article outline, user-facing release notes), calendar, UTMs, Results table filled after 7 days. No vote solicitation, no fabricated proof, maker disclosed.- Commands
/gtm-plan→plan-gtm,/gtm→workflow-gtm,/launch-kit→docs-launch-kit.
Changed
rules/skill-workflows.mdc— bundle row "take a shipped product to market →workflow-gtm".docs/PLAN-LOOPS.md—plan-gtmin the execute-after-approval table, the plan skill map (22 plan skills), launch gates, slash aliases, and a copy-paste prompt; related-loops row.docs/CATALOG.md,docs/TRIGGER-CHEATSHEET.md,skills.sh.json, README ("What should I say?" row, commands table, plan → apply callout) — the five skills and three commands are listed where their neighbors are.- MCP pins re-verified against the registries on 2026-09-17
(
mcp/pinned-versions.json, both templates,.mcp.json,mcp/VERSIONS.md,mcp/README.md):@upstash/context7-mcp3.2.2 → 4.1.1 (v2 SDK, 2026-07-28 protocol; stdio flags unchanged),firecrawl-mcp3.21.3 → 3.24.0,@supabase/mcp-server-supabase0.8.2 → 0.12.0 (--project-ref/--read-only/--featuresstill accepted),chrome-devtools-mcp1.3.0 → 1.9.0,@playwright/mcp0.0.76 → 0.0.81,@notionhq/notion-mcp-server2.4.1 → 2.5.1,server-sequential-thinking2025.12.18 → 2026.8.31,server-memory2026.1.26 → 2026.8.31,awslabs.lambda-tool-mcp-server2.0.19 → 2.1.1,awslabs.aws-api-mcp-server1.3.45 → 1.5.5,awslabs.cloudwatch-mcp-server0.0.8 → 0.2.1.server-github,-postgres,-redis,-slackhad no newer release. scripts/test-install.mjsreads the expected firecrawl pin frommcp/pinned-versions.jsoninstead of a hard-coded3.21.3, so a pin bump cannot pass the pin check and fail the install smoke test for the same fact.
[1.34.0] — 2026-09-12
A prompt audit of the shipped surface (rules, agents, commands, skills)
against current-model behaviour: instructions written for older, less
steerable models were rewritten at normal volume, facts that had rotted were
corrected, and text the agent cannot act on (memory of previous projects,
"production test" scores) was removed. No skill changed name, trigger, scope,
or stance; no [LOW freedom] step lost its exactness.
Changed
skills/backend-patterns— the Edge Function example usesDeno.serve; thedeno.land/std@0.168.0/http/serverimport it carried is deprecated.agents/perf-monitor—npx bundle-phobia <pkg>;npx bundlephobiaresolves to a package with no CLI and silently does nothing.agents/deploy-checker—npm audit --omit=dev(--productionis the deprecated spelling).agents/db-migrator— migration files come fromsupabase migration new(14-digit timestamp) instead of a hand-written 8-digit date prefix that cannot order two same-day migrations.agents/debugger— dropped "don't explain before acting" and "ship first, explain after", which contradicted the agent's own required output format (root cause, evidence, verification, prevention).commands/update-deps— the dated Next.js cadence paragraph ("15 LTS ends 2026-10-21, current line 16.3.x") is now a check-at-run-time instruction; the codemod advice and thenext lintremoval stay.commands-portable/research§7 — re-synced withskills/research: "plan before writing code". The portable copy still said "think step-by-step", which reasoning models already do and which had drifted from the skill.skills/audit-langfuse-llm— model examples are tier-based (frontier / small) instead of pinned IDs that age; its prompt-improvement suggestions recommend the provider's native reasoning mode andmax_tokensrather than "Think step by step" text and word caps in the prompt.skills/enhance-web-landing— register pass: removed eighteen "(mandatory)" tags and the all-caps shouting, the three "most-tested AI tell / #1 violated rule in production tests" grader lines, the cross-project palette and serif rotation rules (the agent has no memory of previous projects), and the one-project "Marrow" references. Every rule and its reason is unchanged; the copy-length limits on page content stay because they constrain the page, not the reply.- Seventeen skills — "CRITICAL:" / "MANDATORY:" / "BEFORE X, you MUST:" step
headings read at normal volume; the
[LOW freedom — run exactly]tag already carries the weight. skills/design-canvas— "CRITICAL Requirements" → "Quality bar".rules/composer-2.5-execution.mdc— title and intro no longer pin Composer 2.5 or a "200k window"; the body already said it binds any implementation model. Filename kept for installer and doc compatibility.rules/shell-first-search.md— dropped the "Windows hang history" / "this Windows host" incident wording from a rule that ships to every consumer.docs/PLAN-LOOPS.md— the execution rule has been on-demand since v1.15.0, notalwaysApply: trueas the doc claimed; the planning model is no longer pinned to Opus 4.8.
[1.33.1] — 2026-09-10
Changed
housekeep-dead-code/references/supabase-hygiene.md— droppedsupabase inspect db table-record-countsfrom the CLI-equivalents line. It rested on a single low-confidence doc extraction from a page that redirected, and never appeared in the observability guide's documented set; a wrong subcommand name in a command position is worse than no name. The line now lists only the six confirmed subcommands, says the set has changed across releases, and points atsupabase inspect db --helpplus the version-independent SQL above it.
[1.33.0] — 2026-09-10
A second independent implementation of the dead-code pair was written against the same base commit. Comparing the two surfaced material worth keeping — and two bugs worth not keeping. Ported the former as reference files; rejected the latter after checking both against the knip 6.35.1 source.
Added
plan-dead-code/references/residue-greps.md— the grep pack as a table with a count command and a fix column per class, plus the false positives that matter: hashed and template-string asset references, and the counting convention (rg -o … | wc -lfor matches vsrg -cfor lines — mixing them makes a ratchet compare nothing). Adds a bundle-leak probe (rg -l "service_role" dist) that stops the pass and routes toplan-secrets-audit; it is the one check here that can find a live credential in shipped output.plan-dead-code/references/output-templates.md— a filledplan-dead-code.mdwith the chain-head row discipline, worked severity × effort rubric, and a root-causes section so the report names the barrel rather than its forty symptoms.plan-dead-code/references/preservation-contract.md— the plan-only non-negotiables and the five verdict buckets, matching theplan-perf-auditconvention.housekeep-dead-code/references/ratchet-ci.md— scripts, tsconfig (noUnusedLocalsenabled during the cascade, not after),eslint-plugin-unused-importsflat config, why Knip belongs in pre-push rather than pre-commit, the aggregator'sjqresult check that makes a skipped job fail, and a Supabase types-drift CI guard.housekeep-dead-code/references/supabase-hygiene.md— the schema arm in full. Three traps a client-side grep cannot see: RPCs called only from an RLS policy, trigger, orpg_cronjob; Edge Functions targeted by external webhooks; andpg_stat_*counters too young to mean anything (checkstats_reset). Tables drop in two steps — rename to_deprecated_<name>, drop a release later — because a rename is reversible.
Changed
plan-dead-code— chain-head reporting is now explicit (a dead file's exports and sole-use dependencies are children of that row, not rows of their own, or a three-item plan reads as twenty-seven); severity is paired with S/M/L effort; adds the verified--trace-export/--trace-file/--max-show-issuesflags; and adds a second gate before triage — unresolved imports and unlisted dependencies both explained, not just hints clean.housekeep-dead-code— new Phase 3 proves the ratchet bites: a fresh-clone probe (deletions can pass on a warmnode_modulesand fail onnpm ci) and a deliberate-violation probe that must turn the aggregator red, not just the job.
Rejected from the ported material
Both were checked against the knip 6.35.1 source rather than its docs:
knip:ci: "knip --production … --treat-config-hints-as-errors"— inert. Knip setsisDisableConfigHints = --no-config-hints || isProduction, so the flag cannot fire on a production run. It belongs on the default-mode script.rules: { classMembers: "warn" }in the recommended config — silently exempts that issue type from the gate, because--max-issuestotals only rules set toerror.ratchet-ci.mdnow documents this as the quietest ratchet bypass instead of shipping it as a default.$schema: ".../knip@latest/schema.json"— floats a major version in a file whose own text says not to floatlatest. Pinned toknip@6.-Was shorthand for--workspace, and a "<50 scans, >5 pages" threshold forsupabase inspect db unused-indexes— neither is in the source or the docs. Dropped rather than repeated.
[1.32.0] — 2026-09-10
Dead-code cleanup was one line inside workflow-housekeep (§2d, "npx knip
(if available) or npx ts-prune") — no config, no production mode, no fix
ordering, no ratchet, and ts-prune has been in maintenance mode with its
author pointing at Knip. Everything else a vibe-coded repo accretes
(duplication, debug residue, suppression debt, orphan assets, env drift,
dead schema) was scattered across five skills as prose with no tool behind
it. This adds the plan→apply pair that owns it.
Added
plan-dead-code— configuration-first dead-code audit. Authorsknip.jsonand resolves every configuration hint before trusting a finding, because on a first run most findings are misconfiguration, not dead code. Baselines both--productionand default runs and reads findings files → unresolved → exports → deps (unused files cascade into phantom export and dependency findings). Builds a keep-working list for glob-imported routes, generated Supabase types, Deno Edge Functions, and deliberate@publicAPI, then counts the seven surfaces Knip cannot see: unused locals,jscpdduplication, debug residue, suppression debt, orphan assets, env drift, dead schema. Emitsplan-dead-code.mdand deletes nothing. Slash/deadcode-plan.housekeep-dead-code— the execution arm. Deletes per the approved keep/kill list one category per commit with typecheck, tests, and build between each, and a bisect-the-batch rule on red: files (--fix-type files --allow-remove-files, pre-reviewed list only) → exports/types → unused-locals cascade (remove-unused-vars, then re-run Knip) → dependencies → residue → assets → suppressions. Duplication is measured and ratcheted but handed toworkflow-refactor, since fixing it preserves behavior rather than removing code. Then installs the ratchet:knipscript, CI job pinned at--max-issues <today>and only ever lowered, wired into the single aggregator gate perhousekeep-gates. Slash/deadcode.plan-dead-code/references/knip-config.md—knip.jsonrecipes for Vite + React + Supabase, Next.js App Router, and monorepos, each line annotated with theignoreit replaces; tag usage (@public,@internal,@alias); companion-tool table; and the ratchet CI job.- Brand mark (
assets/logo.svg+ PNG/OG/favicon). Geometric spine + first-line tick; fal.ai candidates were generated and rejected when they drifted. - Claude Code marketplace manifests:
.claude-plugin/plugin.json+.claude-plugin/marketplace.json. - Static landing page in
site/(canonical stay on skills.sh until an owned host is live). FAL_KEYdocumented in.env.example; secret scanner now flags fal key-id:secret pairs.
