I have read a lot of SKILL.md files this year. Good ones, broken ones, ones that were clearly a ChatGPT prompt pasted into a file with a new name. The gap between a skill that works and one that quietly does nothing usually comes down to structure, which is why searching for skill.md examples and templates is the right instinct. You learn this format faster by copying a working file and editing it than by reading any specification. So this article gives you exactly that: the anatomy every skill shares, five worked examples for jobs people actually do, a blank template you can copy into a text editor right now, and the mistakes I keep seeing in files people share online.
Quick context if you are new to this. A SKILL.md file is a plain Markdown document that teaches an AI assistant a repeatable procedure. Claude reads it natively through its Skills feature, and the same file drops into a ChatGPT Project or a Gemini Gem as instruction text. Write once, use in all three. If you want the full beginner explanation, I wrote one in What is a SKILL.md file. This piece assumes you know roughly what a skill is and want working material to start from.
The anatomy every good skill shares
Strip away the subject matter and almost every effective skill file has the same five parts. The order matters less than the presence. When one of these is missing, you can usually predict how the skill will fail.
| Section | What it does | What happens without it |
|---|---|---|
| Frontmatter (name + description) | Tells the AI when to activate the skill | The skill never triggers, or triggers on everything |
| Role and stance | Sets who the AI is being and what it optimizes for | Generic assistant voice, hedged answers |
| Numbered procedure | The steps, in order, with decision points | The AI improvises a different process every time |
| Output format | Exact structure of the deliverable | You reformat the output by hand, every time |
| Rules and boundaries | What to always do, what to never do | The failure modes you were trying to prevent come back |
The frontmatter deserves one extra note because it is the part people get wrong most often. In Claude, the description field is what the model reads when deciding whether your skill applies to the current request. "Helps with meetings" is too vague to trigger reliably. "Converts raw meeting notes into a structured summary with owners, deadlines, and a draft follow up email" triggers exactly when it should. Write the description like you are telling a new hire when to use this playbook.
Five skill.md examples and templates you can copy
These are compressed versions of real skills, one per job. Each shows the skeleton so you can see how the five parts map onto an actual use case. Expand any of them with your own vocabulary, your own edge cases, and your own examples and they get better fast.
1. Meeting notes to actions
---
name: meeting-notes-actions
description: Converts raw meeting notes or transcripts into a
structured summary with decisions, action items with owners
and deadlines, and a draft follow up email.
---
You are a chief of staff who turns messy notes into accountability.
## Procedure
1. Read the full notes before extracting anything.
2. Identify decisions made. A decision has an outcome, not a discussion.
3. Extract action items. Each needs an owner and a deadline.
If either is missing from the notes, flag it as UNASSIGNED
rather than inventing one.
4. Note open questions that were raised but not resolved.
5. Draft a follow up email under 150 words.
## Output format
**Decisions:** numbered list
**Actions:** table with owner, task, deadline
**Open questions:** bullet list
**Draft email:** plain text, no subject line theatrics
## Rules
- Never invent owners or dates that were not stated.
- Keep the summary under one page regardless of meeting length.
This one is the exact skill we give away free as a fully built file, with the edge cases and examples filled in: the free starter skill. If you want to see what a production version of the skeleton above looks like, that is the fastest way.
2. Objection handling prep
---
name: objection-prep
description: Prepares a salesperson for a specific upcoming
conversation by generating the likely objections, the
underlying concern behind each, and a tested response path.
---
You are a negotiation coach. You do not give scripts to recite.
You prepare the user to think under pressure.
## Procedure
1. Ask for: the deal, the person, what they likely fear.
2. Generate the 5 most probable objections for THIS deal,
not generic sales objections.
3. For each: name the underlying concern, then give a
label-and-reframe response, then the follow up question
that moves the conversation forward.
4. End with the one objection the user is most likely to
handle badly, and why.
## Rules
- Every objection must be specific to the context given.
- No response longer than 3 sentences. Real conversations
do not have room for paragraphs.
The step that does the work here is the last one in the procedure. Asking the AI to predict where you personally will stumble turns a generic prep document into coaching. The full version of this thinking, with pressure testing, counterparty modeling, and live roleplay modes, is what Meta Negotiator is built around.
3. Proposal and closing email writer
---
name: closing-email
description: Writes the follow up email after a sales
conversation, structured to advance the deal to a
concrete next step rather than to summarize politely.
---
You write emails that close. Warm, direct, never desperate.
## Procedure
1. Ask what was discussed, what was agreed, what the
ideal next step is.
2. Open with the specific thing that mattered to them,
not with "great speaking with you."
3. Confirm agreements in one line each.
4. Propose ONE next step with a date. Not options. One.
5. Keep the whole email under 120 words.
## Rules
- Ban these phrases: "just checking in", "touching base",
"hope this finds you well", "any thoughts?"
- If no next step was discussed, say so and ask the user
what they want it to be. Do not write a vague email.
Notice the banned phrase list. This is one of the highest leverage things you can put in any writing skill, because the model's default register is exactly the register you are trying to escape. Meta Closer extends this into the whole closing sequence: proposal structure, pricing conversations, silence handling, and the awkward final week of a deal.
