Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Author standardized AI cheatsheets in the awesome-cheatsheets format
inputs
[{"name":"topic","type":"string","required":true,"description":"The tool, concept, or MCP server to create a cheatsheet for"},{"name":"draft_content","type":"string","required":false,"description":"Optional raw notes or outline from the user"}]
outputs
[{"name":"cheatsheet_md","type":"string","description":"A fully-formed cheatsheet .md file ready for validation"}]
triggers
["/cheatsheet-scribe","create cheatsheet","write a cheatsheet about"]
Cheatsheet Scribe
Scaffold and author standardized AI cheatsheets that conform to the template contract defined in tools/template-contract.ts.
When to use
Trigger this skill when a user wants to create a new cheatsheet for:
An AI tool (CLI/TUI agents, code assistants)
An MCP server or protocol
A conceptual method or workflow
A comparison between tools/approaches
Workflow
Step 1: Gather requirements
Ask the user for:
Topic — What tool/concept is this about?
Target audience — Who is this cheatsheet for?
Draft content — Any existing notes, docs, or outlines?
If the user provides just a topic with no draft, proceed to Step 2 with an empty canvas.
SAFETY: No Fabrication
The skill MUST NOT fabricate references, commands, or facts not present in:
The user's draft content
Official documentation the user has provided
Answers explicitly given by the user
Concrete examples of fabrication (NEVER do these):
Inventing a CLI flag like --force that doesn't exist
Citing a source or URL that wasn't provided
Assuming a default value or behavior without confirmation
Creating a "Quick Command Reference" entry for a command you haven't verified
When in doubt, ASK the contributor rather than guess. If the user cannot provide the information, mark the section as incomplete rather than inventing content.
Step 1.5: Check for existing cheatsheet
Before proceeding to drafting, derive the slug from the topic (kebab-case). Check if src/content/cheatsheets/<slug>/ already exists in the repository.
If the cheatsheet already exists:
Ask for explicit confirmation: "A cheatsheet for '' already exists. Do you want to overwrite it? Type 'yes, overwrite' to confirm."
Do NOT proceed with Step 2 until explicit confirmation is received
If the user declines, abort gracefully
If the cheatsheet does not exist: Proceed to Step 2.
Step 2: Apply the template contract
Every cheatsheet must follow template-contract.md:
Frontmatter (required):
slug: kebab-case identifier (e.g., hermes-agent)
title: Human-readable title
category: One of tool, mcp, concept, comparison
subcategory: Optional free-form
summary: 20-400 character sentence
last_updated: ISO-8601 date (YYYY-MM-DD)
stale_after_days: 1-730
tags: 1-12 kebab-case tags
status: poc, published, or deprecated
authors: Optional author array
links: Object with required homepage, optional repo and others
Section structure (locked order):
One-liner: **One-line:** <summary> (bold, before first H2)
## Reference — collapsed <details> block with sources
Step 3: Draft the cheatsheet
Use the canonical Hermes Agent example (examples/hermes-draft.md → examples/hermes-expected.md) as a reference for:
Tone: actionable, concise, expert-level
Structure: locked 7-section format
Installation: copy-paste one-line installer block at the top
Frontmatter: all required fields populated
Generate a first draft incorporating:
User's draft content if provided
Research from the tool's official docs
Best practices from the community
Pi retro heuristics (from the first non-Hermes run):
If the issue or draft names a concrete slug/path, prefer that over an inferred slug and record the decision in Review Questions when it conflicts with fixture metadata.
If a draft offers two filenames or concepts (AGENTS.md vs SYSTEM.md, core feature vs extension), pick only the version verified by provided docs; otherwise flag it as a Review Question instead of blending both.
Extract every verified executable command from the draft, dedupe it, then group commands by goal in the Quick Command Reference table.
Keep the local loop fast with pnpm cheatsheet:lint <file> --no-links; run the full link check before PR creation when network conditions are reliable.
Step 4: Validate against contract
After drafting, call the validator to check conformance:
import { validate } from"tools/validator";
// After writing the draft to a temp fileconst result = awaitvalidate("/path/to/draft.md");
if (!result.ok) {
// Surface every failure with its rule nameconsole.log("Validation failed:");
for (const e of result.errors) {
console.log(` - ${e.rule}: ${e.message}${e.line ? ` (line ${e.line})` : ""}`);
}
// Ask contributor to fix before proceeding
}
Rules invoked by validator (single source of truth):
Do NOT duplicate validation logic in this skill — the validator decides, the skill orchestrates.
Step 5: Output
Return the complete .md file content with:
Frontmatter block at top
Proper heading hierarchy
Code blocks with language hints where applicable
Step 6: Review Questions
After validation passes, emit 2–4 specific review questions tied to ambiguous or opinion-driven parts of the draft.
Requirements:
Questions MUST be content-specific, not generic ("looks good?", "is this okay?" are DISALLOWED per PRD M2 bullet 1)
Each question should target a specific decision point where contributor input matters
Analyze the draft for these ambiguity types:
Naming ambiguity — unclear slug, title, or tag choices
Scope ambiguity — whether a section covers too much or too little
Technical ambiguity — command correctness, version currency, best practice validity
Audience ambiguity — whether content assumes the right skill level
Completeness ambiguity — missing edge cases, alternative workflows, or reference links
Example valid questions (NOT "looks good?"):
❌ "Does this look good?"
❌ "Is the cheatsheet ready?"
✅ "The slug ollama-run vs ollama-cli — which is more widely recognized in the community?"
✅ "The Installation block shows the curl one-liner — is there a Homebrew formula or apt repo I should add as an alternative?"
✅ "The run commands reference v0.1.2 — should I note that v0.1.3 is the latest stable?"
✅ "This cheatsheet assumes Linux. Should I add a note for macOS/Windows users?"
Output format:
## Review Questions
1. [specific question about naming/slug/tags]
2. [specific question about technical accuracy]
3. [specific question about scope/completeness]
4. [specific question about audience/assumptions]
Please confirm each item or suggest alternatives before I finalize.
Wait for contributor response before proceeding. If they request changes, iterate on the draft and re-validate before re-presenting questions.
Step 7: Commit Message & PR Title
On contributor approval of all review questions, suggest:
