latestaiagents/skill-frontmatter

Skill Frontmatter

The frontmatter is what the model reads to decide whether to load your skill. Get it wrong and the skill never fires.

When to Use

• Authoring a new SKILL.md
• Debugging "my skill isn't activating"
• Reviewing a PR that adds a skill

Required Fields

---
name: kebab-case-name
description: |
  Sentence 1: what the skill does.
  Sentence 2: when to use it (specific trigger conditions).
  Sentence 3: activation keywords (phrases users say).
---

That's it. name and description. Nothing else is required by most clients.

The name Field

• Kebab-case — merge-conflict-surgeon, not MergeConflictSurgeon
• Matches directory name — directory holding the SKILL.md must have the same name
• Verb-led or role-led:
  • Action: resolve-merge-conflict
  • Role: owasp-auditor
  • Topic: mcp-server-authoring
• Unique across your skill set — two skills with the same name collide

The description Field

This is THE activation signal. Model reads it and matches against the user's message.

Three-sentence structure

1. What — one sentence on the skill's purpose (action-first)
2. When — specific trigger conditions
3. Keywords — phrases users actually say

description: |
  Systematic merge conflict resolution with context analysis and safe resolution.
  Use this skill when the user has merge conflicts, needs help resolving conflicting changes, or is stuck on a rebase.
  Activate when: merge conflict, conflict resolution, git merge failed, rebase conflict, CONFLICT markers.

Length

• Minimum: 2 sentences. Shorter = under-specified; misses activations
• Sweet spot: 3-5 sentences, ~60-150 words
• Maximum: don't exceed ~200 words; longer dilutes the signal

Voice

• Second-person or imperative
• Specific, not abstract: "when user is debugging a null pointer in Java" beats "helps with bugs"
• Name the tools/frameworks — tsc, Next.js, LangGraph etc. Users and models match on these

Writing Activation Keywords

Keywords are the easiest win. Think about what the user might literally type:

Bad:

Activate when: development, coding, software.

Too vague. Matches everything → matches nothing well.

Good:

Activate when: git merge failed, CONFLICT markers, resolve conflicts, rebase conflict, auto-merge failed.

Specific phrases, variants, error strings users actually see.

Include:

• Domain jargon (prompt caching, OAuth 2.1, RAG)
• Error messages (null pointer, CORS error)
• Tool names (@modelcontextprotocol/sdk, LangChain)
• Task verbs (deploy, debug, refactor)

Common Mistakes

1. Vague description — "helps with code" doesn't match anything specifically
2. Only describing WHAT, not WHEN — missing trigger conditions
3. Keywords no user would type — "Activate when: utilities, helpers"
4. Name ≠ directory — name: foo but directory is bar → skill may not load
5. Multiline string without | — YAML parses as a single collapsed line, loses formatting
6. Typos in YAML — wrong indentation, missing ---, unquoted colons
7. Duplicate descriptions — 5 sibling skills with identical descriptions collide

Validating YAML

Every CI run should parse your frontmatter:

# Quick check
find skills -name SKILL.md -exec sh -c \
  'head -n 20 "$1" | grep -A 20 "^---$" > /dev/null || echo "BAD: $1"' _ {} \;

Or a proper script:

import matter from "gray-matter";
import { readFileSync } from "fs";
const { data } = matter(readFileSync(path, "utf-8"));
if (!data.name || !data.description) throw new Error(`Missing fields: ${path}`);
if (data.description.length < 50) console.warn(`Short description: ${path}`);

Worked Example — Refining a Frontmatter

Draft 1 (bad)

name: test-generator
description: Generates tests.

Too short. No activation signal. Model never picks it.

Draft 2 (better)

name: test-generator
description: Generate unit tests for JavaScript and TypeScript code.

Better, but still missing WHEN and keywords.

Draft 3 (good)

name: test-generator
description: |
  Generate unit and integration tests for JavaScript/TypeScript code using Jest or Vitest.
  Use this skill when the user asks for tests, has added untested code, or needs to improve coverage.
  Activate when: write tests, add tests, test coverage, generate test cases, Jest, Vitest, missing tests.
---

Now the model has enough signal to activate confidently.

Best Practices

1. Write the description from the USER'S angle — what they'd say, not what you'd call your skill
2. Include domain keywords users actually type (errors, tool names, jargon)
3. Keep total length under 200 words; concentrate signal
4. Validate YAML in CI — typos waste hours of debugging
5. Re-read after writing: if you can't tell when this skill should fire, neither can the model
6. Audit for duplicate descriptions across related skills — they should differentiate