Changed
workflow-housekeep§2d now hands off instead of sweeping. It keeps dead artifacts (logs,.bak, build output, committed screenshots) — things dead by inspection — and routes unused source files, exports, types, and dependencies to the new pair. Thets-prunerecommendation is gone. Trigger phrase "remove dead code" moved off this skill.plan-dependency-provenance— boundary drawn: it owns the manifest fact (does the package exist, is it licensed); the module-graph proof and removal of unimported deps go toplan-dead-code. Warns against hand-verifying with grep, since a dep reached only from a config string looks unused and is not.plan-antislop§3 — boundary drawn against provable unreachability: commented-out scaffolding and guards that can't trigger stay slop findings (a taste judgment), while counts and deletion move toplan-dead-code.test-mutation— "delete dead code" now routes to the pair rather thanworkflow-housekeep.workflow-refactor— the "Dead Code → delete it" smell row was the last open side door around the approval contract; it now routes to the pair. Duplication findings arrive there; deletions do not leave from there.package.jsonhomepage is the live skills.sh page.- Cursor plugin manifest:
displayName,logo,category, homepage. - Distribution / promotion docs: Claude plugin install, localskills reject, Show HN draft. No fake “listed” claims.
[1.31.0] — 2026-08-29
Added
enhance-web-instant-nav— Speculation Rules (prefetch/prerender with a safety contract), bfcache fixes, cross-document View Transitions, 103 Early Hints. Slash/instant-nav. First-load CWV stays onaudit-performance.web-performance.mdcproject rule — always-on budgets, LCP priority, INP yield, no-unload.audit-performance§Loading Priority & Speculation, field-data Step 0, verification protocol, andreferences/loading-priority-2026.md.
Changed
audit-performance,audit-bundle-size: React Compiler first; manual memo only with Profiler evidence.plan-perf-auditscope: Navigation layer and loading-priority lines.enhance-web-seo§1d hands CWV fixes toaudit-performance.perf-monitoragent: LCP-lazy, slow-second-page, bfcache, INP rows.- Stack currency: Next.js 15 → 16.x (15 LTS EOL 2026-10-21; monthly
security releases noted in
/update-deps). check-skill-countnow ratchets the README Project Rules count againstrules/project-starter/*.mdc.
[1.30.0] — 2026-08-21
Added
enhance-readability— information-design pass for comprehension (CLT + Gestalt): CPL / reading level, grouping, deadspace, icons, and visuals that replace verbose prose. Slash/readability. Breakpoints stay onaudit-responsive; looks-good stays onenhance-web-ui. CJK 40 CPL from WCAG 2.2 SC 1.4.8.
[1.29.1] — 2026-08-21
Fixed
- T3/T4 headings —
meta-skill-creatorandenhance-skill-promptshad T1–T6 tables/phrases but no## Worked example/## Self-critique.test-exploratoryandworkflow-release-preppromoted T4 to##. - Validator — T3/T4 now require those H2 headings so a table mention cannot pass the gate.
[1.29.0] — 2026-08-21
Added
- Family-first naming policy —
docs/CONTRIBUTING.mddocuments<family>-<topic>, slash aliases, and the four unprefixed exceptions (research,handoff,complete-everything,burndown-full). iterate-family — registered in the skill index (FAMILY_HEADINGS,FAMILY_ORDER, blurbs).- Installer
RENAMED_SKILLSprune — deletes old skill directories on Cursor, the~/.agents/skillsmirror, and Claude after a rename. Honors--dry-run; skipped when--skillis scoped. Tests may inject pairs viaKENJI_RENAMED_SKILLS. OLD_ALIASESgate —check-skill-refsfails leftover backticks of renamed skills except inCHANGELOG.md,docs/CONTRIBUTING.md, andaudit-skill-conflicts.
Changed
housekeep-taxonomy — CATALOG / CONTRIBUTING / index blurb now say "one drifted register (gates / backlog / design tokens)", not "design system only".- Renames —
domain-modeling→docs-domain-modeling;grilling→workflow-grilling./grill-mestill points atworkflow-grilling. "grill me" and glossary triggers unchanged. - Prompt waves (T1–T6) — remaining first-party families after Wave 7
(
housekeep,docs,deploy,debug, then the rest). Validator require-list extended per family.thirdparty-*skipped.
[1.28.0] — 2026-08-21
Added
audit-doctrine— read-only audit of custom lint/ratchet content: is each rule right on the merits (named remedy, reachable token, teaching failure, Tier-D practice), not merely enforced. Distinct fromaudit-gate-logic(bypass/gaming). Un-remedied axes →housekeep-backlog. Slash:/doctrine.
[1.27.0] — 2026-08-20
Added
workflow-release-prep— take the local working tree (uncommitted + staged + untracked + unpushed) to a merge-ready PR against main. Orchestratesaudit-code-review,split-to-prs,workflow-git-commit,workflow-pr, andbabysit. Does not merge. Slash:/release-prep.
Changed
- Skill-conflict routing pass — carved overlapping triggers across the
PR/release, UI/slop, error-handling, tests, backend, and skill-authoring
clusters. Completed truncated descriptions (
burndown-full,backend-patterns,mobile-emulator-start). Movedthirdparty-emil-design-engbody intoreferences/. Validator now fails unmatched quotes/parentheses and trailingi.e.truncation scars.
[1.26.0] — 2026-08-20
Added
researchskill — the production research protocol is now a first-party skill (skills/research/) sonpx skills addand Cursor/share the same playbook./researchis a thin command wrapper (same pattern ascomplete-everything).- Installer
--verify— read-only hash check of every packaged file against the destination. Extra personal files are allowed; missing or stale packaged files fail closed. Merge install now verifies each copy before reporting success.
Fixed
- Install-channel docs —
npx skills addinstalls skills only and does not write slash commands.npx @kensaurus/cursor-kenji --allis the full-pack one-liner. The two--allflags are not the same. - Cursor
/babysitshadow —babysitis a Cursor-managed builtin. The installer no longer copies it into~/.cursor/skillsand deletes stale copies on upgrade (Claude still gets the portable copy). - Dual-name CI gate —
validate-skillserrors on an undeclared skill+command pair so new/picker duplicates cannot ship unnoticed.
[1.25.0] — 2026-08-20
Added
housekeep-backlog(/housekeep-backlog) — apply-now inventory of parked work (unfinished plans, deferred phases, TODO/FIXME, skipped tests, open audit findings) into a livingBACKLOG.mdthat regenerates and diffs (new / done / stale). Collector that feedscomplete-everything/burndown-full. Thedocs-adrpattern applied to parked work.
[1.24.0] — 2026-08-19
Added
enhance-skill-prompts— meta-skill to upgrade an existing SKILL.md prompt (degrees of freedom, structured CoT, one worked example, self-critique rubric, term consistency) without changing what the skill does. Playbook:docs/PROMPT-ENHANCEMENT-PLAYBOOK.md. Annotated exemplar:skills/enhance-skill-prompts/references/exemplar-audit-auth-flows.md.- Prompt-enhancement playbook — T1–T6 applied to every first-party
audit-*/plan-*/test-*skill plus the Wave 4 apply-now set (frontmatter and stance unchanged). Liveaudit-auth-flowsis the reference implementation. Factory (meta-skill-creator,docs/CONTRIBUTING.md) now inherits T1–T6 for new skills. validate-skillsT1/T3/T4 gate — first-party audit/plan/test skills fail CI if they lack Degree of freedom, a Worked example, or a Self-critique rubric.
[1.23.1] — 2026-08-19
Fixed
- Windows
npxfrom a clone —npx @kensaurus/cursor-kenjifrom the repo folder failed withcursor-kenji is not recognizedbecause npm exec uses the local package andcmdlooks for a cwd shim that did not exist. Addedcursor-kenji.cmdplus abin/cursor-kenji.jsentry so the same command works from the clone and from any other directory. Registrynpxfrom a non-clone folder already worked.
[1.23.0] — 2026-08-19
Added
test-exploratory— unscripted guest vs logged-in wander of a live non-prod app (headed playwright-cli, SBTM charters, guest/auth/post-logout isolation). Discovery only; tickets viaworkflow-feedback-to-closure, lock viatest-playwright.- Highest-impact combo pipelines in README, GETTING-STARTED, and CATALOG:
test-exploratory→workflow-feedback-to-closure→test-playwright, plus wander-then-workflow-quality-gate.
Changed
test-qa/test-playwright/test-red-teamdescriptions no longer claim monkey / guest-vs-authed wander;/testand the trigger cheatsheet route those phrases totest-exploratory.
[1.22.0] — 2026-08-19
Added
audit-auth-flows(/auth-flows) — read-only app-layer auth audit. Route×gate matrix, SupabasegetSession()vsgetUser(), middleware-as-only-gate (CVE-2025-29927), refresh-token reuse detection, OAuth/IDOR. Data-layer half stays onplan-rls-audit.
Changed
audit-security/plan-security-audit/plan-rls-auditno longer own session/route-gate / getSession triggers.
[1.21.0] — 2026-08-18
Added
housekeep-gates(/housekeep-gates) — apply-now consolidation of accreted CI gates into one aggregator required check. Execution arm ofaudit-gate-logic.test-mutation(/test-mutation) — StrykerJS / mutmut harness; kills assertion-free coverage theater.enhance-arch-boundaries(/arch-boundaries) — dependency-cruiser / eslint-boundaries as CI fitness functions.docs-adr(/adr) — lightweight ADRs with rejected alternatives; agent-readable INDEX.md.
