kenji skills
You say the job. The playbook runs. Agent skills, slash commands, and subagents for Claude Code, Cursor, Codex CLI, Gemini CLI, and any agent that reads SKILL.md.
162 agent skills · 63 slash commands · 6 subagents · MCP templates. Written on React, Next.js, and Supabase projects; most of it works on any stack.
After you install, describe the job in chat. The matching playbook runs, checks its own result, and opens the PR.
Quickstart
You need Node.js 18 or newer. Run these three commands:
npx @kensaurus/skills --auto --dry-run # 1. preview: lists every file it would write, changes nothing
npx @kensaurus/skills --auto # 2. install for each agent found on this machine
npx @kensaurus/skills --auto --verify # 3. check every installed file against the package
--auto looks for ~/.cursor, ~/.claude, ~/.codex, and ~/.gemini and installs for each one it finds. If it finds none, it installs for Cursor. To pick the tools yourself, use --claude, --cursor, --codex, --gemini (combine freely) or --all.
Then restart your agent. Stop there; you are set up.
Your first session
You do not need to learn the skill names. Remember one command:
/kenji the login page is broken for some users
/kenji reads the job in your own words, names the one playbook that fits, and runs it. If two fit, it asks you one question first.
You can also skip /kenji and describe the task, for example "audit this repo for security problems". The agent picks the playbook from your words. Phrases that work are in What should I say?, and the plain-language guide walks through a full session.
Install a profile
The full pack is 162 skills. Your agent reads a one-line description of each skill in every session, and a longer list makes it more likely to pick a near-miss. If you want less, install a profile:
npx @kensaurus/skills --list-profiles # show the profiles
npx @kensaurus/skills --auto --profile core # the 31 skills of the daily loop
npx @kensaurus/skills --auto --profile web,mobile # core plus two more sets
| Profile | Skills | What it covers |
|---|---|---|
core | 31 | Orient, grill, build, fix, review, test, ship. Part of every profile. |
web | 44 | Design, polish, and audit web front ends |
backend | 35 | APIs, databases, auth, resilience, LLM features |
mobile | 16 | React Native, Capacitor, emulators, store readiness |
growth | 22 | Go to market, onboarding, SEO, pricing, launch copy |
maintain | 28 | CI gates, dead code, docs sync, agent setup, pack authoring |
all | 162 | Everything. The default. |
A profile install is a merge: skills you installed earlier stay. Add --prune to remove this pack's skills that are outside the profile. It removes by folder name and first copies each removed folder to .kenji-pruned/<timestamp>/ beside the skills folder, so an edit you made in place is not lost. Folders this pack does not ship are never touched. Commands that only point at skills outside the profile are skipped, and --prune removes the ones an earlier install left. The lists live in profiles.json.
Note: the default install also adds an opt-in Stop hook to Claude Code and Cursor. It does nothing until you start a
complete-everythingorburndown-fullrun. For Claude Code it also adds a desktop Notification hook, a status line (only if you have none) and a skill-listing budget. To skip the hooks and the status line, add--only skills,commands,agents,rules. Details: The Stop hook and Notifications and the status line.
Which agents are supported
| Agent | Install with | Skills | Slash commands | Subagents | Rules | Stop hook |
|---|---|---|---|---|---|---|
| Claude Code | --claude | ~/.claude/skills/ | ~/.claude/commands/ | ~/.claude/agents/ | ~/.claude/rules/ (.md) | yes, plus a Notification hook and a status line, in ~/.claude/settings.json |
| Cursor | bare command or --cursor | ~/.cursor/skills/, plus a copy in ~/.agents/skills/ | ~/.cursor/commands/ | ~/.cursor/agents/ | ~/.cursor/rules/ (.mdc; see below) | yes, in ~/.cursor/hooks.json |
| Codex CLI | --codex | ~/.agents/skills/ | no; type $ or /skills to pick a skill (see below) | no | merged into ~/.codex/AGENTS.md | no |
| Antigravity CLI | --antigravity | ~/.gemini/config/skills/ | each skill is a /name command | no | no | no |
| Gemini CLI | --gemini | ~/.agents/skills/ | portable commands in ~/.gemini/commands/ | no | merged into ~/.gemini/GEMINI.md | no |
Any agent that reads SKILL.md | npx skills add kensaurus/skills | that agent's skills folder | no | no | no | no |
How each path was chosen:
- Codex CLI reads user skills from
$HOME/.agents/skills(Codex skills docs, "Where Codex loads local skills"). It reads global instructions from~/.codex/AGENTS.md(AGENTS.md docs). Codex has deprecated custom prompts (custom prompts page), and a Codex maintainer says they were removed in codex-cli 0.117.0 (openai/codex#15941). So the installer no longer writes them. On current Codex, use the skills. For a Codex release older than 0.117.0, add--codex-prompts. - Antigravity CLI replaced Gemini CLI for most users in June 2026. It reads user skills from
