The short answer. A skill is a folder with one text file called SKILL.md. It tells your AI how your team does one job: the steps, the rules, one good example and the format you want back. Write it once and everyone gets the same answer, in Claude, ChatGPT, Codex, Copilot, Cursor, Gemini and 40+ other tools.
01. Five People, Five Answers
Day one. Five people on your team open the same AI and type five different prompts.
Make it shorter. Add emojis. Keep it formal. Use bullet points. Sound like me.
Five answers come back. Same tool, five standards. Then someone spends the afternoon fixing them.
"Write better prompts" never fixes this, because a prompt is a request. It lives in one person's head and disappears when the chat closes. The next person starts from zero.
02. Your AI Is Unemployed
Your AI is not dumb. It already knows how to write a quote, a report or a polite reminder.
What it does not know is your job. Your prices. Your order of work. The thing you never say to a customer. Nobody wrote it down, so it guesses the job from whoever typed last.
You would never hire a person and skip the job description. With AI, almost everyone does.
Want this set up for your team? On my AI Systems Strategy Call we map your repeat jobs and pick the first skills to write. Book the call (Rs 1999)
03. What a Skill Is
Anthropic introduced Agent Skills in October 2025 and later released the format as an open standard. The idea is simple:
- A skill is a folder.
- Inside it is one file called SKILL.md: a short header plus plain instructions.
- It can also hold extras: reference files, templates, even small scripts.
The clever part is how the AI reads it. At the start it only sees each skill's name and description, about 100 tokens each. When your request matches a description, it opens that one skill and reads the full instructions. So you can keep dozens of skills around and the AI stays fast.
// what a skill looks like on disk
payment-chaser/ ├── SKILL.md the job description (required) ├── references/ optional: price list, tone guide └── scripts/ optional: small helper code
04. The 6 Lines
A job description for AI is six parts. The first two are the header, which the spec calls frontmatter. The other four are the instructions underneath, written in normal words.
| Line | What it is for | The common mistake |
|---|---|---|
| name | What you call it. Lowercase, hyphens, same as the folder name. | Capitals or spaces. The spec rejects them. |
| description | What it does and when to use it. The AI reads this to decide when to open the skill. | "Helps with invoices." Too vague, so it never triggers. |
| Steps | Your order of work, 1, 2, 3. | Skipping the step where it asks for missing info. |
| Rules | What it must always do and never do. | Rules nobody on the team agreed on. |
| Example | One real job you already did well. | No example. This one line changes the output most. |
| Output | The exact format you want back. | "Make it good." Say the sections, length and tone. |
Only name and description are real fields. Steps, rules, example and output are just headings in the body. The spec puts no format limits on the body, so write whatever helps.
// the skeleton, copy this
--- name: your-skill-name description: What it does. Use when [the words your team actually types]. --- # Your skill name ## Steps 1. 2. 3. ## Rules - Always - Never ## Example Input: Good output: ## Output [sections, length, tone]
05. Before and After
Same customer enquiry, handled two ways.
| 5 people, 5 prompts | 1 skill |
|---|---|
| Five prices, because two people guessed | Price comes from your price list, or it asks |
| Some quotes include GST, some don't | Tax handled the same way every time |
| One promises a date you can't hit | Rule: never promise a date that isn't confirmed |
| Five layouts, five tones | One layout you would actually send |
| Someone fixes them before sending | Someone checks it and sends |
The model did not get smarter. It got a job description.
06. One File, Every Tool
This is the part most people miss. Agent Skills is an open standard, so the same SKILL.md works across tools. The client list at agentskills.io shows 40+ of them, including:
- Claude and Claude Code
- ChatGPT and Codex
- GitHub Copilot and VS Code
- Cursor, Gemini CLI, JetBrains Junie
- Snowflake Cortex Code, Databricks Genie Code, Mistral Vibe, Goose
Why it matters for a business: the part you wrote, your process, is not locked to one vendor. Switch tools next year and your job descriptions come with you.
Where the file goes
| Tool | Where to put it |
|---|---|
| Claude app | Zip the folder. Customize, Skills, +, Create skill, Upload a skill. Code execution must be on in Settings, Capabilities. |
| Claude Code | ~/.claude/skills/your-skill/SKILL.md for you, or .claude/skills/ inside a project to share it. |
| ChatGPT desktop and Codex | ~/.agents/skills for you, or .agents/skills in the repo. In the ChatGPT desktop app, open Skills in the sidebar. |
Paths from each vendor's own docs, checked 28 September 2026. Other tools: see their page on agentskills.io.
07. Don't Start From Zero
Anthropic publishes a free library at github.com/anthropics/skills, 178.8k stars when I checked. Eight of the skills in it are ordinary business jobs:
- docx, xlsx, pptx, pdf: create and edit Word files, spreadsheets, decks and PDFs properly.
- internal-comms: status updates, newsletters and internal notes in a set format.
- doc-coauthoring: a structured way to write a document together with the AI.
- brand-guidelines: applies Anthropic's own brand. Use it as a pattern for writing yours.
- skill-creator: helps you write and test new skills. Start here.
