free-ai-lab free-ai-lab

SKILL.md writing guide

Advanced tutorial

Field guide + 10 copy-and-paste templates + advanced tips to make Skills more stable.

In one sentence: A SKILL.md is just three parts — the header "tells AI who you are and when to use it", the body is "exactly how to do it", and examples show "what the result looks like". This article breaks down every field and then gives you 10 templates you can copy directly.

For absolute beginners, start with the four-step creation tutorial; this article is the advanced reference.

Structure overview

--- name: skill-name ← lowercase kebab-case id description: what it does + when to use it ← most critical; AI judges activation by this --- # Main title (body: steps, rules, boundaries) ## Examples (input → output, shows AI what to aim for)
frontmatter (header)

name + description, in YAML format

Body

concrete steps + boundaries + rules; the more specific, the more stable

Examples

input → ideal output, anchoring the standard AI should produce

Writing the best header fields

description is the linchpin

It decides "when AI remembers this Skill". The closer it matches the user's own words, the higher the hit rate.

❌ Too vague (AI won't know when to use it)
description: Help write copy
✅ Precise (full coverage of trigger scenarios)
description: Write product-note posts in the style of viral short-form content. Use when the user says "write a post for social media", "help me draft a product tie-in piece", or pastes product info.
name naming convention

Lowercase letters + hyphens (e.g., weekly-report-writer), so its function is obvious at a glance. Don't use spaces, uppercase, or non-Latin characters.

10 copy-and-paste scenario templates

Save any of the following as a folder (containing a same-named SKILL.md) and it works — just replace the examples:

1️⃣ Meeting notes Skill
--- name: meeting-notes description: Turn meeting transcripts into structured minutes. Use when the user pastes meeting notes or says "organize today's meeting". --- # Meeting minutes 1. First identify and output: topic, time, attendees 2. Organize into three sections: "decisions / to-dos / open issues" 3. One item per person for to-dos, format: owner + what to do + due date (missing info marked [TBD]) 4. Don't add information that isn't in the notes ## Example Input: Today's meeting decided the product should add an export feature; I own the backend, Xiao Zhang owns the frontend, demo due Friday... Output: ### Decisions - Add an export feature to the product ### To-dos - Ming: backend export implementation (deadline [TBD])
2️⃣ Email polisher Skill (business tone)
--- name: email-polisher description: Rewrite rough email drafts into professional business emails. Use when the user pastes an email draft or says "polish my email". --- # Email polishing 1. Keep all key information points 2. Improve the greeting, opening, and closing etiquette 3. Split long sentences; one idea per paragraph 4. Output: polished email + change summary in 3 lines or fewer ## Example Input: Hi boss, that collaboration we discussed is confirmed, when do you want to sign? Output: Dear Mr. Wang, regarding the collaboration discussed previously, we have confirmed moving forward. Please let us know when a convenient time might be, and we can arrange the signing at your earliest convenience. Best regards!
3️⃣ Weekly report Skill
--- name: weekly-report description: Turn scattered work logs into a clear, structured weekly report. Use when the user pastes a jumble of notes or says "write my weekly report". --- # Weekly report generation 1. Sort the notes into "this week's results / problems encountered / next week's plans" 2. Order results by importance, each with supporting data (add [TBD] if missing) 3. Offer 2 suggested titles to choose from 4. Stick to facts; don't fabricate events that didn't happen ## Example…
4️⃣ Resume polisher Skill
--- name: resume-polisher description: Polish resume descriptions to be more professional and quantifiable. Use when the user pastes a resume or says "optimize my resume". --- # Resume polishing 1. Start with action verbs (e.g., "led", "implemented"), avoid first person 2. Quantify results where possible (numbers, percentages); keep the original if none 3. Keep each experience entry to 1–2 lines; cut fluff and filler 4. Output: polished list + a note on what changed and why
5️⃣ Front-end code review Skill
--- name: frontend-code-review description: Review front-end code for maintainability and accessibility. Use when the user says "review this code" or pastes code. --- # Front-end code review 1. Check each item: readability / accessibility (aria) / performance (repeated renders) / security (XSS) 2. For each issue, give: file + line, the problem, and suggested fix code 3. Sort output by severity 4. If no issues, clearly say "no issues found"
6️⃣ Image prompt generation Skill
--- name: image-prompt-writer description: Turn a scene description into a high-quality English AI image-generation prompt. Use when the user says "write an image prompt for me". --- # Image generation prompt 1. Output in English, structured: subject + environment + lighting + composition + style + quality tags 2. End with fixed quality tags: high quality, 4k 3. Automatically pick an appropriate style descriptor based on the scene (realistic / illustration / 3D / poster) 4. After the prompt, add a 2-line explanation
7️⃣ Study notes organizer Skill
--- name: study-notes description: Organize scattered study materials into structured notes and memory cards. Use when the user pastes materials or says "organize my study notes". --- # Study notes 1. Extract core concepts, opening each with a "one-sentence definition" 2. Organize hierarchically by topic, adding examples 3. Generate 5 Q/A memory flashcards 4. Output Markdown suitable for importing into a notes app
8️⃣ Data analysis interpreter Skill
--- name: data-interpreter description: Interpret a dataset (table/numbers) and give insights and recommendations. Use when the user pastes data or says "analyze this data". --- # Data interpretation 1. Summarize the overall picture first, then list notable trends/outliers 2. Distinguish "data facts" from "inferences"; label any inferences 3. Give 3 or more actionable next-step suggestions 4. Output comparisons in table form
9️⃣ Short-video script Skill
--- name: short-video-script description: Turn a topic into a 30–60 second spoken short-video script. Use when the user says "write a short-video script" or "turn this copy into a video script". --- # Short-video script 1. Structure: 3-second hook + 15-second core + closing call-to-action 2. Make it spoken: short sentences, casual tone, pause markers 3. Give a visual cue every 10 seconds (shot / subtitle) 4. Keep the runtime under 45 seconds, around 300 words
🔟 Travel planning Skill
--- name: travel-planner description: Output a travel itinerary based on duration/budget/preferences. Use when the user says "help me plan a trip" or "make a travel guide". --- # Travel planning 1. First confirm the inputs: number of days, budget, origin and destination, preferences (food / nature / culture) 2. Schedule each day morning/noon/evening: attractions + transport + dining 3. Mark the expected stay time at each attraction 4. Include: a budget estimate table + 3 practical tips (like booking/weather)

The templating mindset: turning "high-frequency repeated workflows" into a Skill is a more advanced use than bookmarking prompts — install once, and AI follows your standard every time afterward.

Advanced tips: making Skills more stable

  • Add "counter-examples": writing "don't do X" in the body noticeably reduces overstepping (e.g., "don't fabricate data", "don't replace names when uncertain")
  • Specify the output format: explicitly demand "table form", "### headings", "3 lines or fewer" — AI's output structure becomes more stable
  • Anchor with examples at the end: examples aren't just for show; they're the reference that aligns AI to the production standard
  • Use scripts/ for mechanical work: hand repeated computation and formatting to scripts; SKILL.md only "directs"
  • Polish iteratively: test 3 inputs → add where AI erred into the body → retest; Skills get more accurate the more you use them

Related reading