~/.gemini/config/skills/(Antigravity skills docs) and turns each skill into a/namecommand.--autoinstalls for it when~/.gemini/configexists. It is not part of--all. - Gemini CLI reads user skills from
~/.gemini/skills/or the~/.agents/skills/alias. Within the user tier, the alias takes precedence (Gemini CLI skills docs, "Discovery tiers"). It reads global context from~/.gemini/GEMINI.md(GEMINI.md docs). - Codex and Gemini share one copy in
~/.agents/skills/. It is written once per run, and other skills already in that folder are left alone, even by--clean.- With only
--codexor--gemini, the installer writes just this pack's skill folders. - When Cursor is also a target, Cursor's mirror fills the folder instead, and the installer reuses that copy. The mirror copies everything in
~/.cursor/skills/, so your own Cursor skills also become visible to Codex and Gemini. To avoid that, install Cursor in a separate run with--no-agents-mirror.
- With only
- Cursor reads skills from
~/.cursor/skills/. Its rules docs describe project rules in a repo's.cursor/rules/and user rules in Customize → Rules. They do not mention~/.cursor/rules/, where the installer puts the global rules. If a rule does not seem to apply in Cursor, copy the.mdcfile into your project's.cursor/rules/. - Claude Code loads every file in
~/.claude/rules/in every project, except rules with apaths:field, which load only when matching files are open (Claude Code memory docs). - Other agents. The skills CLI copies
SKILL.mdfolders into the folder each agent reads, for example GitHub Copilot, OpenCode, or Antigravity CLI. Pass-gfor your user folder and-a <agent-id>for the agent. The IDs are in the skills CLI README under "Supported Agents". Example:npx skills add kensaurus/skills -g -a github-copilot. You get the skills inskills/only, without commands, subagents, rules, or the hook. - Claude Code plugin. Instead of the installer, you can run
/plugin marketplace add kensaurus/skillsand then/plugin install kenji@kenji. - VS Code and GitHub Copilot. The root
plugin.jsonfollows Agent Plugins 1.0.0, so Copilot can add this repo as a plugin marketplace. In VS Code, openvscode://chat-plugin/add-marketplace?ref=kensaurus%2Fskills. You get the skills inskills/, without commands, subagents, rules, or the hook.
What the installer writes
The installer is one Node script, bin/install.mjs. It has no dependencies or npm lifecycle scripts. It opens no network connection and starts no child processes; npx downloads the package before it runs. It writes only under your home directory:
| Path | What | When it exists already |
|---|---|---|
~/.cursor/skills/<name>/, ~/.claude/skills/<name>/ | One folder per skill | Same-named files are overwritten. Your other skills stay. |
~/.agents/skills/<name>/ | Skills for Codex and Gemini, and Cursor's mirror (which copies all of ~/.cursor/skills/, yours included) | Same as above. Skip it with --no-agents-mirror, which also means Codex and Gemini get no skills. |
~/.cursor/commands/, ~/.claude/commands/ | Slash commands (.md). Claude Code does not get a command whose name is also a skill, because the skill wins there. | Same-named files are overwritten. |
~/.cursor/agents/, ~/.claude/agents/ | Subagent definitions | Same-named files are overwritten. |
~/.cursor/rules/, ~/.claude/rules/ | Global rules (.mdc for Cursor, .md for Claude Code) | Same-named files are overwritten. |
~/.codex/AGENTS.md, ~/.gemini/GEMINI.md | Global rules merged into one file | The old file is copied to <file>.bak-<timestamp> first. An unchanged file is not rewritten. |
~/.gemini/commands/*.toml | plan-mode, research, fix-issue | Backed up like GEMINI.md. |
~/.codex/prompts/*.md | The same three, only with --codex-prompts | Backed up like AGENTS.md. |
~/.gemini/config/skills/<name>/ | Skills for Antigravity CLI, only with --antigravity or when --auto finds ~/.gemini/config | Same-named files are overwritten. Your other skills stay. |
~/.claude/kenji-hooks/, ~/.cursor/kenji-hooks/ | completion-gate.mjs, the Stop hook script; for Claude Code also notify.mjs and statusline.mjs | Overwritten. |
~/.claude/settings.json, ~/.cursor/hooks.json | One Stop hook entry. For Claude Code also one Notification hook entry, statusLine (only if you have none) and skillListingBudgetFraction (only if unset) | Merged in one write. Your other settings and hooks are kept. A backup is written when the file changes. If the file is not valid JSON, the merge is skipped with a warning. |
~/.cursor/mcp.json | MCP server template | Written only if missing. Never overwritten. |
A normal run merges: it adds or overwrites this pack's files and never deletes your own. It also removes leftovers that this pack itself renamed or stopped installing. After every write, the installer hash-checks each copied file.