Changed
audit-gate-logic— Phase 2.5 gate archaeology (duplicate vintages, competing baselines, hook-vs-CI, dead gates) and winner selection forhousekeep-gates.- Neighbor carve-outs:
plan-test-coverageno longer owns mutation execution;audit-backend-architecturedefers mechanical rules;plan-docs-sync/docs-coauthordefer decision memory;test-unitdefers assertion theater.
[1.20.0] — 2026-08-18
Added
audit-gate-logic(/gate-logic) — read-only audit of CI/CD gate logic: silent bypass, ratchet gaming, conflicting workflow conditions, required-in-name-only checks. Distinct fromaudit-cicd(cost/speed/safety) andworkflow-quality-gate(which runs gates).audit-codemod-safety(/codemod-safety) — read-only audit of a mechanical source transform for behavior-preservation. Distinct fromaudit-code-review(diff quality) andplan-data-integrity(SQL/data migrations).
Changed
- Neighbor descriptions now carve those owners:
audit-cicd,workflow-quality-gate,audit-code-review,plan-data-integrity,enhance-agent-guardrails,workflow-green-repo. check-skill-countnow rewrites leftover PROMOTION.md count tokens (One install, N skills, article title).
[1.19.1] — 2026-08-18
Routing pass from audit-skill-conflicts (no new skills). Description carve-outs
so overlapping triggers pick one owner; audit-* is no longer claimed as
universally read-only; dangling mobile-mobile-* refs fixed; /test and /uiux
routers expanded; check-skill-refs gate added to npm test and CI.
Changed
- Skill descriptions now name a single owner for LLM quality vs security, RLS/secrets vs the OWASP plan, UI/motion/housekeep, QA vs stubs, and third-party name-gating.
audit-*default remains present-then-stop. Audit-and-fix exceptions:audit-responsive,audit-code-quality,audit-performance,audit-security,audit-i18n,audit-bundle-size(plusenhance-email-deliverability)./testroutestest-loadandtest-visual-regression./uiuxholds the enhance surface router.plan-data-integrity//integrity-planno longer own restore drills (those stay onplan-backup-dr).workflow-spec-tddno longer referencesmobile-mobile-*.
Added
check-skill-refs— CI gate for doubled-prefix typos and the staleaudit-responsive-layoutalias.
[1.19.0] — 2026-08-18
Adds the ops / launch gap pack (13 skills): product analytics, UI state matrix, LLM attack surface, IAP, env parity, infra cost, visual + load tests, privacy/DR/ASO plans, email deliverability, and a pack self-audit. Skips the pack's audit-responsive-layout — already shipped as audit-responsive in 1.18.0. audit-llm-security is new (source tarball was missing that file).
Added
audit-llm-security— OWASP LLM Top 10 on app-facing AI (not Langfuse quality, not coding-agent policy).plan-privacy-compliance+/privacy-plan— collection-vs-claimed, consent, deletion/export, GDPR / APPI / store labels. Cross-linked withaudit-analytics.plan-backup-dr+/backup-plan— restore capability, RPO/RTO, drill evidence (not destructive-op gates).audit-analytics— funnel coverage, taxonomy, consent-gated firing.audit-ui-states— empty / loading / error / offline / overflow matrix.test-visual-regression— Playwright screenshot baselines + CI diffs (playwright-cli, not MCP).test-load— k6/Artillery journeys; never hits prod unsigned.audit-monetization-iap— StoreKit / Play / RevenueCat (not Stripe/web).plan-aso+/aso-plan— store listing find + install conversion.enhance-email-deliverability— SPF/DKIM/DMARC + hygiene (keptenhance-*besidedesign-email).audit-env-parity— config agreement across envs (local runnability stays onworkflow-environment-ready).audit-infra-cost— hosting/egress (CI minutes →audit-cicd; tokens →plan-llm-cost-guardrails).audit-skill-conflicts+/skill-conflicts— pack self-audit: contradictory directives, overlapping triggers, stale cross-refs, context bloat. Per-file spec stays onvalidate:skills.
Changed
design-email— no longer claims deliverability / SPF / "why is my email in spam".plan-data-integrity— no longer claims backup drills / disaster recovery; those route toplan-backup-dr.- Neighbors —
audit-responsive,audit-langfuse-llm,plan-llm-cost-guardrails,plan-mobile-readiness,audit-payment-system,audit-cicd,workflow-environment-ready,iterate-post-launch,plan-stub-checker,enhance-agent-guardrails,test-red-team,audit-resilience,data-visualizationnow point at the new skills.
[1.18.0] — 2026-08-18
Adds audit-responsive: the missing layout/IA audit for "desktop is a linearized phone." Existing skills covered mobile-up (design-mobile-first), cross-page IA (audit-ux-journeys), tokens (audit-uiux-design-system), and polish (enhance-web-ui / enhance-web-ux) — none owned the 10 breakpoint anti-patterns + wireframe-then-implement loop.
Added
audit-responsive— detect stack, score 10 layout/IA anti-patterns (file:line+ severity), extract or propose a token sheet, ASCII wireframes at 375 / 768 / 1440, then implement in reviewable chunks (report-first if scope is large). Live verify via playwright-cli. References:references/checklist.md./responsive-audit— thin slash entry. Optional path scope.
Changed
design-mobile-first— description no longer claims generic "responsive" (that stole this skill). Routes linearized-desktop complaints toaudit-responsive.- Neighbors —
plan-uiux-unification,audit-ux,audit-ux-journeys,audit-uiux-design-system,/uiux,docs/PLAN-LOOPS.md, catalog compositions, and the trigger cheat sheet now point at the new skill.
[1.17.0] — 2026-08-15
Cursor/Claude config housekeep: cut always-on rule weight, stop shipping Cursor builtin skill dupes, slim default MCP, interpolate secrets.
Changed
- Skill descriptions — house budget 320 characters (was 1024). 86 descriptions shortened; bodies unchanged.
pathson RN/Capacitor skills;disable-model-invocationon slash-onlyskills-cursorutilities. - Installer — Cursor skips 11
skills-cursornames that Cursor already ships as managed builtins (keepsbabysit). Claude still gets portable copies.--no-agents-mirrorand--quietadded.install.shis a wrapper aroundbin/install.mjs.cursor-syncfinds~/Documents/GitHub/cursor-kenjiand calls the Node installer. - Rules —
full-stack-ship-disciplineis agent-selected (no auto-push).composer-2.5-executiondrops emptyglobs.shell-first-searchis a hang-fallback, not a Grep/Glob ban. - MCP — essential template is Firecrawl (authenticated, feedback flags off) + Context7 + Supabase with
${env:NAME}interpolation. Sequential Thinking and Playwright MCP are full-template only. Research command no longer requires Sequential Thinking MCP.
Added
scripts/shorten-skill-descriptions.mjs— description budget rewriter used for this release.
[1.16.0] — 2026-08-04
Patterns adopted from mattpocock/skills (MIT) after a full comparison of the two packs: alignment-first interviewing, shared domain language, session handoff, intent-first conflict resolution, and the "writing great skills" verbosity levers.
Added
grillingskill +/grill-mecommand — relentless one-question-at-a-time interview to close the user↔agent alignment gap before any code: recommended answer with each question, facts looked up vs decisions asked, no edits until confirmed, closes with a decision log that feedsdesign-prd//plan.domain-modelingskill — build aCONTEXT.mdglossary and sparing ADRs (hard-to-reverse + surprising + real trade-off only); challenges conflicting terms, sharpens fuzzy language, cross-references claims against code.handoffskill +/handoffcommand — compact the session into a handoff doc (state, ordered next steps, suggested skills, gotchas) in the OS temp dir; secrets redacted, artifacts referenced not copied, verification claims held to theverification-before-completionladder.disable-model-invocation: true, so it costs zero always-on context.workflow-merge-conflictsskill — resolve in-progress merges/rebases by tracing each hunk to its original intent (commits, PRs, issues) before resolving; repo checks re-run before finishing.
Changed
meta-skill-creator+create-skill(Cursor) — added the verbosity levers: no-op test (delete lines the model already obeys), positive phrasing over prohibition, leading words ("fast, deterministic, low-overhead" → tight), single source of truth, one trigger per branch in descriptions, checkable completion criteria; plus model-invoked vs user-invoked (disable-model-invocation) cost guidance.debug-error— Phase 1 now requires a tight feedback loop before any hypothesis: one red-capable, deterministic, fast command already run once, asserting the user's exact symptom; Phase 2 gains repro minimisation (every remaining element load-bearing); Phase 6 requires the loop re-run green and the minimised repro converted to a regression test at a correct seam.workflow-spec-tdd— TDD phase now tests only at pre-agreed seams (named in the plan, confirmed with the user), works in vertical slices (tracer bullets, not all-tests-first), bans tautological tests (expected values from an independent source of truth), and treats refactor-broken tests as implementation coupling to fix at the seam..cursor-plugin/plugin.json— re-synced version withpackage.json(was left at 1.14.1 by the 1.15.0 release).
[1.15.0] — 2026-08-02
Always-on rules audit (2026-08-02): the shipped rule set cost ~11k tokens per chat once installed globally, against a ~2–4k healthy budget, and per-project template bundles leaked into every repo. Verified against current Cursor docs (alwaysApply: true ignores globs/description; rules ≤500 lines; skills load on demand) and Claude Code docs (CLAUDE.md ≤200 lines; paths: frontmatter is the only conditional loader in .claude/rules/; hooks for enforcement).
Changed
- Installer (
bin/install.mjs,install.sh) — per-project rule bundles (native-rn-monorepo/,project-starter/) are no longer copied into global~/.cursor/rulesor~/.claude/rules; they are documented as copy-into-project bundles and their always/glob rules were firing in every repo (e.g. RN no-Mac conventions loading in Vite web projects). rules/skill-workflows.mdc— demoted from a 172-line always-on catalog (~2.1k tokens/chat) to a 46-line agent-requested routing index (alwaysApply: false); skills self-describe and load on demand, so the full table was dead weight.rules/composer-2.5-execution.mdc— demoted to agent-requested for Cursor andpaths-scoped (**/plan-*.md,**/plans/**) for Claude Code, so plan-execution guardrails load only during plan runs.rules/senior-engineer.md→.mdc— cut from 122 lines of generic protocol to 26 lines of stack defaults + security non-negotiables, demoted to agent-requested; renamed because current Cursor ignores plain.mdrule files.rules/full-stack-ship-discipline.mdcandrules/verification-before-completion.mdc— kept always-on but rewritten at roughly half the length; intent (deploy-schema-with-UI, evidence-before-completion ladders, completion-judge gate) preserved.scripts/test-install.mjs— asserts the bundle exclusion (nonative-rn-monorepo/orproject-starter/in installed rules) and the.mdcrename.
