Build Your First Claude Skill
Most PMs keep their best prompts in a notes file and rely on memory to use them. Strategic PMs write the situation down once and create a skill so they can easily do it again and again.
Here are two ways to write the line that tells Claude when a skill applies. A weak one:
For searching memory
The second comes from the recall skill in Module 6 of my project context system course:
When a problem or situation feels like something handled in a past sessionBefore /plan on any recurring output type
The first gives Claude nothing to match a moment against. The second names two situations Claude can recognize while you work. The rest of this issue shows you how to write the second kind and check that it fires.
A Skill Is a Reflex
A skill is a folder with a SKILL.md file in it. Claude reaches for the skill on its own when your conversation matches what the file describes. You can also run it by hand by typing /skill-name.
A saved prompt waits until you remember it. A skill watches the conversation and speaks up when its moment arrives.
The Prompt You Have to Remember
Picture a PM with a good stakeholder-update prompt saved in a notes file. It asks for the outcome and the untested assumptions before any drafting starts. It's Thursday afternoon, the update is due in an hour, and she types "draft the update for the vendor migration" straight into Claude. The notes file stays closed. (This scene is made up.)
Most PMs keep their best prompts in a notes file and rely on memory to use them. Strategic PMs write the situation down once and let a skill speak up when it arrives.
In Mirror, Telescope, Hammer terms, the draft request is Hammer Mode. The skill you're about to build puts Mirror Mode in front of it.
Build It in Five Steps
1. Pick one situation you repeat. Choose a moment where you always wish you had paused first. This walkthrough uses the moment you ask Claude to draft a stakeholder update.
2. Create the folder. In your project, make .claude/skills/outcome-check/SKILL.md. Put the folder in ~/.claude/skills/ instead to use it in every project. Give the folder and the name field the same name. That name is the command you type.
3. Write the frontmatter. The block at the top of the file tells Claude what the skill is and when to use it:
---
name: outcome-check
description: Before drafting a stakeholder update, status report or steering-group message, confirm the outcome it supports and surface untested assumptions.
when_to_use: |
- "Draft a stakeholder update for..."
- "Write the status report for..."
- Before any message to a sponsor
---
Claude matches your conversation against the description and when_to_use lines together. Write both in the words you would type, as situations.
This walkthrough is for Claude Code. Skills uploaded to claude.ai accept a narrower set of fields (name, description, license, compatibility, metadata, allowed-tools), so skip when_to_use there and put the situations in the description.
4. Write the procedure. Below the frontmatter, say what Claude does when the skill fires:
# Outcome Check
Before drafting, ask me:
1. What outcome does this update support?
2. What would a skeptical sponsor challenge first?
3. What am I assuming that I haven't tested?
Wait for my answers. Then draft the update around them.
The last line carries the weight. A skill that asks first and writes second keeps you in charge of the thinking.
5. Test both doors. Type /outcome-check and confirm it asks your three questions. Then open a fresh conversation and type "draft a stakeholder update for my vendor migration" without naming the skill. If Claude asks the questions on its own, the trigger works. If it jumps straight to drafting, your trigger lines are probably too abstract. Add the exact sentence you typed.
Where First Skills Go Wrong
Module 6 of my course lists three common mistakes. Trigger lines written as a description of the skill, like "for searching memory," give Claude little to match. A skill that writes files or drafts text without asking first takes the judgment away from you. A skill with trigger lines so broad that it fires after every message teaches you to dismiss it.
Keep the situations specific and the first move a question.
What to Do This Week
Open Claude Code in one of your project folders and paste this:
Help me write my first Claude skill. The situation I repeat: [describe the moment, for example "I'm about to ask for a stakeholder update"]. What I wish I did first: [the pause you want, for example "confirm the outcome and list the assumptions I haven't tested"]. Create it at .claude/skills/[skill-name]/SKILL.md. Write the description and when_to_use as the exact phrases I would type. The procedure should ask me questions first and write nothing until I answer.
Then run the two-door test from step 5. Notice whether the skill speaks up in a conversation where you never mentioned it.
Talk soon,
Paul
P.S. If you build one, hit reply and tell me which situation you wrote it for. I'm collecting first-skill situations from PMs, and the best ones become future issues. Contact Me