--clean is different. It deletes the whole skills/, commands/, agents/, and rules/ folders under each targeted ~/.cursor and ~/.claude, your own files in them included, and then installs fresh. It first copies those folders to <base>/.kenji-backups/<timestamp>/, unless you pass --no-backup. It does not touch AGENTS.md, GEMINI.md, or the Codex and Gemini command files. In ~/.agents/skills/ it removes only this pack's skill folders.
All installer flags
npx @kensaurus/skills # Cursor only (the default)
npx @kensaurus/skills --auto # every tool found under ~ (falls back to Cursor)
npx @kensaurus/skills --all # Cursor + Claude Code + Codex CLI + Gemini CLI
npx @kensaurus/skills --claude --codex # pick targets; combine freely
npx @kensaurus/skills --only skills,commands # subset of: skills, commands, agents, rules, hooks, statusline
npx @kensaurus/skills --skill audit-ux # one skill
npx @kensaurus/skills --profile core # only some profiles' skills (core is always included)
npx @kensaurus/skills --profile core --prune # also remove this pack's skills outside the profile
npx @kensaurus/skills --list-profiles # show the profiles and exit
npx @kensaurus/skills --antigravity # skills for Antigravity CLI (~/.gemini/config/skills)
npx @kensaurus/skills --codex --codex-prompts # also write prompts, for Codex CLI before 0.117.0
npx @kensaurus/skills --dry-run # preview, no writes
npx @kensaurus/skills --verify # read-only hash check (add the same target flags)
npx @kensaurus/skills --clean # delete skills/commands/agents/rules folders first, yours too (backed up)
npx @kensaurus/skills --restore [stamp] # put a --clean backup back (latest if no stamp)
npx @kensaurus/skills --link # symlink instead of copy, for editing skills in a clone
npx @kensaurus/skills --no-agents-mirror # write nothing to ~/.agents/skills
npx @kensaurus/skills --help # full help
From a clone, ./install.sh with no arguments installs for Cursor and Claude Code. With arguments, it passes them to node bin/install.mjs.
The Stop hook
The completion gate, hooks/completion-gate.mjs, is registered as a Stop hook in Claude Code and Cursor. It reads the open workspace for .cursor/complete-everything-state.md or .cursor/burndown-state.md. Those files exist only during a complete-everything or burndown-full run. If neither file exists, or the checklist was last updated more than a day ago, the hook exits without doing anything. Otherwise it sends the agent back to the unfinished items instead of letting it stop. After three follow-ups with no item closed, it gives up. Each run has a 5-second timeout and does not block the agent if it fails.
- Skip it at install time:
npx @kensaurus/skills --all --only skills,commands,agents,rules. Any--onlyalso skips themcp.jsontemplate. - Remove it later: delete the entry whose command contains
kenji-hooks/completion-gate.mjsfromhooks.Stopin~/.claude/settings.jsonand fromhooks.stopin~/.cursor/hooks.json. Then delete~/.claude/kenji-hooks/and~/.cursor/kenji-hooks/. A later install without--onlyadds it back.
Notifications and the status line
Claude Code only. These are merged into ~/.claude/settings.json in the same single write as the Stop hook.
- Desktop notifications.
hooks/notify.mjsis registered as aNotificationhook forpermission_prompt(a tool call has waited about six seconds for approval),idle_prompt(Claude finished about a minute ago and you have not typed),agent_needs_inputandagent_completed(background sessions). It callsnotify-sendon Linux,osascripton macOS and a WinRT toast on Windows, spawns no shell, and always exits 0. If the notifier is missing, nothing happens. Source: hooks reference. - Status line.
hooks/statusline.mjsshows model and effort level, folder, branch, context used, session cost, five-hour usage and session name. It is set only when you have nostatusLine; your own is never replaced. It runs no commands: the branch is read from.git/HEAD, so large repos do not slow it down. Source: status line docs. - Skill-listing budget. Claude Code shows the model a listing of your skills against 1% of the context window and, over budget, keeps only the names of the least-used ones. A full install of this pack goes over (a
--profileinstall does not, and leaves the setting alone), soskillListingBudgetFractionis set to0.03when it is unset. A value you set is kept. The installer prints the listing size against the budget in force. Source: skills docs. - Skip them:
--only skills,commands,agents,rulesinstalls no hooks and no status line.--only hooksinstalls the Stop and Notification hooks without the status line;--only statuslineinstalls only the status line. The budget rides with theskillsgroup, because that is what fills the listing. - Remove them later: in
~/.claude/settings.json, delete thehooks.Notificationentry whose command containskenji-hooks/notify.mjs, thestatusLinekey if its command containskenji-hooks/statusline.mjs, andskillListingBudgetFraction. A later install without--onlyadds them back.
The stack-specific parts, and how to skip them
These skills and rules were written on React, Next.js, and Supabase projects. Here is what assumes that stack:
| Piece | What it assumes | Loads when |
|---|---|---|
rules/senior-engineer.mdc | Next.js + React + Supabase defaults: TanStack Query, Zustand, Zod, RLS on every table | Claude Code: only with files under app/, supabase/, or next.config.* open (its paths: field). Cursor: when the agent judges its description relevant. Codex and Gemini: every session, because it is merged into AGENTS.md / GEMINI.md. |
rules/full-stack-ship-discipline.mdc | Deploy schema with the feature; uses Supabase as the worked example | Claude Code, Codex, Gemini: every session (no paths: field). Cursor: when the agent judges its description relevant. |
rules/project-starter/, rules/native-rn-monorepo/ | Per-project rule sets for Next.js/Supabase and React Native | Never installed. Copy them into a project yourself. |
Skills such as plan-rls-audit, backend-realtime, mushi-* | The tool in their name or description | Only when your request matches |
The other global rules (verification-before-completion, approved-plan-execution, shell-first-search, skill-workflows) do not depend on a stack.
To install without the global rules: npx @kensaurus/skills --all --only skills,commands,agents. This also skips the Stop hook, the Notification hook, the status line, and the MCP template. To drop a single rule, delete its file after installing. A later install without --only puts it back.
Verify, update, uninstall
Verify. npx @kensaurus/skills --verify plus the target flags you installed with (--auto, --all, --codex, …). It makes no changes. It fails if a packaged file is missing or differs from the package, and lists each one. Extra files of your own are allowed. Line endings are normalized before hashing, so CRLF copies on Windows still match.
Update. Run the same install command again. Unchanged files are left as they are, and renamed skills from older releases are removed.
Roll back.
--cleanruns keep a backup. Restore it withnpx @kensaurus/skills --restore(Cursor) or--restore --claude, plus a stamp from.kenji-backups/to pick an older one.AGENTS.md,GEMINI.md,settings.json,hooks.json, and the Codex/Gemini command files keep.bak-<timestamp>copies next to them. Copy one back by hand.- A normal merge does not back up skill, command, agent, or rule files it overwrites. If you edited one of this pack's files in place, copy it somewhere first.
Uninstall. There is no uninstall command. Remove the pieces by hand:
# Skills, commands and agents this pack ships (bash; uses a fresh clone for the names)
git clone --depth 1 https://github.com/kensaurus/skills.git /tmp/kenji
for d in /tmp/kenji/skills/*/ /tmp/kenji/skills-cursor/*/; do
n=$(basename "$d"); rm -rf ~/.cursor/skills/"$n" ~/.claude/skills/"$n" ~/.agents/skills/"$n"
done
for f in /tmp/kenji/commands/*.md; do rm -f ~/.cursor/commands/"$(basename "$f")" ~/.claude/commands/"$(basename "$f")"; done
for f in /tmp/kenji/agents/*.md; do rm -f ~/.cursor/agents/"$(basename "$f")" ~/.claude/agents/"$(basename "$f")"; done
Then remove what is left:
- the global rules listed in Rules, from
~/.cursor/rules/and~/.claude/rules/; - the Stop hook, as described in The Stop hook, and the Claude Code Notification hook,
statusLineandskillListingBudgetFraction, as described in Notifications and the status line; ~/.codex/AGENTS.mdand~/.gemini/GEMINI.md(or restore a.bak-copy);- the
plan-mode,research, andfix-issuefiles in~/.codex/prompts/and~/.gemini/commands/.
