Turn a task you repeat every week into a Claude skill

Take one routine you already do the same way every time, write it down as steps, package it as a Claude skill, and test it on real cases until it works without you re-explaining. Built from Anthropic's official help pages and skill-writing guide, with one worked example.

Start here

How to use this canvas

What a skill is

What you'll end with

Pick the routine

Which routine to start with

Good first skills have all three:
- You do it at least every week.
- The steps are mostly the same each time.
- The result has a fixed shape: an email, a table, a checklist, a summary.
Examples: follow-up after a client call, weekly report, turning interview notes into findings, checking a draft against your style rules.
Keep it to one routine. The official guide says to keep each skill focused on one workflow.

Write the routine down as steps, before any AI

Open a blank note and answer:
1. What do I start with? (for example: rough notes from a call)
2. What are the steps, in order?
3. What does the finished result look like? Paste one real example you were happy with.
4. What mistakes do I always have to fix? Those become rules.
This is the part that takes experience. Claude can tidy the wording later, but only you know which steps matter.
Keep these steps and your three test cases on the same put·here canvas, so they're at hand whenever you change the skill.

Turn it into a skill

Turn on code execution first

The skill file, part by part

At the top, between two lines of three dashes:
- name: lowercase letters, numbers and hyphens only, up to 64 characters, the same as the folder's name. Anthropic's skill-writing guide adds that it can't contain the words "claude" or "anthropic".
- description: what it does AND when to use it, written in the third person. This is what Claude reads to decide whether to use the skill, so put in the words you'd actually type. Keep it under 200 characters (see "Where the official pages disagree").
Below that, plain text: steps, rules, and the output format. Keep it short; Anthropic's guide says to assume Claude is already smart and only add what it doesn't know.
- Sources: How to create custom skills, Claude Help Center (official); Agent Skills specification (official spec site) https://agentskills.io/specification

Example: the skill file

A follow-up after a client call. Copy it, then change the steps to yours.

---
name: client-call-follow-up
description: Turns rough client-call notes into a follow-up email, an action list with owners and a client-file update. Use after any client call or when the user pastes call notes.
---

# Client call follow-up

## Steps
1. Read the notes. List anything you can't tell for sure (a date, who owns a task, a number) under "Check before sending". Don't guess.
2. Write the follow-up email: subject "Recap:" plus the topic of the call, under 120 words, thanks in one line, then what we agreed, then next steps with dates.
3. Write the actions as a table: Action | Owner | Due. If the notes don't say, write "?".
4. Write a 3-line update for the client file: what changed, what's decided, what's open.

## Rules
- Use only what's in the notes. You MUST NOT invent dates, prices or promises.
- If the notes mention money or scope changes, put them first in the email.

## Output format
### Check before sending
### Email
### Actions
### For the client file

Tried in 2026-11: it passed our format check after we cut the description from 228 to 168 characters.

Let Claude draft it from a real conversation

Anthropic's guide suggests this: do the routine once with Claude in a normal chat, correcting it as you go, then ask:

We just did my client-call follow-up together. Turn what worked into a skill: a SKILL.md with a name, a description that says what it does and when to use it, numbered steps, the rules I had to correct you on, and the output format. Keep it under 60 lines.

Then read every line. Delete anything Claude would do anyway.

Package and upload

Make the zip

Upload and switch it on

Test it until it holds

Three real cases before you trust it

Anthropic's guide says to build at least three test cases before relying on a skill. Pick three past examples:
- a normal one,
- one with something missing (no date, no owner),
- one with a tricky part (money, a scope change, an angry client).
Paste each into a new chat the way you normally would, and check the result against what you'd have sent.
Tried in 2026-11 on three made-up call notes: the missing owner came out as "?" and was flagged; the fee change went to the top of the email; nothing was invented.

If Claude doesn't use the skill

If the result drifts

Honest bit / Try it this week / Sources

Where the official pages disagree, and what to watch

Make one skill from your most repeated task

Write down one routine (see "Write the routine down as steps, before any AI"), turn it into a skill, and run your three test cases. Next time you catch yourself explaining the same thing twice, add it.

Sources

put·here
made by puthere