cursor/plugins/local/cheese-grok/skills/design-doc/SKILL.md
Use when the user wants to draft, revise, critique, or shepherd a design document, RFC, ADR, tech spec, or design proposal. Triggers on "write a design doc", "draft an RFC", "ADR for X", "design proposal", "tech spec", "review my design doc", "tighten this design doc", "critique my draft". Drives a spine-first, anti-slop workflow where YOU author Context / Problem / Goals / Non-Goals / Alternatives / Risks and the AI is restricted to critique, expansion, and copyedit. Use even when the user says "help me write a design doc" — the help is procedural, not authorial.
npx skillsauth add paulnsorensen/dotfiles design-docInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
You are about to help the user produce a design doc that a senior reviewer will respect. The default failure mode of AI-assisted design docs is silent fluency without ownership — generic prose that fills sections without saying anything specific. Your job is to make that impossible.
You may not propose, draft, or pre-fill Goals, Non-Goals, Alternatives, or Risks. You may only critique what the user wrote.
If the user asks you to "fill in" any of those sections, refuse and re-route them to the spine pass. The full list of forbidden authorship moves:
You may suggest the kind of thing missing (e.g. "you have no risk about queue backpressure — is that intentional?") but the user writes the substance.
Direct the user to docs/designs/<slug>.md using
templates/00X-template.md. They write, bullet-point style, fine:
Refuse to start Pass 2 until the spine exists.
The user pastes the draft. You apply this exact four-step critique, in order:
Forbidden in Pass 2: producing a "polished version", writing "here is an improved draft", or filling in any spine sections.
The user requests one specific thing per ask: a named technique, a
benchmark number, a citation, a code reference (use
codebase_search / mcp__serena__find_symbol). Never "make this section
better" — only specific facts. Push back if the user asks for prose.
The user pastes the doc. You return a numbered list of every hedging
phrase with line numbers: perhaps, might, could potentially,
generally, in many cases, it's worth noting, it's important to,
probably, arguably, relatively, somewhat. Do not rewrite. The
user decides which to delete.
Tell the user to draw one diagram by hand or in Mermaid. The act of drawing is the comprehension test. If they ask you to draw it, refuse: "If you can't draw it, the design isn't done."
You may help them format a hand-drawn description into valid Mermaid syntax — that's transcription, not authoring.
Confirm the doc has:
Offer to open a draft PR (via gh pr create) if applicable. End.
If the user says "just write it for me" or "give me a draft", restate Rule Zero and explain why: AI-authored spine sections are the #1 reason design docs read as generic. Offer to help with Pass 2–5 instead.
If they insist after one explanation, comply — but flag explicitly what
you authored so a reviewer can audit it later. Add a line at the bottom
of the doc: > §Goals / §Risks drafted by AI on YYYY-MM-DD — review with suspicion.
## Slop sentences, ## Spine integrity,
## Weakest paragraph (3 questions), ## Missing risk four-section
reply.For the template, see templates/00X-template.md.
tools
Reconstruct what a past coding-agent session was doing so you can resume it — goal, files touched, last verified state, and the next step — by querying the session logs. Use when the user says "what was I working on", "recover that session", "reconstruct where I left off", "resume my last session", "what did that session change", "rebuild context from logs", or invokes /work-recovery. Report-only — it never scores or judges. Do NOT use for usage scoring (that is /skill-improver, /tool-efficiency, /prompt-analytics) or one-off interactive log queries (that is /session-analytics).
development
Curate this repo's hallouminate wiki (.hallouminate/wiki/, the repo:dotfiles:wiki corpus) — add or update architecture pages, per-harness docs, and gotchas. Use when the user says "update the wiki", "document this in the wiki", "refresh the harness docs", "add a wiki page", "curate the wiki", "the wiki is stale", or invokes /wiki-curator. Also use at session end to write back a non-obvious decision or gotcha worth preserving. Grounds the existing wiki first, follows one-topic-per-file conventions, verifies every external doc URL before writing, and reindexes. Do NOT use for general code search (that is cheez-search) or for editing AGENTS.md command reference.
tools
Audit how a tool, command, or MCP server is actually used across coding-agent sessions and produce calibrated recommendations — tool-vs-task fit, error forensics, fix recommendations, permission friction, MCP health, and token economics. Use when the user says "tool efficiency", "am I using X efficiently", "audit tool usage", "why does X keep failing", "how do I fix this error", "what should I change", "permission friction", "is this MCP worth it", "tool error rate", "fix recommendations", or invokes /tool-efficiency. Do NOT use for auditing a skill or agent definition (that is /skill-improver) or for one-off interactive log queries (that is /session-analytics).
tools
Analyze how prompts and skill routing behave across coding-agent sessions and produce calibrated recommendations — prompt-pattern analysis, routing accuracy, and knowledge gaps. Use when the user says "analyze my prompts", "prompt patterns", "is routing working", "which skill should have fired", "knowledge gaps", "what do I keep asking", or invokes /prompt-analytics. Do NOT use for auditing a single skill/agent definition (that is /skill-improver), tool/MCP efficiency (that is /tool-efficiency), or one-off interactive log queries (that is /session-analytics).