Skip to main content
Forty minutes, one deliverable: a skill built from your real work, installed in your real Claude, triggered by a real request, and tightened once based on what happened. You leave with the skill and, more importantly, with the loop for making the next one.

​
Step 0: Pick the victim (5 min)

The best first skill is something you have explained to Claude, or to a new colleague, at least three times. Pick one from these archetypes, or bring your own:
A format

Your memo structure, weekly status shape, client email conventions, deck outline. The skill carries the template and the rules.

A procedure

How your team abstracts a lease, reviews a contract, QAs a report. The skill carries the checklist and the order.

A behavior

A standing way of answering: terse replies, always cite sources, British spelling, numbers to one decimal. Small, fun, surprisingly sticky.

Write one sentence: “Claude should ______ whenever I ______.” If you cannot fill both blanks, pick a different victim.

​
Step 1: Mine a real thread (5 min)

Find a past conversation where you coached Claude through this task. Then make Claude do the archaeology:
No usable past thread? Do the task once now, coaching as you go, then run the same prompt. The corrections you just made are the skill’s core content.

​
Step 2: Draft the folder (10 min)

Claude knows the format natively. Hand it the mined rules:
Then review like an editor, against this bar:
  1. Description names both the what and the when.
  2. Nothing in the body explains general knowledge.
  3. Any template or example lives under its own heading so it can move to a separate file later.
  4. If a step must never vary, it is stated as an exact command or exact text, not a suggestion.

​
Step 3: Install it (5 min)

Claude apps: ask Claude to package the folder as a zip (folder name matching the skill name, SKILL.md inside). Then go to Customize, then Skills, click the add button, choose upload, and select your zip. Toggle it on. Claude Code: save the folder as ~/.claude/skills/your-skill-name/ and it is live in your next session. Use .claude/skills/ inside a repo to share it with everyone who clones.

​
Step 4: The three-request exam (10 min)

Open a fresh conversation. No warm-up, no mentioning the skill. Run:
  1. A request that should trigger it, phrased the way you naturally would.
  2. A second should-trigger, phrased differently, maybe the way a teammate would say it.
  3. A request in the same neighborhood that should NOT trigger it.
Score honestly. Fired on 1 and 2, stayed quiet on 3, output followed your rules: shipped. Anything else, continue to step 5 with a diagnosis in hand.

​
Step 5: Tighten one screw (5 min)

Match the failure to the fix: Re-run the exam. One tightening pass is usually the difference between a demo and a tool.

​
Ship report

Done means you can check every box:
  • The skill fired on a natural request in a fresh conversation.
  • It stayed quiet on the near-miss.
  • The output followed your rules without coaching.
  • A named teammate is going to receive this skill this week, by share, provisioning, or repo.

​
Stretch goals

  • Add a resource: drop your real template file into the folder and reference it from SKILL.md, so the instructions stay lean and the template loads only when used.
  • Provision it: on Team and Enterprise plans, have an owner push it organization-wide and become the person who taught Claude how your company writes.
  • Build a behavior skill such as compressed replies or always-cite mode. Include a persistence line, an escape hatch, and an off switch. See how many turns it survives.