Leave ~/.cursor/mcp.json alone if you have edited it. To remove the Claude Code plugin, run claude plugin uninstall kenji@kenji (docs). To remove skills installed by the skills CLI, run npx skills remove.
Troubleshooting
Windows: EPERM or EBUSY during install
Another program is holding a file, usually an open editor watching ~/.cursor or ~/.claude, or an antivirus scan. The installer retries each file for about 3 seconds, then stops and names the file. Close the editor and run the command again; a merge install is safe to repeat.
Windows: line endings (CRLF)
--verify normalizes CRLF to LF before hashing text files, so a checkout made with core.autocrlf=true verifies cleanly. The repo's .gitattributes checks text files out with LF. If ./install.sh still fails with $'\r': command not found, the script was checked out with CRLF line endings: clone again, or run node bin/install.mjs with the same arguments. install.sh needs bash (Git Bash or WSL); the Node installer does not.
"I already have skills, rules, or settings in those folders"
A merge install keeps them. It only overwrites files with the same name as one of this pack's files. Run --dry-run first to see the exact list. Hook config files are merged entry by entry, and a backup is written when they change. To keep your own AGENTS.md / GEMINI.md instead of the merged one, install Codex or Gemini with --only skills,commands (or --only skills).
Codex shows only some skills, or shortened descriptions
Codex lists skills within a budget: 2% of the model's context window, or 8,000 characters when the window is unknown. With many skills installed, it shortens descriptions first and may leave some skills out with a warning (Codex skills docs). To install fewer skills, use --codex --skill <name> once per skill. To turn one off, add [[skills.config]] with enabled = false in ~/.codex/config.toml. Claude Code has a similar listing budget; the installer prints its current size after a --claude install.
Codex: the /prompts: commands are missing
Current Codex releases no longer load ~/.codex/prompts/ (see the links in Which agents are supported). Use the skills instead: type $ or run /skills.
Codex: I set CODEX_HOME
The installer always writes ~/.codex/AGENTS.md. Copy them into your CODEX_HOME. Codex reads AGENTS.override.md instead of AGENTS.md when that file exists, so check for one. Skills are not affected; Codex reads $HOME/.agents/skills either way.
Gemini CLI asks before a skill runs
That is how Gemini CLI works: activating a skill asks for consent and grants the skill's folder read access (docs). In a running session, /skills reload picks up new skills and /memory reload re-reads GEMINI.md. The same page says Gemini CLI was replaced by Antigravity CLI for unpaid-tier and Google One users on June 18, 2026. This installer does not target Antigravity CLI. The skills CLI lists an antigravity-cli agent ID.
A skill does not trigger
Skills are matched against what you say. Name the skill to force it, for example "use audit-security on this repo", or call it directly: /audit-security in Claude Code and Cursor, $audit-security in Codex. Trigger phrases for every skill: docs/CATALOG.md.
How one request lifts the repo
Every playbook follows the same five steps. It assesses before it changes anything, and proves the result before it ships.
| Rung | What runs | What the repo gains |
|---|---|---|
| Orient | workflow-onboard, /research | A decision log |
| Assess | audit-*, plan-*, /grill-me | A failing test, or a plan you approve |
| Change | design-*, enhance-*, backend-*, housekeep-* | A diff you can read in one sitting |
| Prove | test-*, complete-everything, completion-judge | Evidence, not a claim |
| Ship | workflow-ship-and-observe, deploy-*, debug-* | A release that stayed up |
The 23 plan-* skills audit first and wait for your approval. See docs/PLAN-LOOPS.md.
What should I say?
| You say… | What kicks in | What you get |
|---|---|---|
| "orient me" | workflow-onboard | A short tour of the codebase |
| "grill me before I build" | workflow-grilling | One question at a time until you're aligned |
| "build this feature" | workflow-build-feature | Spec → tests → code → smoke → PR |
| "fix this bug and ship it" | workflow-fix-and-ship | Debug → fix → verify → PR → deploy |
| "this mobile app feels like a website" | enhance-mobile-native-feel | Native chrome, haptics, motion, lists, in one pass |
| "fix our mobile UI/UX end to end" | workflow-mobile-native-uiux | Native pass → accessibility → device QA → PR |
| "audit my security" | audit-security | OWASP-style findings with file:line |
| "is this production-ready?" | audit-resilience + audit-realworld | Timeouts, retries, parity checks |
| "complete everything" | complete-everything | No parked leftovers; a judge verifies "done" |
| "ship it and watch it" | workflow-ship-and-observe | Deploy → verify live → observe / rollback |
| "take this to market" | workflow-gtm | GTM plan → approve → analytics, hero, onboarding, SEO, launch kit |
Not sure which line fits? Type /kenji and the job in your own words. It picks the playbook and runs it.
To force a skill: "use enhance-web-ux on /dashboard". Every trigger phrase → docs/CATALOG.md.
Tour
Each time, you describe the job and a named playbook runs. The part that sets this pack apart is the completion gate: it checks the moment the agent tries to stop. The capture below is real output; only the agent line was added. More in A gate that does not take the agent's word for "done" (日本語).
Four more, as stills:
What's inside
- Skills start from a plain request.
- Commands start with
/. - Subagents take one task in a separate context.
- Rules are standing instructions the agent loads.
- MCP templates pin exact server versions against package-hallucination attacks.
skills-cursor/ holds Cursor's own built-in skills, ported so Claude Code can use them. The installer gives them to Claude Code only, because Cursor already ships them. Every skill follows the Agent Skills spec, and npm test checks all of them on every change.
Skill families at a glance
| Family | Count | In one sentence |
|---|---|---|
| Audit — inspect; some then fix | 32 | Check the codebase: security, UX, analytics, IAP, the skill pack |
| Plan — audit first, change only after you approve | 23 | Write a fix plan you approve before any code changes |
| Enhance — improve what already exists | 23 | Polish UI, forms, motion, SEO, PWA, email deliverability |
| Design — build something new | 10 | Create new UI, APIs, emails, themes from scratch |
| Backend — server & data patterns | 5 | Auth, caching, queues, realtime, observability |
| Mobile — React Native / Capacitor | 5 | RN screens, emulators, Capacitor, App Store prep |
| Data — charts & pipelines | 2 | Charts, dashboards, ETL / cron jobs |
| Docs — write it down clearly | 6 | READMEs, PRDs, RFCs with a reader-first voice |
| Housekeeping — consolidate or clear one drifted register | 5 | Consolidate one drifted register (gates, backlog, design tokens, dead code) |
| Workflows — multi-step recipes | 22 | End-to-end recipes (build, fix, ship, green the repo) |
| Test & QA — prove it works | 8 | Unit, Playwright, visual regression, load, red-team |
| Deploy — ship & verify | 2 | npm release + post-deploy smoke tests |
| Debug — find & fix what's broken | 3 | Errors, Sentry, frontend-backend mismatches |
| Iterate — close the loop after launch | 3 | Post-launch feedback loops and agent-harness iteration |
| Mushi Mushi — bug triage helpers | 2 | Integrate the Mushi Mushi bug-report pipeline |
| Protocols — session guardrails | 1 | Keep browser automation from freezing |
| Authoring — build skills & MCP | 2 | Author new skills or MCP servers |
| Third-party (upstream-maintained) | 3 | Vendored upstream skills (Emil, UI/UX Pro Max, Vercel WIG) |
| Core & cross-cutting | 5 | Close everything, burndown, research, handoff |
| Cursor IDE skills | 12 | Canvas, hooks, rules, PR splitter, CLI helpers |
| Total | 174 | Every skill, one line each |
Generated from each skill's SKILL.md by npm run gen:skill-index. 174 skills. The full list is docs/SKILLS.md; trigger phrases are in docs/CATALOG.md.
Commands, subagents, MCP, rules
Commands: type / in chat to see them all
There are 63 slash commands. Claude Code and Cursor get all of them. Codex and Gemini get the portable ones in commands-portable/.