Hooks, skills, commands, and subagents are unchanged — on-demand surfaces were not the problem.
[1.14.1] — 2026-07-29
Follow-up audit of every browser-touching file after the 1.14.0 migration. Two classes of leftover slipped through: the prose pass used single-line regexes, so any "Playwright MCP" that wrapped across a line break survived; and the token pass renamed browser_* verbs without being able to tell that a block of bare verb names was meant to be runnable. All 26 files that reference playwright-cli or the anti-stall protocol were re-read by hand.
Fixed
test-playwright— Phase 3, the skill's core per-step interaction loop, was still bare pseudo-code carrying the pre-migration "wait 2s" phrasing. Rewritten as a runnable$PW $S …sequence. A stale "hardgotoreload" also becamereload.test-qa/references/details.md— the equivalent execution-cycle block, same fix.audit-langfuse-llm— a leftover MCP-shaped```jsonblock (gotoplus a{ "url": … }object) is now a realopen --headedinvocation.audit-uiux-design-system— the skill description still advertised "browser MCP tools", line-split across the YAML fold. Description text feeds skill triggering, so this was stale metadata, not just prose.commands/test.md— "drive a visible browser through the Playwright / MCP" spanned a line break; now names playwright-cli with the-s=and--headedform inline.audit-ux/references/details.md— bare arrow-stub keyboard-navigation block converted to real commands.
Changed
enhance-web-ux— the three-viewport pass instructed "always with viewport set BEFORE navigate", written for the MCP. Verified against@playwright/cli@0.1.17thatresizepersists acrossgoto(a 1024×700 viewport survived navigation intact), so the ordering is sound; the stub is now a runnable loop over the three widths with a comparison table for what each one exposes.
Notes
Skills showing no $PW invocations are intentional, not gaps. Spokes such as enhance-motion and enhance-web-forms reference commands by name and defer to protocol-browser-anti-stall for the invocation form — only the hub carries the PW= boilerplate.
[1.14.0] — 2026-07-29
Browser automation moves off the Playwright MCP and onto playwright-cli. The MCP serves one browser per server, and a persistent Chrome profile can only be locked by one process at a time — so several agents on the same repo either contend for the profile lock or fight over each other's tabs. The CLI gives every agent its own isolated browser through -s=<session>, runs natively in parallel shells, and costs far fewer tokens (no tool schemas or verbose accessibility trees loaded into context). Verified end-to-end against @playwright/cli@0.1.17.
Changed
protocol-browser-anti-stall— rewritten aroundplaywright-cli. Mandatory-s=<session>naming,npx --yes @playwright/cli@latestinvocation (a globalnpm i -gis unreliable underfnm/nvm, whose global prefix is per-shell), and explicit--headedbecause the CLI is headless by default where the MCP was headed. Adds a session-lifecycle section (open/goto/close/list/close-all/kill-all), a parallel-agents section replacing the old tab-claiming etiquette, and artifact rules covering both.playwright-mcp/(screenshots via--filename) and the CLI's auto-written.playwright-cli/snapshots. All anti-stall guarantees are preserved: navigation guard, ≤3s waits, max 4 attempts per goal, evidence-before-retry, timeout budget, blocker format.- Waiting model — there is no
waitcommand in the CLI. Playwright auto-waits for actionability onclick/fill/select; explicit waits becomesleep N(≤3s) orrun-codewithwaitFor({ timeout }). Documented in the protocol and the command map. - 22 skills +
commands/test.md+ README migrated — 191 tool-token replacements plus 52 prose updates acrosstest-playwright,test-qa,test-red-team,audit-accessibility,audit-uiux-design-system,audit-ux,audit-langfuse-llm,audit-bundle-size,audit-i18n,audit-realworld,deploy-verify,iterate-post-launch,enhance-readme,enhance-web-ux,enhance-web-ui,enhance-web-forms,enhance-web-seo,enhance-motion,enhance-pwa,enhance-capacitor-ui,design-email,housekeep-design,plan-perf-audit,plan-stub-checker,plan-uiux-unification. The 19 structuralplaywright:browser_*JSON call sites (axe-core injection, keyboard walks, deploy smoke checks) were rewritten by hand into real shell invocations. mcp/README.md— the Playwright entry now documents the CLI as the supported path, keeping the MCP only as an--isolatedfallback (which loses saved logins). The pin inmcp/VERSIONS.mdis retained for that fallback..gitignore— added.playwright-cli/alongside.playwright-mcp/.
Added
references/mcp-to-cli-map.md— completebrowser_*→ CLI command map (actions, waiting, evidence, tabs, storage, session management), plus the concepts that no longer apply (browser_lock, tab claiming,session.json,--isolatedstorage-state injection) and the gotchas that bite in practice.- Google sign-in workaround —
references/playwright-session-coordination.mddocuments the one case that needs care: Google detects CDP and blocks sign-in from any Playwright-launched browser with "This browser or app may not be secure." The working sequence is a one-time manual login in real Chrome against a dedicated--user-data-dir(no--remote-debugging-port), then reuse viaopen --persistent --profile <dir>. Verified: Gmail loads straight into the inbox, and the session survivesclose→ reopen. scripts/migrate-browser-mcp-to-cli.mjs— the reviewed migration script, kept in-tree as an audit trail of how the rewrite was performed.
[1.13.0] — 2026-07-25
UX journeys audit — a new cross-page skill for the layer audit-ux (per-page heuristics) doesn't cover: can users actually find things and finish their stories? A site can pass every per-page heuristic and still fail because the feature is buried four clicks deep, the nav mirrors the DB schema, or the checkout dead-ends on mobile. Grounded in the 2026 practitioner consensus: a UX audit derives from user stories, audits IA against mental models, walks the real journeys, and grounds findings in behavioral evidence — never taste presented as data.
Added
audit-ux-journeys— cross-page user-story, task-completion & IA audit (Skill 110). Phase 0 derives 5–10 real stories from routes/nav/CTAs (reusesdesign-prd/plan-test-coverageinventories, never invented personas) and scopes depth by audit trigger (KPIs dropping / complaints / redesign / pre-launch-no-data → qualitative, labeled as assumptions). Phase 1 audits IA structurally (IA1–IA9): click depth per story target, orphan pages, dead ends, label consistency (nav ≈ title ≈ H1), grouping vs mental model, first-click logic, wayfinding, search/filter, URL sanity. Phase 2 walks every story end-to-end in a headed browser (desktop + 390px mobile) → task-completion matrix with friction logs (hesitation/mislabel/backtrack/surprise/stall) and error-recovery probes (invalid input, Back, refresh mid-flow); a BLOCKED story is automatically a Blocker. Phase 3 evidence discipline — every finding tagged[data]/[observed]/[judgment]; validates against GA4/PostHog/Clarity/Amplitude funnels when present, otherwise recommends minimal instrumentation so the next audit has evidence. Phase 4 impact×effort report: quick wins vs roadmap vs deprioritized, neutral language, always names what works. Detailed methods inskills/audit-ux-journeys/references/checklist.md.
Changed
audit-ux— description and Step 0c now scope it explicitly to the per-page experience lens and hand cross-page journeys/IA toaudit-ux-journeys(its 0c IA map is context-building, not a structural audit). CATALOG, TRIGGER-CHEATSHEET, README (family table 17→18, full list, Assess bullets, "Start here" row), andskill-workflows.mdcupdated; skill count 109 → 110.
[1.12.1] — 2026-07-24
Fix broken Tsumagoi image on the npm README (the private-repo raw GitHub URL returned 404). Point at the public live-site OG image and drop the private repo link.
Fixed
- Tsumagoi ad image —
srcnowhttps://tsumagoi.kensaur.us/og-image.png(public) instead ofraw.githubusercontent.com/...(404 on the private repo). - Also by table — Tsumagoi row links only to the live site; no GitHub repo link.
[1.12.0] — 2026-07-24
README discoverability overhaul — ELI5 intro, always-visible skill lists with family counts, npm-safe workflow diagram (no mermaid), and a playful Tsumagoi Work&Camp invite. No skill/command behavior changes; counts unchanged.
Added
- "Explain it like I'm five" opener with a use-case table (you say… → what kicks in) and an ASCII workflow loop that renders on npm (replaces the mermaid block that showed as raw code).
- Skill families at a glance table (counts + one-line blurb per family) plus a full always-visible skill table per family — no collapsed
<details>you have to click open. - Wanna code in the mountains? — playful ad for Tsumagoi Work&Camp (1,444 m coworking camp) with live-site link + screenshot.
Changed
- README section order: inventory + full skill list sit above the long Quick Start installer details.
scripts/generate-skill-index.mjsemits markdown tables instead of collapsed details; warmer family headings.- Workflows section simplified to a stage cheat-sheet that points at the top loop diagram.
[1.11.2] — 2026-07-24
Move the auto-generated skill index up so npm/GitHub visitors see every skill right after the inventory table — no 12-screen scroll past Workflows. No skill behavior changes.
Changed
- README section order —
Every skill, in plain Englishnow sits immediately afterWhat's Inside(beforeWorkflows), so the skill list is discoverable on the npm Readme tab without deep scrolling.
[1.11.1] — 2026-07-24
README rewrite — the npm/GitHub landing page now lists every skill with a one-line summary, and the prose is warmer throughout. No skill, command, rule, or hook behavior changes; counts unchanged.
Added
- Auto-generated skill index in README — new "Every skill, in plain English" section lists all 121 installable skills (109 agent + 12 Cursor) grouped by family, each with a one-line summary pulled from that skill's
SKILL.mdfrontmatter. Collapsed<details>groups keep the page scannable. scripts/generate-skill-index.mjs— regenerates the README block between<!-- SKILL-INDEX:START -->/ENDmarkers.npm run gen:skill-indexwrites;npm run check:skill-index(wired intonpm test) fails if the block drifts from the filesystem.
Changed
- README tone — benefit-first intro, shorter sentences, plain-English lead-ins before dense tables, and the giant Workflows "Assess" paragraph broken into scannable sub-bullets. FAQ grammar fix ("A installable" → "It's an installable").
[1.11.0] — 2026-07-24
Payment system audit — a new read-only skill for the code that fails differently from normal CRUD: a retried charge is a double-charge, a lost ledger write is vanished money, a logged PAN is PCI liability, an unverified webhook is a spoofed "payment succeeded". Grounded in the 2026 consensus (idempotency + double-entry ledger + reconciliation, with PCI DSS v4.0.1 as the floor and webhooks as the source of truth). Scope-gated so a Stripe-Checkout shop and an in-house gateway each see only relevant findings, and uses the Stripe MCP, when it is connected (not bundled), for version-anchored provider checks.
Added
audit-payment-system— read-only, scope-gated audit (Skill 109). Phase 0 detects the payment surfaces + provider and classifies the repo (P0 merchant integrator / P1 platform-marketplace / P2 gateway-PSP-fintech) so an in-house-ledger control isN/A(not "Missing") on a shop that offloads it to Stripe. Phase 2 marks each in-scope control Implemented/Partial/Missing/N-A withfile:lineacross seven groups: A money-movement correctness (idempotency on every mutation + DB unique constraint, dedup/payload-guard, payment state machine / no double-capture, integer minor units, multi-currency+FX), B ledger & data integrity (double-entry balanced/sum-zero, append-only immutable, derived balance snapshots, auditability, date partitioning), C async & webhooks (sync-auth vs async-everything, HMAC verify, event-id dedup + 200-then-process, atomic state+ledger+outbox, pull-based recovery for stuck payments, refund/dispute/payout saga), D reconciliation & settlement (daily 3-way match ledger↔PSP↔bank, break report, discrepancy handling, halt-on-unreconciled brake), E fraud/risk/SCA (velocity/geo/device + rules/ML score, 3DS2/SCA + exemptions, fraud-service fail policy, chargeback/VAMP monitoring, AML), F PCI DSS v4.0.1 (never store/log PAN or CVV, tokens-only scope reduction, key rotation, access audit), G resilience (PSP/bank timeout + breaker, bulkhead, partial-write safety, graceful degradation). Detailed per-control detection inskills/audit-payment-system/references/checklist.md.- Stripe MCP wiring (optional) — when the PSP is Stripe and the Stripe MCP is connected, Phase 1 uses
search_stripe_documentationfor concepts/best-practice andstripe_api_search+stripe_api_detailsfor exact API params, anchoring findings to the current API (PaymentIntents, not the legacy Charges API). Non-Stripe providers use/researchagainst official docs.
Changed
- No severity below Critical for a double-charge, lost-money, or PAN-exposure finding. Read-only: payment code is a STOP-and-confirm surface, so findings feed a human-reviewed remediation rather than an unattended edit. Delegates without overlap — per-call resilience →
audit-resilience, PCI/secrets →audit-security, ledger schema →audit-db-schema, outbox/saga structure →audit-backend-architecture. CATALOG, TRIGGER-CHEATSHEET, README, andskill-workflows.mdcupdated; skill count 108 → 109.
[1.10.0] — 2026-07-24
Decision lens for audit-backend-architecture — v1.9.0 answered "is pattern X present?" (conformance). This turns it into a decision advisor that answers "which pattern should this codebase adopt next, and which would be over-engineering right now?" — grounded in the 2026 "start simple, earn every pattern / modular-monolith-first" consensus. No new skill; the existing audit gets a second lens plus three patterns it was missing.
Added
- Fit/decision lens (Phase 3) in
audit-backend-architecture— for each pattern, classify Adopt now / Adopt when [trigger] / Defer (premature) from detected symptoms, with a maturity ladder (Stage 0 modular monolith → cache-aside/outbox → async/BFF/CQRS → db-per-service/saga → mesh/cell-based, each gated by a trigger) and a symptom→pattern decision table. Over-engineering is now a first-class finding (CQRS without read/write divergence, mesh/microservices below the size threshold, distributed monolith). - Core principle — "start simple, earn every pattern" — modular-monolith-first default; a pattern is only "Missing" when a measurable trigger justifies its cost; decide sync-vs-async per interaction; distributed monolith flagged as worse than a monolith.
- Three patterns added (audit rows #17–19 + detection in
references/patterns.md): communication style (sync request/response vs async event-driven, per interaction, with a signal table), cache-aside (TTL + invalidation + stampede guard), database-per-service / data ownership (+ distributed-monolith anti-pattern). backend-patterns/references/architecture-patterns.md— implementation for the three: sync-vs-async decision tree + hybrid/outbox example, cache-aside read+invalidate, and data-ownership (owned schemas at T1, split only under real pressure).
Changed
- Report template gains Adopt now / Adopt when / Defer sections (renumbered to Phase 4).
audit-backend-architecturedescription, CATALOG, TRIGGER-CHEATSHEET, README, andskill-workflows.mdcnow surface the decision lens ("which pattern should I use", "am I over-engineering", "sync vs event-driven"). Skill/command/agent counts unchanged (enhancement, not a new skill).
[1.9.0] — 2026-07-24
Backend architecture audit — a new read-only skill that checks a repo's distributed-systems architecture (not just per-call resilience), plus an implementation reference so the fixes it delegates actually exist. Answers "is my backend production-grade?" for the patterns agents systematically skip: API gateway, BFF, outbox, saga, bulkhead, hexagonal, service mesh, and more. Topology-gated so a Next.js/Supabase monolith and a Kubernetes fleet each see only relevant findings.
Added
audit-backend-architecture— read-only, topology-gated audit (Skill 108). Phase 0 classifies the repo into a tier (T1 serverless/monolith → T2 containers → T3 k8s/microservices/event-driven), then marks each in-scope pattern Implemented/Partial/Missing/N-A withfile:line: API gateway (auth, rate-limit, CORS, transform, logging, monitoring, caching), BFF / API composition / GraphQL federation, circuit breaker, bulkhead, backpressure/load-shedding, outbox + CDC (the dual-write fix), saga (orchestration/choreography + compensation + saga-pivot), CQRS + event sourcing, hexagonal / ports-and-adapters, anti-corruption layer, strangler-fig migration, sidecar / service mesh (incl. ambient/sidecarless), cell-based architecture, zero-trust/mTLS, distributed tracing + SLOs, and contract testing. Detailed per-pattern detection lives inskills/audit-backend-architecture/references/patterns.md.skills/backend-patterns/references/architecture-patterns.md— implementation guidance + code for the patterns the audit flags (gateway centralization, BFF, bulkhead, circuit-breaker placement, outbox+relay/CDC, saga with the saga-pivot rule, hexagonal/ports-and-adapters, anti-corruption layer, strangler-fig), plus a "when NOT to reach for these" section.
Changed
audit-resilienceboundary kept clean — the new architecture audit checks whether patterns exist (structural); per-call runtime tuning (timeouts, retry backoff+jitter, idempotency keys, cancellation) still belongs toaudit-resilience, which the new skill explicitly defers to instead of duplicating.backend-patterns— description gains triggers ("circuit breaker", "outbox pattern", "saga", "bulkhead", "hexagonal architecture", "API gateway", "BFF") and a new "Architecture patterns (distributed systems)" section pointing to the reference; pairs withaudit-backend-architecture.- Routing/docs — added to
skill-workflows.mdc(repo +.cursor),docs/CATALOG.md,docs/TRIGGER-CHEATSHEET.md, andREADME.md(Assess stage, use-case table). Counts synced to 108 skills.
[1.8.3] — 2026-07-23
Command-collision fix — three of our slash commands shared names with Claude Code's own commands. Because Claude Code 2.x merges commands/ into the skill namespace, a file at commands/<x>.md creates /<x>: when <x> matches a built-in it shows a duplicate, and when it matches a bundled skill our file silently overrides Claude's. Renamed all three and added a CI guard so it can't regress. No other skill, command, rule, or hook behavior changes; command counts are unchanged.
Changed
/mcp→/mcp-guide— was duplicating Claude Code's built-in/mcp(the MCP server manager). Renamedcommands/mcp.md→commands/mcp-guide.md./review→/review-code— was duplicating Claude Code's built-in/review. Renamedcommands/review.md→commands/review-code.md(still delegates toaudit-code-review)./debug→/debug-issue— was overriding Claude Code's bundled/debug. Renamedcommands/debug.md→commands/debug-issue.md(still delegates todebug-error).README.md+docs/CATALOG.md— command tables updated to the new names with a note on why each was renamed.
Added
- CI guard in
scripts/validate-skills.mjs— failsnpm testif anycommands/*.mdname collides with a known Claude Code built-in command or bundled skill, so future commands can't reintroduce a duplicate/override.
Upgrade note: the natural-language triggers are unchanged (say "debug this", "review my PR", etc. and the underlying skill still fires). Only the explicit slash aliases changed: use
/mcp-guide,/review-code,/debug-issue. After updating, restart Claude Code — it live-watches skill directories but notcommands/, so newly added or renamed command files only register on the next session start.
[1.8.2] — 2026-07-23
Documentation-skill release — teach the skills to write the way the 1.8.1 README now reads. Bakes the reader-first, plain-language 5W1H approach into the documentation-authoring skills themselves, so every doc they produce orients a newcomer before diving into reference detail. No other skill, command, rule, or hook behavior changes.