4. Persuasion review
---
name: persuasion-review
description: Reviews a draft (email, landing page copy, pitch,
post) and rewrites it for persuasive force using named
principles, showing what changed and why.
---
You are a persuasion editor. You explain your edits so the
user learns the principle, not only the fix.
## Procedure
1. Read the draft. State in one line what it is asking
the reader to do.
2. Score it 1 to 10 on: clarity of ask, cost of saying yes,
proof, and specificity.
3. Rewrite it. Change as little as possible while fixing
the lowest scores first.
4. List each change with the principle behind it in
one sentence.
## Rules
- If the draft has no clear ask, stop and say so before
rewriting anything.
- Never add fake urgency or invented social proof.
The scoring step matters more than the rewrite. It forces the model to diagnose before it prescribes, which is the difference between an editor and an autocomplete. This skeleton comes from the same family as Meta Persuader, which adds audience modeling and the ethical guardrails that keep persuasion on the right side of manipulation.
5. Strategy memo builder
---
name: strategy-memo
description: Turns a business question into a one page
decision memo with options, tradeoffs, a recommendation,
and the evidence that would change it.
---
You are a strategist who writes for decision makers with
four minutes of attention.
## Procedure
1. Restate the decision in one sentence. Confirm it.
2. Lay out 3 real options. "Do nothing" counts as one.
3. For each: cost, upside, main risk, reversibility.
4. Recommend one. Say what evidence would change your mind.
5. One page. Always one page.
## Rules
- No option may be a strawman built to make another look good.
- The recommendation must name a first step doable this week.
Rule one in that file exists because models love strawmen. Give them three options to generate and, unprompted, two will be filler. Naming the failure mode in the rules section is how you prevent it. Meta Strategist is the full sized version, with market analysis frames and pre-mortem routines built in.
A blank template to start from
If none of the five fit your job, start here. This is the minimum viable skill, and it is deliberately boring:
---
name: your-skill-name
description: One or two sentences saying exactly what this
does and when it should activate. Be specific enough that
it would NOT trigger on unrelated requests.
---
You are [role]. You optimize for [the thing that matters].
## Procedure
1. First, gather: [what inputs you need from the user].
2. Then: [the actual steps, in order].
3. [Include at least one decision point: "if X, do Y."]
## Output format
[Exact structure. Sections, lengths, what to include and skip.]
## Rules
- Always: [the behavior that must happen every time]
- Never: [the failure mode you are guarding against]
- Ban these phrases: [the cliches of your domain]
## Example
[One worked input and output pair. This single section
improves consistency more than anything else you can add.]
Save it as SKILL.md in its own folder, drop the folder into Claude's skills directory or upload it through Settings, and iterate. In ChatGPT, paste the body into a Project's custom instructions. In Gemini, paste it into a Gem. The frontmatter block only does its automatic triggering work in Claude, but the rest of the file carries everywhere.
Where most homemade skills go wrong
Three failure patterns account for nearly every broken skill file I get sent. The first is the vague description, covered above, where the skill either never fires or fires constantly. The second is writing goals instead of procedures. "Provide insightful analysis" is a wish. "Score the draft on these four dimensions before rewriting" is a procedure. Models follow procedures. The third is skipping the example section. A single worked input and output pair anchors the model harder than three paragraphs of description, because it shows the standard instead of describing it. If your skill produces inconsistent output, add an example before you touch anything else.
There is also a subtler failure worth naming: the skill that tries to do four jobs. A file that handles meeting notes and email drafting and project planning does all three at the level of a general prompt, which defeats the point. One skill, one job. If you find your file growing multiple unrelated procedures, split it. This is also why multi skill bundles exist as separate files that hand off to each other rather than one giant document. Our Vault bundle is structured exactly that way for that reason: every skill in it does one thing, and you install the ones the current job needs.
FAQ
What is a SKILL.md file?
It is a Markdown text file containing named, structured instructions that teach an AI assistant a repeatable task: a role, a step by step procedure, an output format, and rules. Claude loads these natively as Skills and applies them automatically when relevant. The same content works as instructions in ChatGPT Projects and Gemini Gems.
How do I create a SKILL.md file?
Create a folder, put a file named SKILL.md inside it, and write frontmatter (name and description) followed by your instructions in plain Markdown. Any text editor works. Start from the blank template above, keep the first version short, test it on a real task, and add rules as you notice failures.
Are Claude skills the same as custom GPTs?
They solve the same problem with different plumbing. A custom GPT is configured inside ChatGPT and lives there. A skill is a portable file: you own it, you can version it, and you can run the same file in Claude, a ChatGPT Project, or a Gemini Gem. If you work across platforms, the file wins.
What is the difference between a skill and a prompt?
A prompt is something you type each time and lose when the chat ends. A skill is stored procedure: it persists, activates when relevant, and produces consistent output because the structure does not depend on how well you phrased things today. Prompts are fine for one-off questions. Anything you do weekly deserves a skill.
Where can I download ready made SKILL.md templates?
Anthropic publishes open source examples in its GitHub skills repository, which are worth reading to see the format in production. For skills aimed at commercial work, we build and sell finished ones in our full catalog, and the free meeting notes skill below costs nothing and installs in about two minutes.
Want a finished file instead of a template? Grab our free Meeting Notes to Actions SKILL.md and see the format working before you write your own.
Leave a comment: