AI Guides › Step-by-step guides
By Nigel Guy · 6 min read
Most people who use Claude daily keep a note somewhere with "the prompt that works", paste it in, then spend two messages re-explaining the parts the note forgot. It feels efficient because each paste is quick. In practice you are the one carrying the instructions from chat to chat, and the quality depends on your memory that day.
The rule: if you have explained the same job to Claude three times, stop explaining it and write it down once as a skill, with a description that says when to use it.
What a skill is. Anthropic describes skills as folders of instructions, scripts and resources that Claude loads dynamically for specialised tasks. Claude sees only each skill's name and description up front, and reads the full instructions when your request matches. That is why you can keep many skills without clogging every conversation.
What you need.
Where skills live today. In Claude on the web and desktop, click Customize, then Skills. Select "+" and "Browse skills" to see the directory. In Claude Code, skills are folders such as ~/.claude/skills/<skill-name>/SKILL.md (personal, all projects) or .claude/skills/<skill-name>/SKILL.md (one repository). Menu names move, so if yours differ, search the help centre for "custom skills".
Pick a job only if it passes all four:
| Check | Pass looks like |
|---|---|
| Repeats | You have done it three or more times this month |
| Stable | The good version looks broadly the same each time |
| Teachable | You can say what "good" is in a page or less |
| Checkable | You can tell in a minute whether the output is right |
Candidates that usually pass (my suggestions, not an official list):
Run the task in a normal chat with a real example. Correct Claude as you go: "shorter", "owners in a table", "don't invent dates". Those corrections are the skill. Anthropic's own authoring guide recommends this order: complete a task without a skill first, notice what you keep supplying, then capture it.
You do not need to hand-write the file. Anthropic's guide says Claude understands the skill format natively and will generate a structured SKILL.md if asked. Paste this at the end of the same chat. Fill in the bracketed parts.
You are an expert at writing Claude skills (a SKILL.md file with YAML
frontmatter and markdown instructions).
Context: in this conversation we just completed [JOB_NAME]. My corrections
and preferences above are the real requirements.
Goal: turn what we did into a reusable skill that a fresh Claude, with no
memory of this chat, could follow and get a result I would accept.
Steps:
1. List the rules you inferred from my corrections and ask me to confirm
or fix them. If anything is unclear, ask before writing.
2. Write the skill with:
- name: lowercase letters, numbers and hyphens only, under 64
characters, no word "claude" or "anthropic"
- description: third person, says what the skill does AND when to
use it, with the phrases I would naturally type, under 200
characters
- body: ordered steps, the exact output format, one short good example,
and a "never do this" list
3. Keep the body short. Do not explain things Claude already knows.
4. Put reference material I am likely to change, such as [REFERENCE_MATERIAL],
in a separate file and link to it from SKILL.md.
Constraints: do not invent facts about my business, names, prices or
dates. Use placeholders where you lack information.
Before you answer, check that the description would trigger on
[THREE_REQUESTS_I_WOULD_REALLY_TYPE] and would not trigger on unrelated
tasks. Then give me the finished files.
Fill in the job name, any reference material (a style guide, a template), and three real requests you would type.
In Claude (web/desktop). Make a folder named after the skill containing SKILL.md. Zip the folder so the skill folder itself is the root of the zip, not loose files. Go to Customize, Skills, upload the zip, then enable it.
In Claude Code. Save it as ~/.claude/skills/<skill-name>/SKILL.md. Type /skills to confirm it is listed, and /skill-name to run it directly.
One conflict to know about: the support article says the description is capped at 200 characters, while the developer documentation says 1,024. I could not reconcile these, so stay under 200 and you are safe on both.
Use three requests phrased differently, including one worded the way you would type it when tired. In Claude, review the thinking to see whether the skill was invoked. If it did not fire, the fault is nearly always the description. Add the words you actually used. If it fires on everything, narrow the description. In Claude Code you can also set disable-model-invocation: true so it only runs when you type its name.
Each time the output is wrong, go back to the chat that wrote the skill and say what happened: "it forgot to flag unverified claims; make that rule more prominent". Anthropic's guide recommends iterating on real use rather than anticipated cases, and keeping the SKILL.md body under about 500 lines.