| Command | When | What |
|---|---|---|
/burndown-full | A refactor stopped early | Drive the plan to 100% repo coverage with a verification gate |
/complete-everything | A plan marked done with deferrals | Close planned, parked, and discovered work; run the test ladder |
/green-repo | Whole-repo debt cleanup | Drive typecheck, lint, test, and build to green |
/ship-and-observe | Deploy to production | Verify the live revision, watch the window, roll back if needed |
/feedback-to-closure | Reports, QA, Sentry | Dedupe into tickets, fix, verify in production |
/plan-mode | Before coding | Research + approved plan (/plan is the host mode switch) |
/commit · /pr · /fix-issue [#] | Git flow | Lint and commit · push and open PR · issue to fix to PR |
/debug-issue · /review-code · /test | Quality | Runtime evidence · review · route to the matching test skill |
/research · /readme · /refactor · /update-deps | Maintenance | Doc research · README sync · modular split · safe dep updates |
/uiux · /readability · /responsive-audit · /instant-nav | UI | Design system · reading load · breakpoints · instant navigation |
/deadcode | Unused files, exports, deps | Knip baseline, then delete by category behind a shrink-only ratchet |
/gtm · /gtm-weekly · /launch-kit | Growth | GTM plan and loop · weekly review · launch copy from verified claims |
/skill-conflicts · /mcp-guide | Pack hygiene | Overlapping triggers · MCP tool reference |
/*-plan (23 aliases) | Audit before changing | Thin pointers to the plan-* skills; audit and plan only |
RN monorepo bundle: copy commands/native-rn-monorepo/ and rules/native-rn-monorepo/ into your project (iOS builds on CI). Full command list → docs/CATALOG.md.
Subagents: Claude Code and Cursor
There are 6 subagents:
| Agent | Triggers on | Output |
|---|---|---|
code-reviewer | "review", code changes | Quality, security, types |
debugger | Errors, exceptions | Root cause + fix |
db-migrator | "migration", "new table" | SQL, RLS, indexes |
deploy-checker | "deploy", "ship it" | Pre-deploy validation |
perf-monitor | "slow", "optimize" | Perf audit |
completion-judge | A closure claim | PASS / CONTINUE / BLOCKED against plan, state, diff, and fresh evidence |
MCP server templates
cp ~/skills/mcp/mcp.json.template ~/.cursor/mcp.json # essential 3
cp ~/skills/mcp/mcp-full.json.template ~/.cursor/mcp.json # every server below
The installer writes the essential template to ~/.cursor/mcp.json only when that file does not exist. Set FIRECRAWL_API_KEY, CONTEXT7_API_KEY, SUPABASE_ACCESS_TOKEN, and SUPABASE_PROJECT_REF in your environment. The Slack and Notion entries in the full template still use YOUR_* placeholders. Never commit real keys (SECURITY.md). Setup → mcp/README.md.
| Tier | Servers | Keys? |
|---|---|---|
| Essential | Firecrawl, Context7, Supabase | Firecrawl + Context7 + Supabase |
| Dev | GitHub, GitHub Official, Playwright, Postgres, Memory, Chrome DevTools | PAT / conn string |
| Cloud | AWS Lambda, S3, CloudWatch, Redis | AWS profile / URL |
| Productivity | Slack, Notion | Bot token / API key |
Rules: global and per-project
Global rules the installer writes: verification-before-completion.mdc (always on), approved-plan-execution.mdc, full-stack-ship-discipline.mdc, senior-engineer.mdc, shell-first-search.mdc, skill-workflows.mdc. Codex and Gemini get all of them merged into one file, except skill-workflows. Stack assumptions and how to skip them: The stack-specific parts. Model and effort routing: docs/MODEL-AND-EFFORT.md.
Per-project rules for a Next.js + Supabase app are never installed globally. Copy them in yourself:
cp ~/skills/rules/project-starter/*.mdc your-project/.cursor/rules/
| Rule | Enforces |
|---|---|
supabase.mdc | Typed clients, RLS, migrations |
typescript.mdc | No any, Zod, ActionResult |
components.mdc | Primitives, Server Components, a11y |
tailwind.mdc | Tokens, mobile-first |
data-fetching.mdc | TanStack Query, RSC prefetch, 'use cache' |
web-performance.mdc | LCP priority, INP yield, bfcache, budgets |
git.mdc | Conventional commits, no secrets |
Project constitution: copy docs/AGENTS.template.md into your app repo as AGENTS.md.
Plan → apply pairs: plan-uiux-unification → housekeep-design; plan-dead-code → housekeep-dead-code; plan-gtm → workflow-gtm. The plan half audits and stops. The apply half changes the repo only against a list you approved. Third-party skills are prefixed thirdparty-* and carry an ATTRIBUTION.md → docs/THIRD-PARTY-SKILLS.md.
