A prompt you paste every week is a tool you have not built yet. Today you build one.
Memory tells the agent about your project. A skill teaches it how to do one specific job. In Claude Code, a skill is a folder with a SKILL.md file inside: a name, a description of when to use it, and the instructions.
Claude Code sees the description of every skill you have installed and loads the full instructions only when a task matches. So skills cost almost nothing until they are needed, and you can also run one directly by name.
Pick the job
The right first skill is a job you already do repeatedly and explain the same way every time. Good first candidates:
- ·Review a change before it ships, against your own checklist
- ·Write release notes from the latest commits
- ·Check a page against your copy rules: banned words, tone, length
- ·Turn a bug report into a reproduction and a fix plan
A bad first skill is "act like a senior engineer". A skill does one job, with steps and a finish line.
The shape of a skill
Skills live in .claude/skills inside your project, shared with anyone who works on it, or in ~/.claude/skills, yours across every project.
---
name: ship-review
description: Review the current change before it ships. Use when asked to review, check, or get a change ready to merge.
---
# Ship review
1. Run git diff against main and read every changed file in full.
2. Check each item below. Quote the line for every failure.
- No secrets, keys or tokens in the diff
- No debug logging left in
- New user-facing text follows the rules in CLAUDE.md
- Every new API route checks who the caller is
3. Run the type check and the tests. Paste the real output.
4. Report what passed, what failed with file and line, and what you could not check.
Do not fix anything unless asked. This skill reports.What makes a skill work
- ·The description is the trigger. Write it as when to use the skill, in the words you would actually type. A vague description means it never fires.
- ·Steps, not vibes. Numbered steps it can follow and you can audit.
- ·A finish line. Say exactly what the output is, so a finished run is easy to tell from an abandoned one.
- ·One job. If a skill grows a second job, split it in two.
Today's task
- ·Pick one job you have explained to the agent more than twice.
- ·Create .claude/skills/your-job-name/SKILL.md with a description, numbered steps and a defined output.
- ·Start a fresh session and ask for the job in plain words, without naming the skill. Check that it picked the skill up.
- ·Run it on real work, then fix whichever step was unclear.
Output: one working skill in your project, used once on real work.
Free, with your email
Open days 2 to 3.
No account, no card. The day opens right here, and the next steps of the path come to your inbox.
Go deeper in the Inner Circle
- The Operator Mindset
