Build your first Claude skill

A skill is a workflow you only explain once

A prompt is a one-off instruction. A skill is a folder Claude keeps, containing a file called SKILL.md: your instructions written in plain English, plus a short description telling Claude when to reach for them. When something you ask matches that description, Claude opens the file and follows your process. When it does not match, the file stays closed and costs you nothing, which is why you can have a dozen skills sitting there without slowing anything down. Anthropic calls that progressive disclosure, and the mechanics are documented in the Agent Skills overview.

The shift is subtle and it is the whole point. Without a skill, every good result you get is a one-time result, and next Tuesday you are re-explaining your format, your tone and your rules from scratch. With a skill, the fifth time is as good as the best time, because the best time got written down.

Five steps, about half an hour

1. ✍️ Pick the thing you redo weekly. Not the most impressive task, the most repeated one. In the reel it was a property listing post. For you it might be the Monday carousel, the Friday client recap, the caption you rewrite for every product drop, or turning your messy notes into a proper meeting summary. Write the name of it down before you open Claude.
2. 🧪 Do it with Claude once, manually, and get it genuinely right. This is the step people skip and it is the one that matters. Start a normal chat, walk Claude through the task, and correct it until the output is something you would actually publish. Every correction you make here becomes a rule in the skill, so the messy back and forth is not wasted time, it is the raw material.
3. 📝 Have Claude write the SKILL.md from what just happened. Paste the capture prompt below into that same chat. Claude turns your conversation, including the corrections, into a proper skill file with a name, a description, your step by step process, and your rules.
4. 📦 Save it as a folder, then zip it. Make a folder named after your skill in lowercase with hyphens, for example weekly-post. Inside it, save Claude's output as a file called SKILL.md. Add any supporting files you want it to use, like a template document or a voice guide. Then compress that folder into a zip. One trap worth knowing: the skill folder has to be the root of the zip, not sitting inside another folder.
5. ⬆️ Upload it and test the trigger. In the Claude app, open Customize, then Skills, then Add, then Upload a skill, and choose your zip. Now start a brand new chat and ask for the task in your own everyday words, without naming the skill. If Claude picks it up on its own, you built it right. If it does not, the description needs work, and the next section fixes exactly that.

The three prompts

First, the capture prompt. Run this at the end of the chat where you got the output you liked, so Claude is writing down what actually worked rather than guessing:

Turn everything we just did into a reusable skill file so you can repeat this exact process without me explaining it again.

Write it as a SKILL.md with YAML frontmatter containing a name in lowercase-with-hyphens and a description. The description has to say both what the skill does and when you should use it, and it should include the everyday phrases I would naturally type when I want this task done, because that is what you match against. In the body, write the process as numbered steps in the order we actually did them, then a rules section capturing every correction I gave you in this chat, then a section describing what a finished result looks like and what would make it wrong. Write it for a version of you with no memory of this conversation, so spell out anything I only said once.

Then the pressure test, before you trust it with anything real:

Read the SKILL.md you just wrote back to me as if you had never seen our conversation, and tell me where it would fail.

Specifically: which steps are vague enough that you would have to guess, which rules contradict each other, and what context you would be missing that I have in my head but never typed. Then tell me honestly whether the description would actually trigger this skill if I typed [paste the sentence you would really write], and if not, rewrite the description so it would. Finish by giving me the corrected full file, ready to save.

And the upgrade prompt, once the skill has run a few times and you know where it drifts:

Here is what the skill produced this week, and here is what I changed by hand before I used it: [paste both].

Work out what rule was missing that would have prevented each of my edits, and write those rules in the same voice as the rest of the file. Do not just add more instructions, that is how a skill file gets bloated and starts contradicting itself. Where a new rule replaces an old one, say so and remove the old line. Give me the updated SKILL.md in full, and a one-line note on what changed so I can see the file getting sharper over time.

The description does all the work

Almost every "my skill never triggers" problem is a description problem. Claude only sees each skill's name and description until something matches, so that one sentence is the entire hiring decision. Anthropic's skill authoring best practices are worth a read, and the short version is this: say what it does and when to use it, in the words you would actually type.

The weekly content skill, written out

This is the reel's workflow, generalised so you can lift it. Swap the specifics for whatever you post. Save it as SKILL.md inside a folder called weekly-post:

---
name: weekly-post
description: Builds my weekly [post type] end to end, from picking the subject through to the finished caption in my voice. Use when I ask for this week's post, say I need something to post, mention my weekly [post type], or ask to prep content for tomorrow.
---

# Weekly post

## What this does
Takes my recurring [post type] from a blank page to a draft I can review in one pass.

## Steps
1. Ask me for the subject, or if I have not given one, propose three options from what we have discussed recently and wait for me to pick.
2. Research the subject properly before writing. Pull the specific details that make the post concrete: real numbers, names, dates. If a detail cannot be verified, flag it rather than writing around it.
3. If the post needs permission from someone else (a client, a partner, a source), draft the request email and show it to me. Never send anything, only draft.
4. List the images or assets the post needs and where each one comes from. If I have to supply one, tell me exactly which.
5. Rebuild my template around this week's subject using the structure and rules below.
6. Hand me the finished draft plus a one-line note on anything you were unsure about.

## My template
[Paste your actual post structure here, section by section, exactly as you use it.]

## Rules
- Sentence case, no em dashes, no corporate filler.
- Lead with the specific thing, never a general statement.
- Every claim traceable to something real.
- Draft, never send or publish. I approve everything.

## Done looks like
A draft I would post after a light read, not a rewrite. If I am rewriting more than two lines, a rule is missing and you should ask me what it is.

Getting it to run while you sleep

Here is the honest version, because the reel makes it look like the skill wakes itself up. It does not. A skill is the recipe, not the timer. What it gives you is a task that runs start to finish from one short message instead of twenty minutes of directing, and that is what makes automating it possible at all. There are three levels, and most people should stop at the first:

1. ☕ Level one: one line, first thing. You open Claude with your coffee and type "do this week's post". The skill does the rest while you read your emails. Free, no setup, and genuinely the version I use most.
2. ⏰ Level two: put it on a schedule in Claude Code. Claude Code reads skills straight from a folder on your computer, personal ones in ~/.claude/skills/ or per-project ones in .claude/skills/, with no zip and no upload. From there you can have it run on a timer so the draft is waiting when you wake up. This is the real "before you wake up" version, and it does mean getting comfortable in a terminal.
3. 🔒 Level three: keep yourself as the approval step. Whatever level you land on, have the skill draft and stop. Emails drafted, not sent. Posts written, not published. The agent in the reel drafted a permission email and left it sitting there for a human to read, which is exactly right, and it is the habit that keeps automation from becoming a story you tell about a mistake.

Two safety notes

Proof

Anthropic, how to create custom skills (plan availability and the upload path)
Agent Skills overview (SKILL.md structure, naming rules, security guidance)
Skill authoring best practices
Teach Claude your way of working using skills
Using skills in Claude Code (the folder paths)