04 · Best practice
Author skills, not prompts
Skills auto-load by trigger phrase. Stop re-typing the same instructions in chat.
Skills are the unit of compounding leverage in Claude Code. Every time you re-prompt the same convention (commit messages, PR descriptions, voice for stakeholder updates, how to write a release note), you're paying tax on a thing you could have written down once.
Aman Khan puts it well: "Skills are a peek into your mind as a builder. A well-written skill encodes how you think about a problem, what conventions you follow, what mistakes you've learned to avoid."
Anthropic's docs are explicit: at startup only the name and description of each skill are loaded. The full SKILL.md only gets read when relevant. That means you can have 50 skills installed without bloating context.
How to do it
The next time you re-prompt yourself, stop. Open Claude and say: "Create a skill for [the task]. It should include [your requirements]. I want it to [your style]."
Save it to ~/.claude/skills/<skill-name>/SKILL.md (global) or <repo>/skills/<skill-name>/SKILL.md (project-local).
Write a specific description with trigger phrases — Anthropic recommends gerund form: writing-release-notes, prioritizing-backlog. Vague descriptions don't get auto-invoked.
Test: ask Claude to do the task without the slash command. Did the skill load? If not, the description is too vague.
The format that works
Every skill in this repo follows the same seven sections so PMs see one shape:
- Goal — what this skill produces, and why a normal prompt can't get there.
- Process — numbered. The order matters.
- Clarifying questions — letter/number options so the user picks instead of writes.
- Output structure — the document the skill produces.
- Target audience — calibrates voice and depth.
- Tiny worked example — one paragraph of what good looks like.
- Final instructions — the three things to do first.
---
name: writing-release-notes
description: |
Write user-facing release notes in our voice from a list of merged PRs.
Use when the user mentions release notes, changelog entries, or "what
shipped this week".
---
Where to start (curated)
- anthropics/knowledge-work-plugins — official from Anthropic. Skills for Feature Specs & PRDs, Roadmap Planning, Stakeholder Updates, User Research Synthesis. If you only install one thing, install this.
- phuryn/pm-skills — 100+ skills across discovery → strategy → execution.
- deanpeters/Product-Manager-Skills — 49 skills + 6 command workflows.
- aakashg/pm-claude-code-setup — Aakash Gupta's opinionated minimal starter.
Full curated index on the Skills page.
Anti-patterns
- A skill called
helper. Anthropic's docs literally call this out. - Pasting an English description with no trigger phrases — Claude won't auto-invoke.
- Skills over 500 lines (split into reference files via progressive disclosure).
- Reserved words in the name (
anthropic,claude). Validation will reject them.