Changed
docs-writer— new "Core principle — write for the reader's mental model first" section with a 5W1H table (what/why/who/how/when·where → how to answer each) and the rules that follow (lead with the goal not the implementation, progressive disclosure, beat the curse of knowledge, show-don't-tell). Adds a "Newcomer on-ramp" section demonstrating the 3-column plain-language building-blocks glossary, a "Kill the Jargon" guideline (plain-first, precise-second), an upgraded README template (plain tagline + why/who + glossary hint), and two new README-checklist items (first-screen what/why/who; jargon defined or glossed).enhance-readme— previously visual-only; adds "The words matter as much as the pixels" coupling the hero/screenshot work todocs-writer's principle, defines the ideal first-screen order (name+tagline → hero → why/who → first-win → depth), adds why/who to the hero template, and adds a stop-condition that the first screen must answer what/why/who in plain language.docs-coauthor— adds a guiding-principle line: orient the reader (what/why/who) from the first sentence, not just at the Reader-Testing stage; points to the canonicaldocs-writerrule.
Kept DRY: the full principle lives in docs-writer; the other two reference it. plan-docs-sync (a drift audit, not an authoring skill) intentionally unchanged.
[1.8.1] — 2026-07-23
Documentation-clarity release — same skills, friendlier front door. Rewrites the released-page surfaces (README on npm/GitHub, llms.txt for AI crawlers, the Cursor plugin manifest, directory listing copy) so a non-technical newcomer can answer what / who / why / how before hitting the counts and taxonomy. No skill, command, rule, or hook behavior changes.
Changed
- README newcomer on-ramp — adds a plain-English hero line, a "New to AI skills?" primer, a "Who it's for" callout, and a five-row building-blocks glossary (Skill / Command / Subagent / Rule / MCP server → plain meaning + how you trigger each) directly under the intro, while preserving the full developer depth below.
llms.txt— leads with the plain value proposition (ready-made playbooks your editor auto-triggers) before the inventory, improving how AI answer-engines and directory crawlers summarize the project..cursor-plugin/plugin.json— benefit-first description for the Cursor marketplace / directory listings.docs/CATALOG.md— skill-taxonomy table gains a lifecycle Stage column and aplan-/housekeep-row, with a newcomer pointer to the Getting Started guide.docs/GETTING-STARTED.md— plain-language example rows for the v1.8.0 skills (motion, forms, design consolidation, guardrails, resilience).docs/TRIGGER-CHEATSHEET.md&docs/PROMOTION.md— first-time-reader pointer and benefit-first listing copy.
[1.8.0] — 2026-07-23
Design, motion, forms, and anti-vibe-coding release. Adds six skills that fill the loop's remaining gaps — full-stack parity (audit-realworld), production resilience (audit-resilience), coherent motion on an existing system (enhance-motion), design-system consolidation to one SSOT (housekeep-design), production-quality forms (enhance-web-forms), and guardrails-as-code against AI regressions (enhance-agent-guardrails). README gains a "mental model" section teaching how the families chain against each other, a by-use-case entry table, and lifecycle-stage tags. Skills 100 → 106.
Added
audit-realworldskill — read-only full-stack gap audit that benchmarks a repo against the RealWorld ("Conduit") reference: its formal API spec, shared Bruno/Hurl E2E contract suite, and closest-stack reference implementation. Auto-detects strict spec Conformance (repo is a RealWorld build) vs Benchmark (maps RealWorld's production-relevant patterns onto the repo's own domain), and bows out cleanly on non-CRUD/non-web repos. Reports Implemented/Partial/Missing/Diverges across frontend, backend, and data layer withfile:lineevidence, then delegates real production hardening toaudit-security/plan-security-audit/plan-perf-audit/plan-rls-audit/full-stack-ship-discipline— RealWorld is treated as a completeness/pattern reference, never a production bar. Wired into the Audits routing table,docs/CATALOG.md, anddocs/TRIGGER-CHEATSHEET.md. Skills 100 → 101.enhance-motionskill — audits an existing app's design system and current motion, then applies coherent,prefers-reduced-motion-safe, 60fps motion using a right-sized 2026 stack (tw-animate-css for CSS utility transitions, Auto-Animate for zero-config list/layout changes, Motion/motion.dev for component transitions·gestures·presence, GSAP only for complex timelines), with Kinetics as a pattern source. Defines a motion-token SSOT so timing/easing stops fragmenting. Distinct fromdesign-motion(from-scratch cookbook); applies changes and verifies via browser MCP.housekeep-designskill — consolidates a design system that has drifted across many vibe-coding sessions and dev handoffs into one source of truth. Detects competing tokens, duplicate components, naming drift, mixed icon libraries, and arbitrary values; reconciles each conflict to a best-of-both canonical form; migrates usages via mechanical codemod (pairs withburndown-full); and installs lint guardrails so drift can't recur. Establishes a 3-layer W3C token taxonomy (primitive → semantic → component). The execution arm ofplan-uiux-unification; the design counterpart ofworkflow-housekeep.enhance-web-formsskill — builds/upgrades web forms to production quality: accessible structure (labels, fieldsets, autocomplete, input types), schema-driven validation with client↔server parity, screen-reader-associated inline errors, complete interaction states (loading/disabled/success/error/empty), multi-step flows, unsaved-changes guards, and optimistic submit feedback. Auto-detects the form + validation stack; applies changes and verifies via browser MCP.enhance-agent-guardrailsskill — installs guardrails-as-code so AI/vibe-coding can't keep reintroducing the same problem classes (leaked secrets, injection, off-system styles, untested code, vulnerable deps, destructive ops). Audits existing protection, then sets up agent policy files (.cursor/rules+AGENTS.md), a pre-commit hook (secret scan + SAST + lint/typecheck), a CI gate that treats agent output as untrusted, and lint-as-policy rules. Verifies the guards block a planted bad pattern; flags merge-blocking CI changes for human review. Grounded in 2026 research on AI-generated-code security debt.audit-resilienceskill — read-only audit for the non-functional "20%" agents systematically skip (the "80% problem"): timeouts, retries with backoff+jitter, circuit breakers, idempotency keys, rate limiting, graceful degradation, cancellation, audit logging, and PII handling in logs. Inventories every external call, mutation, webhook, and payment path and marks each concern Implemented/Partial/Missing withfile:line, severity, and the exact fix skill; delegates remediation tobackend-error-handling/backend-patterns/backend-observability/plan-*.
All six new skills wired into both skill-workflows.mdc routing copies, docs/CATALOG.md, and docs/TRIGGER-CHEATSHEET.md. Skills 100 → 106.
Changed
- README mental-model rewrite — the Workflows section now opens with a "how the families work against each other" lifecycle diagram (Orient → Assess → Change → Prove → Ship & operate, with always-on Guardrails), a "Start here — by use case" entry table that chains the new skills, and lifecycle-stage tags added to the skill-taxonomy table (Assess / Change / Prove / Ship / Operate / Guardrail). Teaches composition, not just the flat skill list.
[1.7.0] — 2026-07-21
Completion-loop release — close the gaps where agents stop early, declare false "done", or leave documented capabilities unwired. Adds an enforcement layer (evidence rule + independent judge + opt-in continuation hook), five closure/ops workflow skills, and repairs every known broken internal reference. Skills 95 → 100, commands 16 → 36, subagents held at 6.
Added
complete-everythingskill +/complete-everythingcommand — close unfinished plan intent and connected work previously parked as out of scope/follow-up/optional, persist observable acceptance evidence in.cursor/complete-everything-state.md, and loop the applicable typecheck/lint/test/API/build/browser gates until completecompletion-judgesubagent — independently reconciles the approved outcome, durable state, full diff, baseline, and fresh evidence before returningPASS,CONTINUE, orBLOCKEDverification-before-completionrule — defines evidence-backed completion levels from implemented through observed-stable and forbids claims above the proven level- Opt-in Cursor completion hook — continues only actionable unchecked
complete-everything/burndown-fullstate; safely merges into existing user hooks and remains inert for completed, errored, or human-gate-only runs - Closure-loop workflow skills + commands —
workflow-green-repo(/green-repo) drives the whole repository to a verified-green baseline;workflow-ship-and-observe(/ship-and-observe) takes merged code to a monitored production release with a live-revision check and rollback tail;workflow-feedback-to-closure(/feedback-to-closure) turns raw feedback into deduplicated durable tickets driven to production-verified closure workflow-environment-readyskill — preflights the toolchain, services, and config before a long autonomous run so it does not fail at the finish line on a missing dependency or credentialiterate-agent-harnessskill — converts an agent failure (premature stop, false done, reward-hacked check, broken handoff) into a durable harness guard plus a regression check that fails before the fix- 17
plan-*pointer commands —/uiux-plan,/security-plan,/rls-plan,/secrets-plan,/validation-plan,/integrity-plan,/error-plan,/deps-plan,/cost-plan,/aeo-plan,/mobile-plan,/capacitor-plan,/stub-plan,/perf-plan,/docs-plan,/test-plan,/slop-plan— thin audit-and-plan entries the docs previously referenced but shipped no file for
Changed
- Composer execution precedence — normal plan runs still stop at phase boundaries; explicit closure modes persist the checkpoint and continue. Approved non-destructive schema work on a confirmed target now follows full-stack deployment discipline without a contradictory permission stop.
- Count synchronization — skill, command, and subagent totals are all derived from the filesystem.
Fixed
- Broken internal skill references — repaired a doubled
workflow-spec-workflow-spec-tddlink in four skills, a wrongreferences/path inworkflow-spec-tdd, deadmushi-debuglinks inmushi-health/mushi-integration(now the mushi MCPdiagnose_setuptool), and non-existent../start-emulator/../test-emulatorlinks inmobile-rn-screen. A from-scratch relative-link scan now resolves clean.
[1.6.0] — 2026-07-18
Universal installer — cursor-kenji now installs to every supported AI coding tool, mapping to what each tool actually loads. Skills-capable tools (Cursor, Claude Code) get the full skills + commands + agents + rules set; context-file tools (Codex CLI, Gemini CLI) get the rules merged into their global instructions file plus portable command ports. Skill count held at 94.
Added
--autodetection — probes~/.cursor,~/.claude,~/.codex,~/.geminiand installs the right artifacts to each installed tool (falls back to Cursor if none are found).- Codex CLI support (
--codex) — mergesrules/into~/.codex/AGENTS.md(auto-loaded) and portsplan/research/fix-issueto~/.codex/prompts/*.md. - Gemini CLI support (
--gemini) — mergesrules/into~/.gemini/GEMINI.md(auto-loaded) and ports the same commands to~/.gemini/commands/*.toml. commands-portable/— new single source of truth for the three tool-agnostic playbooks the installer transforms per tool.- Install smoke test now exercises the Codex/Gemini paths (merged-rules content, command ports, TOML shape, idempotency,
--autodetection).
