writing-great-skills
GitHub指导编写高质量 Agent Skill (SKILL.md) 的指南。涵盖元数据、触发词设计、输出契约及质量检查，确保 AI 能准确触发并生成专业成果，避免骨架代码和模糊指令。
Trigger Scenarios
Install
npx skills add mohitagw15856/pm-claude-skills --skill writing-great-skills -g -y
SKILL.md
Frontmatter
{
"name": "writing-great-skills",
"description": "Author a high-quality Agent Skill (SKILL.md) that an AI reliably triggers and executes well — strong frontmatter, a sharp description with trigger phrases, a clear output contract, quality checks, and anti-patterns. Use when asked to write a skill, create a SKILL.md, improve a skill, review a skill for quality, or contribute to a skills library. Produces a complete, SkillCheck-passing SKILL.md plus a short rationale for the key choices."
}
Writing Great Skills Skill
A skill is a promise: given this kind of request, produce this kind of professional output, every time. The best SKILL.md files win on two things — the model triggers them at the right moment, and once triggered it produces the right artifact without hand-holding. This skill helps you write one that does both.
Working from a brief
Given a rough idea ("a skill for writing changelogs"), produce the full SKILL.md anyway — infer the deliverable, inputs, and structure, and mark genuinely open choices. Never hand back a skeleton with <!-- TODO --> left in; fill them.
Required Inputs
Ask for (if not already provided), else infer and label:
- What the skill should do and the concrete artifact it produces
- When it should trigger (the phrasings a user would actually type)
- The inputs it needs from the user
- Any framework or standard it encodes (for attribution)
The anatomy of a great SKILL.md
1. Frontmatter (this is what gets your skill found)
---
name: kebab-case-name # matches the folder; short, specific
description: "<one rich sentence>"
---
The description is the most important line in the file — it's all the model sees when deciding whether to load the skill (progressive disclosure: only names + descriptions are in context until one is invoked). A strong description has three parts:
- What it does + the concrete deliverable.
- A "Use when …" trigger clause listing the real phrasings ("Use when asked to write a postmortem, do a root-cause analysis, or document an incident").
- A "Produces …" clause naming the output ("Produces a blameless postmortem with timeline, root cause, and action items").
Write triggers the way users speak, not the way you'd categorise the skill. Cover synonyms.
2. One-line value statement
Open the body with a single sentence on the value, in the voice of a senior practitioner.
3. Working from a brief
State that the skill delivers a complete artifact even with thin input — infer and label assumptions, never leave bracketed placeholders, never refuse for missing context. This is what separates a skill that works from one that nags.
4. Required Inputs
A short list of what to ask for — and an instruction to proceed with labelled inferences if they're missing.
5. Output Format / Structure
The heart of the skill: a concrete template — real headings, tables, and sections — of the final artifact. Show the shape, don't describe it abstractly. This is where most of the quality lives.
6. Quality Checks
A short checklist the output must satisfy (the rubric a reviewer would apply). Make them observable.
7. Anti-Patterns
The specific failure modes to avoid — the lazy or generic outputs a weaker model would produce.
Process
- Nail the deliverable in one sentence before writing anything else.
- Write the description and stress-test the triggers ("would the model pick this over a neighbouring skill?").
- Draft the Output Format as a real template.
- Add Quality Checks and Anti-Patterns that target this skill's specific failure modes.
- Validate:
npm run skillcheck(structure) and run it against a thin brief to confirm it doesn't beg for inputs.
Output Format
Return:
- The complete SKILL.md in a fenced block, ready to save to
skills/<name>/SKILL.md. - A 3–5 bullet "why this works" note: the trigger phrases chosen, the deliverable, and the sharpest anti-pattern it guards against.
Deeper Materials
This skill ships with support files — use them when they are available:
references/description-engineering.md— Description Engineering: the 300 Characters That Decide Everything. Apply it while producing the output; it carries the calibration and judgment calls the method summary above compresses.templates/skill-scaffold.md— a fill-in version of the deliverable with the quality gates inline. Offer it when the user wants to work the document themselves rather than have it generated.
Scoring Rubric (0–40)
Score any output of this skill before handing it over; 32+ is ship-quality.
| Dimension | 0 | 5 | 10 |
|---|---|---|---|
| Description & trigger engineering | Description says what the skill is about, with no trigger phrases or named deliverable | Has "Use when…" and "Produces…", but triggers are category labels, not phrasings users actually type | All three parts present; triggers cover the real synonyms users would say, and are distinct enough that the model wouldn't confuse this skill with a neighbour |
| Output contract concreteness | Output Format describes the artifact ("a structured report with sections") | A partial template — some real headings, but key sections still described abstractly | A full template with real headings, tables, and field formats — two independent runs would produce recognisably the same product |
| Guardrail specificity | Quality checks are unobservable ("output should be clear"); anti-patterns missing or generic | Checks are observable but generic to any document; anti-patterns don't name this skill's failure modes | Every check is verifiable at a glance, and each anti-pattern names a specific lazy output a weaker model would actually produce for this artifact |
| Thin-brief resilience | The skill begs for inputs or would return a skeleton full of [placeholders] |
Working-from-a-brief section exists, but there's no instruction to label inferences, so gaps get silently invented | Explicitly instructs: build the complete artifact anyway, infer and label assumptions, never leave TODOs — and the inputs list says what to ask for only when it changes the output |
Quality Checks
-
nameis kebab-case and matches the intended folder - Description states what it does, has a "Use when …" trigger clause, and names what it Produces
- Body has: value line, working-from-a-brief, inputs, a concrete Output Format template, Quality Checks, Anti-Patterns
- No
TODO/placeholder text left in - Triggers are distinct from neighbouring skills (won't mis-fire or get skipped)
- Would pass
npm run skillcheckwith no errors
Anti-Patterns
- A vague description with no trigger phrases — the skill never gets picked
- An Output Format that describes the artifact instead of templating it
- Quality Checks that aren't observable ("output should be good")
- Leaving
<!-- TODO -->or[bracketed]placeholders in the final file - Overlapping so heavily with an existing skill that the model can't choose between them
Version History
- 54fad50 Current 2026-07-19 12:57
- a38bc30 2026-07-05 11:19