Shell helpers and repository layout
source ~/skills/shell-aliases/cursor-helpers.sh
newskill <name> · lsskills · cursor-sync · cursor-dev · newrule <name> · newagent <name> · gc <type> <msg> · gp. Definitions in shell-aliases/cursor-helpers.sh (clone only).
skills/
├── skills/ # 162 Agent Skills (SKILL.md each)
├── skills-cursor/ # Cursor's built-in skills, ported for Claude Code
├── commands/ # 63 slash commands
├── commands-portable/ # commands ported to Codex and Gemini
├── agents/ # 6 subagents
├── hooks/ # Stop hook, Claude Code Notification hook and status line
├── rules/ # global rules + per-project bundles
├── mcp/ # MCP templates and pinned versions
├── docs/ # SKILLS, CATALOG, PLAN-LOOPS, GETTING-STARTED, …
├── notepads/ # context templates (clone only)
├── shell-aliases/ # bash helpers (clone only)
├── scripts/ # validators, count and index checks, install tests
└── bin/install.mjs # the installer
Upgrading from cursor-kenji
2.0.0 renamed the pack: repo kensaurus/cursor-kenji → kensaurus/skills, npm @kensaurus/cursor-kenji → @kensaurus/skills, plugin cursor-kenji@cursor-kenji → kenji@kenji, slash namespace /cursor-kenji: → /kenji:. Old GitHub URLs redirect.
What to run once
| You installed with | Do this once |
|---|---|
npx @kensaurus/cursor-kenji | npx @kensaurus/skills --all. The installer moves its Stop hook from cursor-kenji-hooks/ to kenji-hooks/ and removes the old copy. |
npx skills add kensaurus/cursor-kenji | npx skills add kensaurus/skills. Old installs still update through the GitHub redirect. |
| Claude Code plugin | Existing installs keep loading under the new name. For the clean ID: /plugin marketplace remove cursor-kenji, /plugin marketplace add kensaurus/skills, /plugin install kenji@kenji. |
| A clone | git remote set-url origin https://github.com/kensaurus/skills.git |
CURSOR_KENJI_GATE_STATE_DIR and CURSOR_KENJI_DIR still work; the new names are KENJI_GATE_STATE_DIR and KENJI_SKILLS_DIR.
FAQ
How do skills trigger? How do I force one?
You talk normally. The agent matches your words against each skill's description. To force one, name it: "use audit-security on this repo". Full trigger list: docs/CATALOG.md.
What is the difference between audit-* and plan-*?
audit-* skills report findings and stop, unless the skill says it fixes what it finds (audit-responsive, audit-code-quality, audit-performance, audit-security, audit-i18n, audit-bundle-size, audit-agent-speed). plan-* skills only write a plan-{name}.md burndown, and you approve each phase before any code changes. See docs/PLAN-LOOPS.md.
How many skills, and where do the counts come from?
162 agent skills in skills/ plus 12 Cursor-specific skills in skills-cursor/ (174 total). Cursor, Codex, Gemini, and the skills CLI get the first set. Claude Code gets both sets. Every count in this README comes from the filesystem: npm run check:skills fails CI when a number here drifts, and npm run fix:skills rewrites it. The family table is generated by npm run gen:skill-index.
Do I need React, Next.js, or Supabase?
No. See The stack-specific parts for what assumes that stack and how to leave it out.
Where do MCP API keys go? Is there an llms.txt?
Copy mcp/mcp.json.template to ~/.cursor/mcp.json and set the environment variables. Never commit real keys (SECURITY.md, mcp/README.md). Machine-readable docs: llms.txt.
Contributing
mkdir -p skills/my-skill && $EDITOR skills/my-skill/SKILL.md
npm test # spec validation, counts, index, install tests, dead-code check
Each skill must pass Agent Skills spec validation (npm run validate:skills): name matches the folder, description is at most 320 characters, and the body is under 500 lines. See CONTRIBUTING.md, docs/README.md, docs/DISTRIBUTION.md, and docs/TRIGGER-CHEATSHEET.md. Releases: docs/PUBLISHING.md.
Related
- agentskills.io: the Agent Skills spec these skills follow
- skills CLI and skills.sh: install
SKILL.mdfolders into many agents - awesome-cursorrules: curated rules collections
- SkillsMP: aggregator listing of this repo
This pack ships skills, slash commands, subagents, rules, and MCP templates in one installable package. Where it is listed → docs/DISTRIBUTION.md.
More from KENSAURUS
All apps on Google Play: kensaurus developer page. Everything else: kensaur.us.
The list comes from https://kensaur.us/apps.json. npm run sync:family rewrites it; do not edit it by hand.