In Claude Code, add the whole library with /plugin marketplace add anthropics/skills. There are bigger community libraries too, like alirezarezvani/claude-skills and wshobson/agents.
How to spot a bad skill before you trust it
- The description is vague, so you can't tell when it fires.
- No example inside.
- It runs scripts and you have not read what they do. Only install scripts from sources you trust.
The rule: start with these. Write your own only for what makes your business different.
08. The 5 Templates
These are the five jobs worth writing down first. Each one is a real SKILL.md in the published format. Swap the [brackets] for your own details, save it as SKILL.md inside a folder with the same name, and add it to your tool (section 06).
Every example inside the templates is illustrative, not a real customer.
1. Quote from an enquiry
// quote-from-enquiry/SKILL.md
--- name: quote-from-enquiry description: Turns a customer enquiry from email, WhatsApp or a DM into a ready-to-send quote using our price list and rules. Use when someone pastes an enquiry and asks for a quote, estimate, pricing or "how much". --- # Quote from an enquiry ## Steps 1. Read the enquiry. Pull out: name, what they want, quantity or scope, location, deadline. 2. If anything needed for the price is missing, stop and list the questions to ask. Do not guess. 3. Match every item to the price list below. Never invent a price. 4. Add tax and delivery using the rules. 5. Write the quote in the output format. ## Price list [Service or product] : [price] per [unit] [Service or product] : [price] per [unit] ## Rules - Prices in [currency]. Tax: [your tax rule, e.g. GST 18% shown separately]. - Quote is valid for [7] days. - Never promise a date we have not confirmed. Say "we will confirm the date". - If the job is not on the price list, write "custom quote needed" and flag it to [owner]. ## Example Input: "Hi, need 3 rooms painted in my 2BHK, can you do it by next Friday?" Good output: a quote for 3 rooms from the price list, tax shown, a note that the date will be confirmed after a site visit, and one question: "Are the walls already primed?" ## Output Greeting with their first name, a line-item table (item, qty, rate, amount), tax, total, validity, one clear next step, sign-off from [business name]. Under 150 words outside the table.
2. Weekly client report
// weekly-client-report/SKILL.md
--- name: weekly-client-report description: Writes the weekly progress report for a client from raw notes, task lists or numbers. Use when asked for a weekly report, client update, status update or "what did we do this week". --- # Weekly client report ## Steps 1. Read everything pasted in: notes, task list, numbers. 2. Sort it into: done this week, in progress, blocked, next week. 3. Pick the 3 things the client cares about most and put them first. 4. Compare numbers to last week if they are given. Never make up a number. 5. Write it in the output format. ## Rules - Plain words, no jargon. - If something is late, say it clearly with the new date and the reason in one line. - Every "blocked" item names who we are waiting on. - No number without a source in the notes. If a number is missing, write "[to add]". ## Example Input: "logo v2 sent, site homepage 70%, waiting on client photos, ads spend 12k leads 34" Good output: Wins first (logo v2 delivered, 34 leads from 12k spend), homepage at 70%, blocked on photos from the client, next week: homepage live once photos arrive. ## Output Subject: "[Client] weekly update, [date]" 1. Top 3 this week (one line each) 2. Done / In progress / Blocked / Next week (bullets) 3. Numbers (small table, only if given) 4. One question or decision we need from them Under 200 words.
3. Call notes into your CRM
// call-notes-to-crm/SKILL.md
--- name: call-notes-to-crm description: Turns messy sales or client call notes or a transcript into clean CRM fields and follow-up tasks. Use when someone pastes call notes, a call transcript or a voice note summary and wants it logged, or says "update the CRM". --- # Call notes into your CRM ## Steps 1. Read the notes or transcript. 2. Fill every field in the output. If something was not said, write "not mentioned". Never guess. 3. Pick the deal stage from our stage list only. 4. Write the follow-up tasks with an owner and a due date. 5. Flag anything that sounds like a risk: budget, timing, a competitor. ## Deal stages [New lead / Qualified / Proposal sent / Negotiation / Won / Lost] ## Rules - Use the customer's own words for their problem, in quotes. - Budget and dates exactly as said, never rounded. - Keep private or sensitive details out unless they matter for the deal. ## Example Input: "spoke to Riya, clinic owner, 2 branches, misses calls after 7pm, budget maybe 20-25k/month, wants demo next week, also talking to another agency" Good output: Contact Riya, clinic owner, 2 branches. Problem: "misses calls after 7pm". Budget 20-25k/month. Stage: Qualified. Task: book demo next week, owner [name]. Risk: comparing with another agency. ## Output Contact | Company | Problem (their words) | Budget | Timeline | Decision maker | Stage | Next step | Owner | Due date | Risks Then a 2-line summary for the team chat.