Changed
--allnow targets all four supported tools (was Cursor + Claude Code). A barenpx @kensaurus/cursor-kenjistays Cursor-only for backward compatibility.- The rules merge deliberately omits the skill-routing index (
skill-workflows.mdc) — the context-file tools have no skills loader, so it would be dead text. Skills and subagents are not written to Codex/Gemini for the same reason. - Generated context files are idempotent (byte-identical on re-run) and backed up (
.bak-<stamp>) before any overwrite. install.shgains--auto/--codex/--gemini/--all; Codex/Gemini delegate to the Node installer so the merge/port logic has one source of truth.
[1.5.1] — 2026-07-18
Prompt-engineering quality pass from docs/PLAN-SKILL-PACK-ENHANCEMENT.md (Phases 3–5). No new skills; skill count held at 94. Dual-runtime (Cursor + Claude Code) portability, staleness, and licensing hygiene.
Changed
- Dual-runtime portability — skills now refer to each other by bare name (no path assumptions), and MCP calls use portable
ServerName:tool_nameprose instead of runtime-specific fenced tool blocks. - Command frontmatter — added
description+argument-hintto 14 slash commands for discoverability. - Progressive disclosure — relocated bulk inline content to
references/inworkflow-housekeep,backend-realtime, anddeploy-npm(references/example-mushi-mushi.md); trimmed oversizedplan-*descriptions and rewrote thin ones to the house standard. - Descriptive prose over pseudo-tools —
audit-uxandaudit-securityreframeGlob/Grepfenced blocks (which read as executable calls) as plain guidance.
Fixed
- Stale model IDs — refreshed illustrative model references in
audit-langfuse-llmto current-generation IDs. - Hardcoded counts — decoupled the style-catalog count from the always-on
thirdparty-ui-ux-pro-maxdescription. - License attribution — pinned verified upstream MIT licenses (with copyright holders) in
thirdparty-emil-design-engandthirdparty-web-interface-guidelinesATTRIBUTION.md, replacing "See upstream repository" placeholders.
[1.5.0] — 2026-07-18
Added
- npm installer: Claude Code target —
npx @kensaurus/cursor-kenji --claude(Claude Code only) and--all(Cursor + Claude Code) install skills, commands, agents, and rules to~/.claude/with.mdcrules converted to.md. One-click on any OS, no clone needed. Newnpm run install:claude/install:allscripts. - install.sh: Claude Code commands —
~/.claude/commands/now installed alongside skills/agents/rules. docs/PLAN-SKILL-PACK-ENHANCEMENT.md— full prompt-engineering audit of all 94 skills + 15 commands with a phased enhancement burndown (token usage, progressive disclosure, dual-runtime portability, description standards).- Third-party UI skills (Jun 2026) —
thirdparty-emil-design-eng,thirdparty-ui-ux-pro-max,thirdparty-web-interface-guidelineswithATTRIBUTION.md,docs/THIRD-PARTY-SKILLS.md, and/thirdparty-web-interface-guidelinescommand burndown-fullskill +/burndown-fullcommand — drive partially-executed refactors/migrations to 100% repo coverage via MATCH/DONE patterns, persistent.cursor/burndown-state.md, batched execution, and a verification gate (framework-agnostic; works in Cursor and Claude Code)
Fixed
- Sentry MCP calls —
naturalLanguageQuery→queryacross 15 skill files (matches current Sentry MCPsearch_issues/search_eventssignature). mobile-emulator-start— display default corrected to1080×2400;1080×4000is now opt-in for scroll-QA with a matching tall skin.test-red-team— description trimmed under the Agent Skills 1024-char frontmatter limit.design-mobile-first— repaired garbled self-referential description.enhance-web-landing— H1 renamed to match the skill (internal codename removed).- Privacy/staleness — private project names removed from
test-qaanddata-pipeline; unverifiable CVE numbers and named-incident claims replaced with hedged failure-class descriptions inplan-input-validation,plan-rls-audit,plan-data-integrity,plan-llm-cost-guardrails,plan-secrets-audit. - install.sh — removed duplicated "Installed N skills" log line.
[1.4.2] — 2026-06-24
Fixed
- npm tarball —
.mcp.jsonincluded in published package (Cursor Marketplace MCP reference) - MCP essential 5 — aligned
mcp/mcp.json.template,.mcp.json, and docs on Playwright (was Chrome DevTools in template)
Changed
- Anti-slop pass — skill YAML descriptions, subagent frontmatter, README badges, CATALOG/PROMOTION/GETTING-STARTED copy
docs/DISTRIBUTION.md— install channel matrix (Cursor vs Claude Code vs Marketplace)- README — Claude Code labeled bash-only; shell helpers table expanded; project rules count clarified
package.json— removedpreparegit-hooks side effect on consumernpm install.gitignore—*.log,*.bak,*.old,*.tmp
[1.4.1] — 2026-06-23
Fixed
package.jsonbin —cursor-kenji→bin/install.mjs(npm publish no longer strips the bin entry)repository.url— normalized togit+https://github.com/kensaurus/cursor-kenji.git
Changed
- npm OIDC Trusted Publisher configured (
kensaurus/cursor-kenji· workflownpm-publish.yml) - CI publish — OIDC-first;
NPM_TOKENonly used when the secret is set
[1.4.0] — 2026-06-23
Added
llms.txt— machine-readable doc index for answer-engine / AI crawler discoverabilitySECURITY.md— MCP secrets hygiene, OIDC-first publish guidance, zero-dep policyscripts/scan-secrets.mjs— pre-commit secret pattern scanner (zero-deps)scripts/check-mcp-pins.mjs+mcp/pinned-versions.json+mcp/VERSIONS.md— semver-pinned MCP templates- README FAQ — install, skill count, audit vs plan, MCP keys, llms.txt
Changed
- MCP templates — all
@latestremoved; npm + uvx packages pinned; AWS servers migrated to valid PyPI names (lambda-tool,aws-api, unifiedcloudwatch-mcp) - Pre-commit — runs secret scan before skill validation
npm test— includes secret self-test + MCP pin checkdocs/PROMOTION.md— OIDC-first npm publish; launch copy skill count 90
[1.3.0] — 2026-06-23
Added
- Eleven new
plan-*skills — security spine, launch gates, antislop, Capacitor hardening, observability/spend (plan-rls-auditthroughplan-capacitor-hardening) docs/AGENTS.template.md— project constitution (mission / stack / roadmap / agent discipline)docs/examples/plan-audits/— sample plan burndown outputs; root/plan-*.mdgitignored
Changed
- README de-duplicated (~340 lines); CATALOG canonical; skill count 90 (17 plan-*)
check-skill-count.mjssyncs README, CATALOG, PROMOTION, package.json- Pre-commit hook re-stages all count-sync files on drift
create-skillbody split — detail inreferences/authoring-guide.md(<500 lines)
[1.2.7] — 2026-06-17
Added
composer-2.5-execution.mdcrule (alwaysApply: true) — execution-time guardrails for Composer 2.5 when implementing an approvedplan-*.md: anti-reward-hacking, anti-feature-deletion, checkpointing, context + terminal discipline, STOP-and-ask on auth/RLS/secrets/payments/migrations- Documents the two-model workflow (strong model plans, Composer 2.5 executes) in README and
docs/PLAN-LOOPS.md
Changed
- Project rules count: 8 → 9
[1.2.6] — 2026-06-17
Added
plan-test-coverage— user-story-driven coverage audit, traceability matrix, fake-green detection (plan only)docs/PLAN-LOOPS.md— six-skill plan loop: link, chain, prompts, execution mapping
Changed
- Six-skill plan loop (was five): adds
plan-test-coverageafter stub-checker, before perf/security
[1.2.5] — 2026-06-17
Added
plan-docs-sync— docs drift audit vs code truth; onboarding-drift checks; docs-as-code guardrails (plan only)plan-perf-audit— measure-don't-guess perf audit across web/mobile/backend/data; Lighthouse CI + RUM guardrails (plan only)plan-security-audit— OWASP + Supabase-first (RLS, service_role scan); no destructive testing; secrets location-only (plan only)- Five-skill plan loop documented in
skill-workflowsand CATALOG
[1.2.4] — 2026-06-17
Added
plan-stub-checker— exhaustive stub/dead-button/fake-component/unwired-handler audit with burndown and phased wiring plan (plan only; no implementation until user approves)
[1.2.3] — 2026-06-16
Added
plan-uiux-unification— exhaustive non-destructive UI/UX + design-system audit that outputs burndown, unification plan, and phased roadmap (plan only; no fixes until user approves)
[1.2.2] — 2026-06-16
Changed
- Playwright test skills: shared signed-in session reuse (Google OAuth, etc.) and multi-agent tab discipline — no fighting over the same browser tab
- New reference:
protocol-browser-anti-stall/references/playwright-session-coordination.md - Updated:
test-playwright,test-qa,test-red-team,protocol-browser-anti-stall
[1.2.1] — 2026-06-12
Fixed
- npm CI publish: use OIDC trusted publishing (remove empty
NPM_TOKENthat blocked auth)
[1.2.0] — 2026-06-12
Added (13 skills since v1.1.0)
- Red team & quality:
test-red-team,workflow-quality-gate - Post-launch & shipping:
iterate-post-launch,workflow-launch-ready,workflow-feature-flag - Enhance & audit:
enhance-web-seo,enhance-pwa,audit-bundle-size,audit-i18n,design-email - Bundled workflows:
workflow-build-feature,workflow-fix-and-ship,workflow-onboard - Routing rule:
rules/skill-workflows.mdc— capabilities catalog with skill-chaining diagram
Changed
- Skill count: 60 → 73
test-red-teamtrimmed to stay under 500-line spec limit.gitignore: ignore.playwright-mcp/and.qa-screenshots/scratch artifacts
Install
npx skills add kensaurus/cursor-kenji
# or
npx @kensaurus/cursor-kenji@1.2.0
Jun 2026 — Bundled Workflows, Gap-Filling, & Verbosity Trim
New Bundled Workflow Skills
Orchestrator skills that chain multiple individual skills into a single tracked loop. Inspired by community patterns from darthlinuxer/Agentic-Skills, kscius/KS-Cursor-Orchestrator, and the Anthropic agent skills ecosystem.
| Skill | What it chains |
|---|---|
workflow-build-feature | spec → TDD → unit tests → smoke test → PR |
workflow-fix-and-ship | debug → root cause → fix → regression test → smoke → PR → deploy |
workflow-quality-gate | red-team → security → bundle → perf → unit tests → go/no-go verdict |
workflow-launch-ready | SEO + PWA + bundle + i18n + quality gate + deploy smoke + iterate |
workflow-onboard | First-contact orientation — reads codebase, produces 5-minute briefing |
New .cursor/rules/skill-workflows.mdc
Always-on routing index that routes intent to bundled workflows vs individual skills, with a full skill-chaining diagram. Reduces the cognitive load of "which skill do I use?" to a single lookup.
Improvements
test-red-teamtrimmed from 515 → 417 lines: security priority table replaced verbose prose sections; Playwright quick-ref removed (already in platform docs).
Jun 2026 — Post-Launch Iteration & Shipping-Quality Skills
New Skills
| Skill | Why |
|---|---|
iterate-post-launch | Closes the post-ship improvement loop — pulls Sentry errors, Supabase slow-query advisors, and live Playwright walkthrough into a ranked backlog, implements fixes, and verifies them against the live app. |
enhance-web-seo | Audits and fixes SEO for any web app: meta tags, OG/Twitter cards, JSON-LD structured data, robots.txt, sitemap, canonical URLs, heading hierarchy, Core Web Vitals. Applies fixes and verifies with Playwright. |
audit-bundle-size | Finds and eliminates JS bundle bloat: detects bundler (Vite/Webpack/Next.js), runs production build with analysis, identifies large chunks/duplicate deps/missing lazy splits, and maps every finding to a specific import. |
enhance-pwa | Adds or upgrades PWA features: Web App Manifest, Workbox service worker with per-asset caching strategies, install prompt, push notifications, offline page. Capacitor-compatible. Lighthouse PWA audit before/after. |
workflow-feature-flag | Feature-flag rollout discipline: design flag contract, gate the feature, stage at 0%→internal→5%→100%, monitor Sentry error rate and Supabase logs at each stage, promote or roll back, then clean up the flag from code. |
audit-i18n | i18n audit with emphasis on human-readable, natural-sounding copy — not machine-translated jargon. Finds hardcoded strings, checks translation completeness, rewrites stiff/literal copy to sound like a real person in each locale, fixes date/number/currency formatting. |
design-email | Full-stack transactional email: React Email templates with dark mode, mobile-first layout, inline styles, and copy that sounds like a person wrote it. Covers SPF/DKIM/DMARC deliverability, Resend/SendGrid/SES integration, and Supabase Edge Function triggers. |
Jun 2026 — Adversarial Red-Team Skill
New Skills
| Skill | Why |
|---|---|
test-red-team | Adversarial full-app sweep — feature-first coverage matrix (feature → surfaces → components+states) attacked across 4 dimensions: UI/UX, data pipeline, security (OWASP Top 10 + MASVS), and performance. Drives Playwright browser MCP for web/PWA, Playwright Android WebView attach for Capacitor hybrid apps, and adb tap-walk for native chrome. Cross-references Sentry for production telemetry, Supabase for DB-layer truth and RLS verification, and Firecrawl for current OWASP/MASVS guidance. Produces a severity-ranked defect list with repro steps, evidence, and launch-readiness verdict. Ships with a references/owasp-attack-checklist.md payload library covering XSS, SQLi, IDOR, MASVS-PLATFORM, auth attacks, and 2026 performance thresholds. |
Jun 2026 — Installer hardening + spec compliance
Installer (bin/install.mjs)
- Fixed a shipped bug: the installer only copied subdirectories, silently
dropping every top-level
.md/.mdcfile —npxinstalls landed 0 subagents, only 1 of 14 commands, and no rules files. It now copies files and directories. --clean/--mirror— make~/.cursorexactly mirror this repo (no overlap/duplicates); takes a timestamped backup first (--no-backupto skip).--restore [stamp]— restore a previous--cleanbackup.--only <csv>and--skill <name>— partial installs.--link— dev mode: symlink (junction on Windows) instead of copy.- MCP template is written only when missing and never overwrites existing keys.
Quality gates
scripts/validate-skills.mjs— validates every skill against the Agent Skills spec: frontmatter present,namematches its directory,description≤ 1024 chars, body length warnings.scripts/test-install.mjs— install smoke test (would have caught the bug above).- Both run in the pre-commit hook, a new cross-OS
validateworkflow, and beforenpm publish. - Trimmed 4 over-long skill descriptions to ≤ 1024 chars (
audit-ux,enhance-web-ui,enhance-web-ux,enhance-capacitor-ui). - Added
.gitattributesto normalize line endings (LF) across machines.
Jun 2026 — Anti-Vibe-Coding Spine + Taste Skills
New Skills
| Skill | Why |
|---|---|
workflow-spec-workflow-spec-tdd | The anti-vibe-coding spine — brainstorm → spec → plan → RED/GREEN/REFACTOR TDD → self-review before "done". Adapted from obra/superpowers |
mobile-mobile-capacitor-platform | Capacitor platform + pipeline depth — plugins, OTA, deep links, push, native CI, store submission + Apple preflight, security scan, migrations. From cap-go/capgo-skills |
mobile-mobile-rn-performance | React Native perf/build/upgrade depth — FPS, Hermes, TTI, bundle size, FlashList, Reanimated, Turbo Modules, 16KB alignment, RN/Expo version upgrades. From callstack/agent-skills |
data-pipeline | Build-time data-pipeline correctness — idempotency, atomic writes, 4-layer staging, windowed backfills, dead-letter, observability. For ETL / edge-function workers / pg_cron / queues |
backend-observability | Build-time observability + logging — error↔trace↔log correlation, structured leveled logs, PII/secret redaction, OTel spans, LLM trace capture, alert/SLO design. Vendor-neutral (Sentry + Langfuse + OTel) |
enhance-web-landing | Anti-slop frontend for landing pages, portfolios, and marketing sites — brief inference, variance/motion/density dials. From Leonxlnx/taste-skill |
enhance-web-redesign | Audit-first upgrade of existing sites — 60-second AI-tell triage, then scan/diagnose/fix generic AI patterns without breaking functionality |
enhance-web-web3d | Audit-first 3D/WebGL + GSAP cinematic-motion elevation of an existing web app (Three.js / R3F + ScrollTrigger + Motion / React Spring). From freshtechbro/claudedesignskills |
Family Rename
enhance-page-ui → enhance-web-ui, enhance-page-ux → enhance-web-ux, enhance-screen-rn → mobile-rn-screen, enhance-web-mobile-ui → enhance-capacitor-ui. New enhance-<surface>-<aspect> convention + a surface-router block added to every enhance skill.
May 2026 — PDCA Browser Testing, npm Release, Native RN Loop
New Skills
| Skill | Why |
|---|---|
test-playwright | Closes the PDCA loop after you ship a change. Drives the live localhost app through the Playwright MCP manually like a real user, and fixes pain points/errors as it goes (full-stack: UI/UX + API + DB). Red-teams the work and suggests enhancements |
deploy-npm | End-to-end release workflow for a Changesets + GitHub Actions + npm Trusted Publisher (OIDC) monorepo — green CI → merge release PR → resolve the github-actions[bot] anti-loop → dispatch publish → verify on npm and GitHub Releases |
mobile-emulator-start | Boots Metro + Android emulator in the correct order — kills duplicate ports, picks fresh-cache vs fast-iteration, defaults to 1080×4000 for tall QA screenshots, polls /status before deeplink to avoid "Cannot connect to Expo CLI" races |
enhance-capacitor-ui | Cross-surface architecture for hybrid PWA + iOS + Android apps. Three orthogonal axes (form factor / platform / pointer) so polish on one surface can't degrade another |
New Rules
| Rule | Why |
|---|---|
full-stack-ship-discipline.mdc | alwaysApply: true — prevents the "local migration file never deployed" failure mode. Forces UI tasks to inventory backend deps in the same chat, deploy via Supabase MCP, and verify against the remote DB |
shell-first-search.md | Workspace-wide rule routing routine search to Shell (grep/find/ls) instead of Grep/Glob tools, which can hang for minutes on some Windows hosts |
Housekeep Pass
- 11 skills renamed so
name:frontmatter matches folder names per Cursor spec - 22 descriptions tightened to direct-tone single-sentence WHAT + concrete trigger list
- 9 commands demoted to skill pointers (no more duplicated playbook)
Apr 2026 — The Enhance Family
| Skill | Why |
|---|---|
enhance-web-ui | Composition over decoration — fix hierarchy, grouping, spacing, motion. NN/g + Laws of UX grounded |
enhance-web-ux | Replace generic "stacked" UI with semantic data. Every change cites a Nielsen heuristic, uses existing primitives, verified at 3 viewports via browser MCP |
enhance-readme | Theme-aware hero + tour grid + optional animated GIF via Playwright MCP |
audit-ux | Deep UX audit grounded in NN/g 10 heuristics, Laws of UX, Intuit Content Design, Google HEART |
split-to-prs (Cursor Skill) | Slice a single chat / branch / PR into small reviewable PRs with safe snapshot |
canvas (Cursor Skill) | Live React canvas beside chat — updated with refreshed SDK primitives |
Earlier 2026
| Addition | Type |
|---|---|
create-hook | Skill |
workflow-spec-workflow-spec-tdd | Skill |
workflow-spec-tdd | Skill |
workflow-workflow-parallel-agents | Skill |
audit-code-review | Skill |
| 20 new skills (audits, debugging, deploy verification, file handling, PRD, QA, housekeeping) | Skills |
| 6 new cursor-skills (babysit, canvas, create-hook, shell, statusline, update-cli-config) | Cursor Skills |
/plan, /pr, /debug | Commands |