The first time I got an AI to do a job really well, I saved the prompt in a notes file. The next week I couldn't find it, typed a shorter version from memory, and got a worse result. The week after that I typed a different version again.
That's the problem a skill solves. A skill is that good prompt, written down properly once and kept where the AI can find it, so it does the job the same careful way every time.
This guide explains what a skill is, how it differs from the other things that get called "memory", and how to make your first one. You don't need to know any code.
A skill is a folder of instructions for one kind of job. At its heart is a single text file. The file says two things:
A skill can also hold extra files the job needs: a template, a checklist, an example of a finished piece, or a small script. Anthropic describes skills as folders of instructions, scripts and resources that Claude loads when it needs them, to do a specific task in a repeatable way.
In Claude, skills are stored in your account. When you ask for something, Claude looks at the descriptions of the skills you have switched on, picks the one that fits and follows it. You don't have to paste it in or remind it. Anthropic describes it this way: Claude reviews available skills, loads relevant ones, and applies their instructions.
These four get mixed up all the time. Here is the simple version.
| What it is | Think of it as | |
|---|---|---|
| Prompt | What you type in one chat | A spoken request |
| Skill | Saved steps for one kind of job, used whenever that job comes up | A written method or recipe |
| Project | A workspace with its own files and instructions, always loaded inside it | A filing cabinet drawer for one piece of work |
| Memory | What the AI keeps about you and your work between chats | Notes about where things stand |
The key difference between a skill and a project: a project's files are always there when you work inside it, but only inside it. A skill switches on only when the job needs it, and it works wherever you are. Anthropic's own help pages put it the same way: projects give background knowledge that is always loaded, skills give procedures that activate when needed.
Put simply: memory and projects tell the AI where things stand. Skills tell it how to do the job. You get the best results when you have both.
The other words you'll see, in plain English:
Don't make a skill for one-off questions, or for facts that change often. Facts belong in memory.
Anything you do more than a few times, where you know what good looks like. Anthropic's own examples are applying brand guidelines, following a team's way of working and automating personal tasks. Some everyday examples:
Anthropic lists these benefits, and they match what I've found:
A skill doesn't make an AI perfect. It reduces drift and mistakes. Anything that matters still gets checked by you.
In Claude, a skill's main file is a plain text file called SKILL.md, with the name and description at the top and the instructions underneath. Here is a made-up example for a small business newsletter. Yours will be different, but the shape is the same.
name: monthly-newsletter
description: Use when I ask for the monthly newsletter, the customer
email, or the "what's new" email.
Steps
1. Read the project's memory file first for this month's news.
2. Ask me for anything missing: offers, dates, new products.
3. Write a subject line under 50 characters and three options for it.
4. Keep the email under 250 words. One main message, one link.
5. Use UK spelling. No jargon. Sound like a person, not a brochure.
6. Check every date and price against what I gave you. If anything
isn't confirmed, mark it [CHECK] instead of guessing.
7. Give me the draft. Never send it.
8. When I approve it, add one line to the memory file: date sent,
subject line used.
Rules
- Never invent offers, prices or testimonials.
- If I say "no", ask what to change. Don't start again from scratch.
That's it. No code, just clear instructions. Notice the two things that make it work well: step 1 reads the memory file before starting, and step 8 updates it afterwards.
This is the part most guides leave out, and it's where the real gains are.
A skill on its own knows how to do the job but not what's happened so far. A memory file on its own knows where things stand but not the method. Join them up and the AI can pick up a job exactly where you left it.
The pattern I use:
Get this right and the skill does the job the same way every time, using the latest facts every time.
You don't have to write it from scratch. The easiest way is to get the AI to write it with you.
The exact prompt we use for this is in the SEO Intel Skills and Prompt Pack.
Signs a job deserves a skill:
No. Anthropic lists skills as available on the Free, Pro, Max, Ultra, Team and Enterprise plans. The one setting to check: skills need code execution switched on (Settings, then Capabilities on personal plans).
What a paid plan adds is mostly room. A long job that uses a skill uses up more of your allowance, so on a free plan you'll reach the limit sooner.
The idea works anywhere, but how you set it up differs.
In those, you may need to open the right project or Gem yourself rather than the AI choosing for you. Instructions written and tested for one assistant usually need a little adapting before they behave the same in another, so test on something small first.
A skill doesn't learn by itself. It only changes when someone edits it. But your memory files are learning all the time: every session adds what happened, what was decided and what went wrong. The trick is to feed those lessons back into your skills on a schedule, so the skill gets better as you do.
This is the habit that made the biggest difference to my results.
What to look at. Once a week if you use a skill a lot, once a month if you don't, ask the AI to read:
What to ask. Ask the AI to compare the skill with what your memory files have learned, and to suggest changes without editing anything until you say yes.
The exact review prompt, which also handles saving the old version and logging what changed, is in the SEO Intel Skills and Prompt Pack.
What to change.
Keep a record. Save the old version before you change it (v1, v2, v3) and add one line at the top of the skill: the date and what changed. If a change makes things worse, you can go back.
Keep it small. A review usually means three or four lines changed, not a rewrite. If a skill needs rewriting every month, it's probably trying to do too many jobs. Split it.
The loop looks like this: do the work, save what happened to memory, review memory against the skill, update the skill, repeat. Your memory is where the learning is stored. The review is how the skill catches up.
A skill is instructions your AI will follow, and some skills include small scripts. Treat them like any software:
Skills aren't clever. They're just your best way of doing a job, written down once. That's exactly why they work.
This guide comes from a real problem on one of my own stores. Here's how to skip it.
Run the free store audit. No account, no card. Built for Shopify; Wix stores get a quick check too.
Run the free auditSkills Pack (Shopify, £14.99): 8 Claude skills that check your store against Merchant Center rules and fix what they find, including suspension recovery.
Skills Pack £14.99I cleared my own suspension without paying anyone. The full story, every mistake and every fix, is free to read.
Read how I did it