4. Support reply in your tone
// support-reply/SKILL.md
--- name: support-reply description: Drafts customer support replies in our brand tone, using our policies. Use when someone pastes a customer message, complaint, review or support ticket and asks for a reply. --- # Support reply in our tone ## Steps 1. Work out what the customer actually wants: a fix, a refund, an answer, or to be heard. 2. Check the policy below. Only offer what the policy allows. 3. If you need information to help (order number, photo, date), ask for it in one line. 4. Write the reply in our tone. ## Our tone [e.g. warm, direct, no corporate words, first name, short sentences, one emoji max] ## Policy Refunds: [your rule] Replacements: [your rule] Response time we promise: [e.g. within 24 hours] ## Rules - Say sorry once, clearly, if we got something wrong. Don't over-apologise. - Never blame the customer. - Never promise anything the policy does not allow. Escalate to [name] instead. - Angry customer: acknowledge the feeling first, then the fix. ## Example Input: "Ordered 5 days ago and still nothing. Useless." Good output: "Hi Aman, sorry, 5 days is too long. Can you send me your order number? I'll check where it is and update you today." ## Output One reply, under 80 words, ready to paste. Then one line for the team: what was offered and whether it needs escalation.
5. Payment chaser
// payment-chaser/SKILL.md
--- name: payment-chaser description: Writes polite, firm payment reminders for unpaid invoices, matched to how overdue they are. Use when asked to chase a payment, follow up on an invoice, send a reminder or when someone mentions an overdue or unpaid bill. --- # Payment chaser ## Steps 1. Get: client name, invoice number, amount, due date, days overdue, how many reminders already sent. 2. If any of these is missing, ask for it. Never guess an amount. 3. Pick the stage: - Stage 1 (1 to 7 days late): friendly nudge. - Stage 2 (8 to 21 days): firm, clear date to pay by. - Stage 3 (22+ days): final notice, next step stated. 4. Write the message for the channel asked (email or WhatsApp). ## Rules - Always include invoice number, amount and a payment link or details: [UPI / bank / link]. - Stay polite at every stage. Firm is fine, rude is not. - Never threaten anything we would not actually do. - Stage 3 names the real next step from our policy: [e.g. work paused until paid]. ## Example Input: "Sharma Traders, INV-104, Rs 18,000, due 10 Sep, 12 days late, 1 reminder sent" Good output (stage 2): "Hi Mr Sharma, a quick follow-up on invoice INV-104 for Rs 18,000, due on 10 Sep. Could you clear it by Friday? Payment details below. Thank you." ## Output Email: subject line plus body under 90 words. WhatsApp: under 50 words. Then one line: stage used and the date to follow up next.
09. What a Skill Won't Fix
A skill locks in your standard. So the standard has to be right first.
- Bad inputs. Messy notes in, messy report out. Fix what goes in.
- No owner. If nobody owns the file, it goes stale in a month.
- Rules nobody agreed on. If two managers want two tones, the skill can't pick for you.
- No review step. The skill drafts. A person still checks anything that goes to a customer or moves money.
10. Rolling It Out to a Team
Writing the file takes ten minutes. Getting a team to use one version takes a bit more thought.
- One owner per skill. They approve every change.
- One shared place. A shared project folder, a repo, or your team's Claude or ChatGPT workspace. Not twelve private copies on twelve laptops.
- One update rule. When the output is wrong, fix the file, not the chat.
- Start with one job. The one your team repeats most and argues about most.
Example math. 6 people redoing a 20 minute task 3 times a week is 360 minutes, about 6 hours a week back. That is an illustration, not a client result. Plug in your own numbers.
This is exactly what I do with teams: pick the jobs, write the first skills together, and train people to keep them sharp. See team training, or book the call below and we map it for your business.
11. FAQ
Is a skill just a saved prompt?
No. A prompt is typed every time. A skill is a file the AI opens by itself when the job matches, and it can carry examples, reference files and scripts.
Do skills work in ChatGPT?
Yes. OpenAI's docs say standalone skills work in the ChatGPT desktop app, Codex CLI and the IDE extension.
Do I need to know how to code?
No. It is a plain text file. Two header lines, then instructions in normal words.
Where do I put the file?
See the table in section 06. Claude app: upload a zip. Claude Code: .claude/skills/. Codex: .agents/skills.
Is my company data safe in a skill?
A skill is only instructions and does not send data anywhere by itself. Keep passwords and customer data out of the file, and read any script before you install it.
Want your team running these instead of re-prompting all day? Book my AI Systems Strategy Call. You walk away with your flow mapped, the first skills picked and clear next steps. Book the call (Rs 1999)
12. Sources
- Anthropic: Introducing Agent Skills. The original announcement.
- Agent Skills specification. Frontmatter fields, name rules, progressive disclosure. Checked 28 Sep 2026.
- agentskills.io. The open standard and the client list (40+ tools). Checked 28 Sep 2026.
- anthropics/skills on GitHub. The free official library, 178.8k stars on 28 Sep 2026.
- Claude Code docs: Skills. Folder paths and how skills trigger.
- Claude Help Center: Using skills in Claude. Uploading a skill in the Claude app.
- OpenAI: Build skills. Codex and ChatGPT skill paths.
- alirezarezvani/claude-skills and wshobson/agents. Community libraries.
Related guides
- The Lead Loop. AI Playbook #01: Claude watches every lead so none go cold.
- Social media skills for Claude Code. Free skill packs, one per platform.
- What is an AI agent. The plain-words version.