agent-kitchen

Designs Claude Code harness artifacts. One triage ladder picks the right surface (skill, hook, path-scoped rule, CLAUDE.md or AGENTS.md entry, workflow script, subagent, MCP); the stance holds the bar every artifact must clear (a gap between the model's default and what the owner wants — a new idea or source worth operationalizing, a miscommunication to prevent, or an observed failure — plus a nameable attention shift); per-surface mechanics live in references. Use for creating, reviewing, or refactoring any of these, for distilling articles or docs into one, for skills that under- or over-trigger, hooks that don't fire or fire too often, bloated CLAUDE.mds, and for deciding which surface a behavior belongs on. Triggers on any work on a skill, hook, rule, CLAUDE.md, AGENTS.md, or workflow orchestration script, including debugging triggering or firing semantics and packaging a /command or plugin.

Always-on design-engineering discipline for shadcn (Base UI) + Tailwind v4 + Next.js + Vercel, applied within the framework's conventions. Slows UI choices down: fluid type and spacing configured at the token layer (never inline arbitrary clamps), layout primitives when simpler patterns break, container queries for varying-width slots, frequency-aware motion (Tailwind utilities before Motion, often none at all), and proactive polish applied unprompted. Use for any request to build, lay out, animate, polish, review, or fix the feel of an interface, from 'build a landing page' to vague asks like 'this feels off' or 'make it feel premium'.

Stack-aware, read-only quality audit for a JavaScript/TypeScript web repo. Detects the stack (Next.js, React, shadcn, Tailwind v4, Motion) from package.json, runs real verification (lint, typecheck, build, react-doctor), then routes a dimensional review (correctness, Next.js, React performance, web vitals, shadcn/Tailwind, design polish, motion performance, accessibility, security and best practices, components, view transitions) into one P0/P1/P2 report with file:line and a concrete fix per finding. Read-only by default; opt into P0-only fixes on a branch with `mode: fix`. Use for a whole-repo quality, design, accessibility, or performance audit, a pre-ship review, or a scheduled quality scan. For a single file or diff, /code-review covers it; for a one-component design pass, design-engineer applies automatically.

Stack-wide UI discipline for shadcn 4.x (Base UI default, Radix supported) + Tailwind v4, covering component architecture and token mechanics. Architecture: edit the source in `components/ui/`, don't build a parallel component or wrapper; compose, don't prop; keep primitives uncontrolled by default; bare data attributes (`data-open`, `data-pressed`) for visual state, `data-slot` for parent-aware targeting. Mechanics: rem and oklch by default, hex never; semantic tokens over raw palettes; map design values to existing tokens and ask when nothing fits; `render` not `asChild` on Base UI. Read `globals.css` and the `shadcn/tailwind.css` base layer for project tokens before writing classNames. Use for any question about shadcn or Tailwind v4 conventions: composing or extending components, variants, controlled vs uncontrolled, wrapping Base UI, tokens, registries, or why a utility class isn't taking effect.

Writes GitHub-Flavored Markdown documents: READMEs, docs, CONTRIBUTING, issue/PR templates, technical .md files. Picks the form whose job matches the content (table for a comparison, alert for a callout, collapsible details for long output, task list, footnote, mermaid/math) and cuts everything decorative. Carries the GitHub-specific traps the default gets wrong: exact alert syntax, swatch and heading-anchor rendering rules. Use for structured documents that are scanned and looked up, and for asks like 'document this', 'create or format a markdown file', 'clean up or fix this README', 'add a table, callout, diagram, or collapsible section', or 'why isn't this rendering on GitHub'.

Encodes ways to produce written tone of voice that articulates ideas in as few good words as possible, built on the idea of 'Saltintesta' put forward in Paul Graham's 'Write Simply'. Use when drafting or editing any prose meant to be read with attention.

Default to optimistic-UI patterns where perceived speed matters. Render from cached state immediately, mutate optimistically with rollback, don't gate render on session validation, and never let heavy synchronous work sit between an input and the next paint. Applies to mutation handlers (save / edit / delete / toggle), list and detail views, dashboards, and any reactive UI measured against Linear / Superhuman / Raycast / Vercel-dashboard-class speed. The coding pattern at the component layer only, not a sync-engine, CRDT, or IndexedDB-architecture skill. Use when an interface feels slow, laggy, or sluggish on input (including a bad INP score), when deciding whether to show a spinner or loading state, or when asked for optimistic updates or Linear-class snappiness.

Audits a Claude Code setup end-to-end: what loads every session, what it costs, and whether the harness's own claims hold. Inventories the standing surfaces (CLAUDE.md chain, personal/project/plugin skills, MCP instruction blocks, agent rosters, hooks, settings at every scope), quantifies their per-session token cost, runs four checks — self-consistency, duplication across scopes, enforcement parity, scope discipline — and reads the CLAUDE.md chain as intent (code outranks the doc; no restated stack, deps, or layout). Output is a quantified triage that stays in-repo: findings at machine scope (anything under ~/.claude, user or enterprise settings, global plugins) are reported for the owner to action, never edited. Use to review, slim, or sanity-check a whole Claude Code setup — including 'why is my context so big' or 'what loads every session' — after major harness changes (a new plugin fleet, a new machine, a model release) or when sessions feel slow from turn one; for a single artifact, the relevant forge o

Mines Claude Code session transcripts for the corrections that earn artifacts — the intake stage of the kitchen's loop. Extracts what the user actually said across recent sessions (via a bundled script that reduces multi-MB transcripts to user turns), dispatches readers to classify corrections, redirects, and repeated failures, dedups against the standing harness, and hands surviving candidates to the forge with evidence counts. Use on a cadence (end of a heavy week, after a project push), when a correction feels familiar, or before claiming a harness has no gaps; skip when recent sessions were light or single-topic.

Reads URLs / articles / docs in parallel subagents (one per source), produces a brief of quoted excerpts with citations, then hands off to the forge skill with the right surface in mind. The brief is the artifact, not a kitchen sink and not a paraphrase; the source material never enters the main thread directly, quoted excerpts do. Use when you have reading material (URLs, papers, blog posts, internal docs, pasted content) and want the next forge step grounded in those sources rather than training-data priors; skip when training already covers the topic well.

Conducts a zoomed-out strategic review of a whole project, harness, or body of work against its actual purpose. Treats the project's own docs as testimony to test rather than authority, strips the reviewer's own context bias with clean-context checks, re-derives what the work really competes with, follows where effort actually went versus where the mission needs it, sorts hygiene from strategy, extrapolates forward from the evidence to what the work could become, actions reversible decisions instead of returning option menus, and compresses the verdict into one sentence. Use when asked to zoom out, step back, reassess what actually matters, question whether a project still serves its purpose, give a fresh-eyes or unbiased assessment, imagine what a project could become, or check the center of gravity after a long stretch of maintenance work.

Fast append to .worklog/LOG.md — log one or a few typed entries (decision/failure/gotcha/question) the way you'd write a commit message, then stop. No BRIEF regen, no graduation. Use for /worklog-minor, or when the user wants to quickly jot a failure mode, gotcha, or update for the next agent without the full /worklog pass.

Project history for agents and humans — a committed .worklog/ that any agent (Claude Code, Cursor, Codex, v0) reads and writes. BRIEF.md is a ~100-line dated orientation surface (what's decided, ruled out, unresolved); LOG.md is the append-only record behind it. Use when the user runs /worklog, asks to log or recall project history, or proposes something the log already tried and rejected — push back before agreeing.