Skills, explained
What is a Claude skill?
A folder of markdown that teaches Claude how to do one kind of work properly, and loads itself whenever that work comes up. No plugin, no API, no account — files on a disk.
Most people meet a model's weak spots one prompt at a time. You ask for a cold email, you get something that reads like a cold email, and you spend four rounds dragging it toward something you would actually send. Next week you start from the same place, because nothing you learned in those four rounds was written down anywhere the model can see.
A skill is where you write it down. It is a directory containing a
SKILL.md and however many supporting files the subject needs.
The frontmatter carries a name and a description of when the skill applies;
Claude reads that description, decides the current task matches, and pulls in
the rest as it goes.
The part that makes it useful: you do not invoke it
This is the difference that matters, and it is easy to miss. A prompt library only works if you remember you have it. A skill advertises what it is for, so when you say "write me something to send to these twelve plumbers", Claude recognises outreach and consults the method without being told to.
The second difference is volume. Nobody pastes six thousand words into a chat before asking a question. A skill can hold that comfortably, split across files the model opens only when relevant — so the deliverability material is there when you are debugging placement, and out of the way when you are writing a subject line.
Skill, prompt, custom GPT, MCP server
These get used interchangeably and they are genuinely different things. The honest comparison:
| Approach | Lives | Best at | Weak at |
|---|---|---|---|
| Prompt | In the message you type | One-off shaping, fast to try | Forgotten, re-typed, size-limited |
| Skill | A folder you own | A method applied consistently without being asked | Static — it knows nothing about today |
| Custom GPT | Inside one ChatGPT account | A packaged assistant you can share by link | Not portable; tied to a vendor |
| MCP server | A running service | Live data and taking actions | Needs hosting, auth and upkeep |
The last row is the one worth dwelling on, because it marks the real ceiling on skills. A skill is knowledge, not reach. It can teach Claude that a three-touch sequence over 7–14 business days is the shape that works, and it cannot tell Claude who replied yesterday, which domains are burnt, or who already asked not to be contacted. The moment you need any of that, you need something that can read live state — for us, an MCP server. Pairing the two is the honest setup: the skill supplies the method, the server supplies the facts.
What a skill is made of
Nothing exotic. Ours is seven files, and the structure is the whole format:
SKILL.md— frontmatter plus the method itself. If somebody only reads one file, it has to stand alone.- Supporting files, one per concern — writing the message, sequencing, deliverability, the funnel arithmetic, compliance.
That is it. You can open the whole thing in a text editor, disagree with a paragraph, change it, and keep using it. Compare that with a configured assistant inside somebody else's product, which you can use but not inspect.
When a skill is the wrong tool
Three cases, stated plainly because the alternative is a page that recommends skills for everything:
- The work needs current data. Pricing, availability, who replied, what is in the CRM. A skill written in October still says October things in March. Use a tool call or an MCP server.
- You need the model to act, not advise. Sending the mail, booking the slot, updating the record. Skills inform; they do not reach out of the conversation.
- You have not got a method yet. A skill is a written-down way of working. Writing one before you have one produces confident prose about nothing, which is worse than no skill, because the model will now follow it.
A real one you can read in full
Abstract explanations of file formats are not very useful, so here is a working skill with nothing held back. The Outreach Skill is about 6,500 words on cold outreach — message structure, CTA design, sequencing, deliverability setup, the funnel maths, and the compliance floor across CAN-SPAM, TCPA, CASL and GDPR. It is what we know from running campaigns, written for a machine to use.
It is free, it needs no email address, and you can modify and redistribute it. It also runs the arithmetic and tells most readers to do their outreach themselves, which is the only reason the cases where it says otherwise are worth reading.
- Download it — .zip for Claude, .md for everything else
- Or just read the whole file in your browser
- Installing it in Claude Code, Claude web, ChatGPT, Cursor or Gemini
- The prompts worth asking it once it is loaded
Questions
What is a Claude skill?
A Claude skill is a folder of markdown files that teaches Claude a method for a particular kind of work. It contains a SKILL.md describing when the skill applies, plus any number of supporting files Claude reads as it needs them. Claude loads a skill automatically when a task matches its description, so you do not have to remember to invoke it.
How is a skill different from a prompt?
A prompt is something you type once and it shapes one reply. A skill lives on disk, applies itself whenever the work matches, and can hold far more material than anyone would paste into a chat — ours is about 6,500 words across seven files. A prompt is an instruction; a skill is a reference the model consults.
Is a skill the same as a custom GPT?
They solve the same problem in different ecosystems. A custom GPT is a configured assistant hosted inside ChatGPT with its own knowledge files. A skill is a portable folder on your own disk that Claude reads. The practical difference is ownership: you can open a skill in a text editor, edit it, version it and pass it to someone else, and it is not tied to one vendor account.
Do I need Claude Code to use a skill?
No. Claude Code loads skills from a folder, and Claude web and desktop accept them as an upload under Settings → Capabilities → Skills, or as files attached to a Project. Because the content is plain markdown, the same material also works as a ChatGPT knowledge file, a Cursor rules file or a Gemini Gem.
Can I write my own skill?
Yes, and for anything specific to your own business you should. A skill is a folder with a SKILL.md in it; the frontmatter needs a name and a description saying when to use it, and the body is ordinary markdown. The hard part is not the format, it is having a method worth writing down.
What running this yourself actually takes
The method is the part people expect to be hard, and it is the part you now have for nothing. The work is everything around it — and almost all of it is a fixed cost you pay whether you send 200 emails a month or 2,000.
| What | To stand it up | Then, forever |
|---|---|---|
| Sending domains | Buy a separate domain, never the one carrying your real mail | Replace any that burn |
| Authentication | SPF, DKIM and DMARC, correct on every sending domain | Re-check after any DNS change |
| Inbox warm-up | Three to four weeks of ramping before a single real send | Re-warm anything that sat idle |
| Lists | Source the companies, find the people, verify the addresses | Rebuilt continuously — lists decay about 2% a month |
| Suppression | One list that holds across email, SMS and calling | Every opt-out, every channel, immediately |
| Placement monitoring | Seed accounts, so you find out before the replies stop | Weekly, and after every copy change |
| Calling or SMS | A2P 10DLC brand and campaign registration, numbers, a dialler | DNC scrubbing, calling windows, spam-flag rotation |
| Writing and sending | The method — which you now have in full | 20–40 hours a month at any real volume |
Below a few thousand sends a month that overhead is the job. And the failure mode is almost never a bad email — it is three good weeks followed by five silent ones, because sequences only produce anything when they actually run, and outbound is the task that slips first when a busy week arrives.
Run it yourself when you have 5–10 hours a week that are genuinely free every week, you are willing to spend the first month warming a domain before a single real send, and somebody will still be triaging replies in month four when it has stopped being novel. Plenty of people meet that. It is completely learnable and nothing is held back.
Hand it off when those hours are worth more spent elsewhere. For most owners of a working service business they are: 20–40 hours a month is $1,500–3,000 of your own time, and the same hours put into delivery or sales calls usually return more than they cost.
Or we run the whole thing
The Outreach Platform sources the leads in your niche, writes and sends the email, SMS, voice and social outreach, owns the deliverability, and forwards replies straight to your inbox. There is no dashboard you have to operate — and because it exposes an MCP server, the same assistant you just installed the skill into can run the account from the conversation you are already in.
| Plan | What runs | Month |
|---|---|---|
| Free | 10 leads and a written pitch for one — nothing sent | $0 |
| Sourcing, writing, sending, deliverability | $29 | |
| + Calling | Adds voice, registration and DNC handling | $59 |
| Full volume | Every channel at full send | $99 |
Set $29 against $1,500–3,000 of your own time and the comparison is the whole argument — run it with your own hourly rate and see where it lands. No contract. And the free tier means you can see the output on your own market before deciding anything.
No credit card for the free tier · Nothing sends until you say so