Cheatsheet Scribe

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:

1. Topic — What tool/concept is this about?
2. Target audience — Who is this cheatsheet for?
3. 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):

1. One-liner: **One-line:** <summary> (bold, before first H2)
2. ## Installation — copy-paste one-line installer(s) + first-run flow
3. ## Step-by-Step Setup & Optimization — numbered steps
4. ## Best Practices — bullets
5. ## Quick Command Reference — code blocks
6. ## Expected Outcomes — what the user gets
7. ## 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 file
const result = await validate("/path/to/draft.md");

if (!result.ok) {
  // Surface every failure with its rule name
  console.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):

• frontmatter-missing, frontmatter-yaml-invalid, frontmatter-schema
• section-missing, section-out-of-order
• one-liner-missing, one-liner-too-short
• reference-details-missing
• mermaid-missing-in-mental-model, mermaid-fence-broken, mermaid-empty
• link-broken (optional, skip via skipLinks: true)

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:

1. Naming ambiguity — unclear slug, title, or tag choices
2. Scope ambiguity — whether a section covers too much or too little
3. Technical ambiguity — command correctness, version currency, best practice validity
4. Audience ambiguity — whether content assumes the right skill level
5. 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:

1. Conventional commit message (lower-case, imperative):

   docs(cheatsheet): add <slug> cheatsheet for <tool/concept>
   

   Examples:

   • docs(cheatsheet): add ollama-cli cheatsheet for local LLM inference
   • docs(cheatsheet): add hermes-agent cheatsheet for autonomous AI agents
   • docs(cheatsheet): add mcp-sequencer cheatsheet for MCP protocol workflows

2. PR title (same as commit message, can be more descriptive):

   docs(cheatsheet): add <slug> cheatsheet — <one-line summary>
   

   Example:

   docs(cheatsheet): add ollama-cli cheatsheet — local LLM inference from CLI
   

3. PR description (recommended, optional for contributor):

   ## Summary
   <summary from frontmatter>
   
   ## Sections
   - Installation: <one-line installer summary, platforms covered>
   - Step-by-Step: <n> steps
   - Best Practices: <n> items
   - Commands: <n> code examples
   
   ## Validation
   - [x] Passes template-contract.md validation
   - [x] Review questions addressed
   

DO NOT stage, commit, or push — those steps stay with the human contributor.

Output the commit message and PR title as a clearly formatted block the contributor can copy.

Examples

See examples/ for input/output pairs:

• hermes-draft.md → hermes-expected.md (canonical reference)

Dependencies

• Consumes tools/validator — the single source of truth for all validation rules
• Output must pass tools/validator before merge
• Section list in this skill MUST drift-test against template-contract.md

See Also

• Contributor Tutorial — 30-minute walkthrough for new contributors: install Claude Code → invoke /cheatsheet-scribe → answer review questions → open PR
• Template contract — structural rules every cheatsheet must obey