Skip to content

hktitan/kheskills

unversioned · 5168196f9e01MIT

Harshit Khemani's 19 agent skills: UI craft, design tokens, SVG, stickers, diagrams, sound, launch films, demos, packaging, the poteto pyramid, a content engine, and brains that answer from one source's public work with the source named. Start with khe-auto.

Changelog

What changed in kheSkills, newest first. One entry per day that something a user can see changed: a skill added, a command renamed, a rule tightened. Fixes to wording inside a skill are in the commit history.

2026-11-17

Added

  • npx skills add HKTITAN/skills installs only the 19 khe skills again. The skills CLI 1.9 checks out git submodules on install, so it had been finding 30 of the linked authors' skills inside skills/ and installing them as this repo's (49 in all). Every link in .gitmodules now carries update = none, git skips them, and scripts/check.mjs fails on a link without it.

  • In Claude Code, kheskills (1.2.1) and khe-brain (1.6.1) no longer load each other's parts. Both were rooted at ./, and Claude Code loads a root plugin's commands/ and agents/ whatever its entry lists (an empty list falls back to the default folder; claude plugin details showed it), so kheskills also served the seven brain commands and khe-brain also served the four kheAgents. Each brain's command now sits in its own folder (skills/khe-brain-varun/commands/varun-mayya.md) and khe-brain is rooted at ./skills. The commands keep their names (/khe-brain:varun-mayya), and the --sparse list drops commands. scripts/check.mjs fails on a root commands/ folder and on a khe-brain rooted anywhere else.

  • The repo is an Agent Plugins 1.0.0 plugin: plugin.json at the root names it kheskills and serves every khe skill in skills/, brains included. Every install command from before works unchanged.

  • scripts/marketplace.mjs now writes every install manifest from the folders: the kheskills and khe-brain entries in the Claude Code marketplace (their skills, agents, commands and descriptions; only version stays hand-set), the author plugins, the root plugin.json and the README install block. Every marketplace entry gains a homepage, and the two own plugins a repository and keywords.

  • scripts/check.mjs holds each manifest to its own spec: skills.sh.json to the skills.sh schema (closed fields, its length limits, the Brains group naming every /khe-brain: command), plugin.json to the Agent Plugins schema, the marketplace to Claude Code's rules (names, sources, existing paths, a semver version named on a CHANGELOG line). It fails when what kheskills or khe-brain serves changed since main and its version did not, since Claude Code keeps an installed copy until the version moves. It fails on a model name in the agents, commands, policy, router or top-level docs, and on an agent model other than inherit. CI fetches main for the version check and runs claude plugin validate . --strict.

  • khe-package, Emil Kowalski's way of packaging work rewritten in this collection's format: it picks the format (an explainer page, a click-through of one live component, a hover reel, a Stageshow scene, a recording page), sends three ideas as stills before any render, renders frame by frame on a paused clock at 60 fps with the short side at 2000 px or more, draws the operating system's own cursor, and writes the post's first line for a stranger. It ships its own renderer, render.mjs, which seeks every CSS and WAAPI animation to Playwright's fake clock each frame, and a gate, package-check.mjs, over the plan and the rendered file. Four gold plans, an explainer page, a live stage and eleven teaching pairs.

  • khe-auto places every correction made during a job with khe-poteto-pyramid before handing over (Non-negotiable 11 and a new step 8), so a fix outlives the chat and the next agent inherits it.

  • README figures in the khe design system: the poteto pyramid with each layer's hand-offs, khe-package's flow, and a map of how the skills hand off, built by .github/showcase/figures.mjs and checked with khe-diagram's gate.

  • Cheatsheets for khe-poteto-pyramid and khe-package.

  • Installing from the marketplace works for everyone: the README install block (generated by scripts/marketplace.mjs, checked by check.mjs) adds it with --sparse, so Claude Code no longer downloads the 68 linked submodules and times out; Matt Pocock's plugin is served whole because his repo ships its own manifest (generation now checks every author repo for one); the kheskills plugin no longer loads the brain commands a second time.

  • kheAgents: khe-router, khe-executor, khe-reviewer and khe-pyramid ship in the kheskills plugin (1.1.0) as kheskills:<name>, and orchestration/CLAUDE.md is the policy that uses them: route first, the top session only orchestrates, one sub-session per task, executors do the work, evidence before done, every correction placed with the pyramid.

  • The marketplace serves the skills I use from other people too, one plugin per author straight from their repo: emil-kowalski, benji-taylor, vercel, vercel-ui-sdk, pstack (Cursor's Claude Code build) and matt-pocock. scripts/marketplace.mjs builds them from the README's links, and scripts/check.mjs fails when they drift.

  • A kheskills plugin in the marketplace: every khe skill except the brains in one Claude Code install, /plugin install kheskills --marketplace HKTITAN/skills. scripts/check.mjs fails when a skill is left out of it.

  • khe-auto has a voice: one friend's voice over every skill, warm and dry, no preamble, matched to the person's length and lowercase, emoji only after they use one. Its gate adds stock-phrase: for help-desk filler such as "let me know if you need anything" and "great question".

  • Lauren Tan's pstack skills are linked and installed through their Claude Code build (harness/pstack/claude-code, called as /pstack:<name>) instead of the Cursor originals. khe-auto's README shows how it runs and a real routed reply.

  • khe-launch-film starts from a five-line brief and the one moment the film must show (new brief, structures, publish and sources references, after Hiten Shah's Product Launch Video Playbook of 16 November 2026); its gate adds moment-missing: and the candidate no-brief:, reviews end with a cold-viewer test, and the router's launch-film playbook opens with the brief.

  • khe-poteto-pyramid, built on Lauren Tan's ("poteto") pyramid from her talk "All Code Is Slop" at AI Engineer Code Summit (11 November 2026), read through Miguel Alvarado's study guide on Artesano and her pstack guides. It puts each correction you give a coding agent at the highest layer that can hold it (codebase, then lints, types and CI, then verification, then prose, with review last), builds it there, proves the check fails on the past mistake, and deletes the instruction it replaces. Four jobs: place, sweep, triage (an AGENTS.md into checks and judgment calls) and audit (ten questions, and whether a loop is ready to automate). Six gold placement records and a triaged AGENTS.md for an invented app, eleven teaching pairs, and a gate, pyramid-check.mjs, that reads placement records and rules files.

  • A new Engineering domain in khe-auto, with eight overlap rows.

  • 21 of Lauren Tan's pstack skills are linked from cursor/plugins and routed (poteto-mode, correct, create-verification-skill, create-formal-verification-skill, statechart, how, why, architect, interrogate and others), with her plugin install line.

  • 13 of Matt Pocock's skills are linked from mattpocock/skills and routed (codebase-design, improve-codebase-architecture, diagnosing-bugs, setup-pre-commit, git-guardrails-claude-code and others); his code-review and tdd are pocock-code-review and pocock-tdd here.

Changed

  • Verification and review, after Addy Osmani's "How the Claude Code team thinks about verification" (17 November 2026). khe-poteto-pyramid 1.1.0: a check the change can edit is no check (Hard rule 3), and a class needs a trigger readable from a diff and must survive an attempt to disprove it (placing). Its verification reference puts end-to-end runs and a project verify skill first and adds the four common properties and differential tests; its checks reference adds observe mode before enforce, diff-triggered reliability checklists, flaky tests (fix or delete with a reason, never retry) and pruning; its audit reviews the reviewers against reverts and clean controls.
  • kheAgents (kheskills 1.2.0): khe-executor runs cheap checks first, drives the change end to end, stops after three failed rounds and stamps how it verified; khe-reviewer tags findings important, nit or pre-existing, names the commit it reviewed, builds on earlier rounds' findings and flags a change that touched its own gate; khe-pyramid tries to disprove a rule before building it and starts new gates in observe mode. No model is hardcoded any more: the four agents inherit the session's model (model: inherit), and orchestration/CLAUDE.md has whoever starts a step pick its tier by the task. khe-auto 1.1.0 and orchestration/CLAUDE.md carry the same rules: three rounds then the person, one tuned reviewer over several rounds, checks out of reach of the change.

2026-11-15

Added

  • khe-brain-enrico, a brain of Enrico Tartarotti's YouTube channel: all 88 long videos on his Videos tab as of 15 November 2026 (17 hours 57 minutes), read in full from captions. It answers how tech products are designed, engineered and built with the video and the minute named, pressure-tests a design or a product plan against his stated positions, and says so when the videos do not cover a question. 1,406 claims in twelve theme files, a core file, a glossary of 375 terms and frameworks, 19 disagreements, 5 changes of mind he states himself and 29 topics the videos are silent on. Every cited line was checked against the captions at its time by a reviewer who had not written it; references/sources.md says what that found. Sponsor reads are left out, and lines by the Figma, Notion and Lovable teams he interviews are never reported as his. In the plugin it is /khe-brain:enrico-tartarotti (1.6.0).

  • khe-brain-harshit, a brain of Harshit Khemani, built from his own X and LinkedIn exports of 13 November 2026 (7,853 posts, threads and replies; 666 cited). It answers with the post named, pressure-tests a plan against his posted views, and drafts khe.money posts in his voice since mid-2025 with one Rests on: line per view, each cited. Ten theme files, a voice file measured on his posts, a story timeline, a glossary, his changes of mind and the topics his posts are silent on. Its gate adds draft lines: no Rests on:, hashtags, emoji, em dashes and sentence case. In Claude Code it is /khe-brain:harshitkhemani, and the khe-brain plugin is 1.5.0.

  • The README has a section on khe, my design system, with its brand page at ds.khe.money. khe ships as a contract, khe.DESIGN, which khe-design-tokens reads like any .design file, so it is linked here rather than added as a skill.

2026-11-13

Added

  • khe-ui has a ui-sdk lane for a screen built with Vercel's UI SDK (@ui-sdk/*, in beta): what the library decides and what this skill still decides, the hooks each value is written on (data-slot, state attributes, data-starting-style and data-ending-style, @layer ui-sdk), and what the optional web.css look leaves unfinished against this skill's numbers. It never restates the library.
  • khe-design-tokens binds a contract to the --ui-* variables of UI SDK's web.css.
  • khe-auto routes to Vercel's own ui-sdk skill for composing, building and testing UI SDK components, with its install line, two overlap rows and a step in the new-screen playbook. The skill is linked from the README under "From others" and is not copied or pinned here.
  • khe-brain-isb 1.1.0 adds 146 session notes from ivi, the I-Venture Immersive programme at ISB (April to August 2025), in ivi/: one note per recording of a class, a bootcamp or a guest talk, each naming its speaker and date. Twelve ivi-* theme files hold 2,300 or so claim lines, each linked to its session note. A session line is reported as its speaker's and never as the school's. Classmates, their startups and anything a speaker asked to keep in the room are left out, and two bootcamps are held back for that reason. 613 pieces in all.
  • Thirteen more of Vercel's skills are linked and routed: nine from vercel-labs/agent-skills (React and Next.js performance, composition patterns, React Native, view transitions, deploying, the CLI with tokens, optimising a deployed project, the web design and writing guidelines) and the four contributor skills in vercel-labs/ui. khe-auto has a row for each and three new overlap rows.
  • Vercel's ui-sdk skill is pinned in skills/ui-sdk like Emil's and Benji's, so it follows vercel-labs/ui automatically.
  • khe-brain-isb 1.1.1 adds seven anonymised notes, titled a bootcamp session, for two ivi bootcamps that were held back: general teaching points only, with the speaker, the date and the speaker's own frameworks, examples and materials left out. 153 session notes and 620 pieces in all.
  • The khe-brain plugin is 1.4.0, so an installed copy picks up the session notes on update.
  • Emil Kowalski's learn-taste is linked in skills/ and routed by khe-auto, with an overlap row against khe-design-tokens. It turns the design corrections made in a session into rules with their reasons.

Changed

  • The pins on Emil Kowalski's skills follow his update of 2026-11-13, which corrected API names and contradictions between skills across all nineteen.

2026-11-12

Added

  • khe-brain-yc: answers startup questions from 88 Y Combinator pieces (14 Paul Graham essays, 28 library posts and 46 talks with a published transcript), with the person who said it named and the piece linked. 1,202 claims in twelve theme files, every one checked against the source text. A claim is never reported as what YC says, only as what a named partner said.
  • /khe-brain:yc in the khe-brain plugin.
  • khe-brain-isb: answers business questions from the public isb-skills notes, with the file named. 467 pieces, 368 claims in twelve theme files, each claim checked against its source file. It reports the notes as the notes and never as what ISB teaches, and marks a benchmark the notes give without a source.
  • /khe-brain:isb in the khe-brain plugin.

Changed

  • The plugin's commands take a person's full name: /khe-brain:emil-kowalski, /khe-brain:benji-taylor, /khe-brain:varun-mayya. An institution takes its usual short name. The skill folders keep their names, so npx skills add installs are unchanged.
  • scripts/check.mjs matches a command to its skill by the path the command reads, and fails when a brain has no command.

2026-11-11

Added

  • khe-brain-varun: answers startup, content, career and AI questions from 18 of Varun Mayya's public videos (the 15 videos of his Kata playlist and three podcasts), with the video and the minute named. 1,378 claims in twelve theme files, a core file, a glossary of 274 terms and a list of what the videos do not cover. Times are good to about two minutes; references/sources.md says how the lines were checked.
  • The khe-brain plugin for Claude Code. /plugin marketplace add HKTITAN/skills, then /plugin install khe-brain@kheskills, puts every brain under one name. Each command reads its skill's folder, so there is one source for both ways of installing.
  • skills.sh.json carries its schema link, and the Brains group names the plugin.

Changed

  • khe-brain-varun's gate checks a time on every video citation and that the time matches its link, and its replies can carry up to three Also: lines for supporting claims.

2026-11-10

Added

  • The collection: khe-auto, khe-ui, khe-design-tokens, khe-svg, khe-sticker, khe-diagram, khe-sound, khe-launch-film, khe-demo, khe-brain-emil, khe-brain-benji and khe-content-engine, each with gold examples, teaching pairs and a gate, and nineteen skills by Emil Kowalski and Benji Taylor linked beside them.