Back

How to create good AI agent skills (with Matt Pocock’s skills repo)

By

How to create good AI agent skills (with Matt Pocock’s skills repo)

What AI agent skills are, how to install mattpocock/skills, and a concrete checklist for writing SKILL.md playbooks that keep coding agents on track.

1 What AI agent skills are

Search “how to develop AI skills as a software engineer” and you mostly get career roadmaps: Python, RAG, agents, evaluation. Fine. Still useful. There is a second meaning I care about more day to day: agent skills. Those are reusable Markdown playbooks that teach a coding agent a disciplined workflow, so every chat does not start from zero.

  • A SKILL.md playbookSays when to wake up and how to behave for one job
  • Survives the tabA one-off prompt dies when the chat closes; a skill stays in the repo
  • Small enough to mixPortable across harnesses that speak the Agent Skills pattern
  • Boringly consistentSame process when the same failure shows up again
Disposable chat prompt dissolving versus a reusable SKILL.md playbook card

Career skills vs agent skills

Classical “AI engineer skills” are about composing models into products. Agent skills are smaller and more selfish: you compose your process into the agent so it ships more like a careful senior and less like a fast junior with amnesia.

Links mattpocock/skills Agent Skills standard Write skills that actually work

2 Why mattpocock/skills is a strong model

mattpocock/skills calls them “Skills for Real Engineers.” Small files. Editable. Not another mega-framework that owns your whole process. The README does not flatter the big process kits. When something breaks, you should be fighting the bug, not the framework.

The repo groups around four failure modes almost everyone hits with coding agents:

MisalignmentThe agent built the wrong thing. Fix with grilling (/grill-me, /grill-with-docs).
Fuzzy languageNo shared vocabulary. Put project language in CONTEXT.md and ADRs.
Code that does not workWeak feedback. Reach for /tdd, /diagnosing-bugs, types, browser checks.
Ball of mudThe agent accelerates entropy. Use design-aware flows like /to-spec and /improve-codebase-architecture.
Skills come in two flavors. User-invoked: you type them; they orchestrate. Model-invoked: the agent can pull them in when the task fits. A user-invoked skill may call model-invoked ones. Do not invert that into a pile of competing slash commands.

3 Install and use mattpocock/skills

Option A: skills.sh (editable copy)

Use this when you want to hack the skills and make them yours.

  1. From your project root run npx skills@latest add mattpocock/skills
  2. In the installer, pick the skills and coding agents you care about. Include /setup-matt-pocock-skills
  3. In agent chat, run /setup-matt-pocock-skills (issue tracker, triage labels, docs location)
  4. First loop: /grill-with-docs, then /to-spec and /to-tickets, then /implement

Option B: Claude Code plugin

Use this when you want Pocock’s set current and mostly read-only.

  • /plugin marketplace add mattpocock/skills
  • /plugin install mattpocock-skills@mattpocock

Or from the shell with claude plugin marketplace add / claude plugin install. Run setup once per repo, same as Option A.

First loop Grill→Spec→Tickets→Implement

Links Repo README skills.sh

4 How to create good AI skills

The repo ships /writing-great-skills. Treat that as the meta-skill. Pocock’s core idea is blunt: a skill wrangles predictability out of a stochastic system. Same process every run. Not identical prose every time.

Checklist before you commit a skill
One job Right mode Trigger verbs Checkable steps
Progressive Composable Repo truth Prune hard
  • One job“Grill a plan” or “red-green-refactor.” Not “do engineering.”
  • Right invocation modeUser-invoked costs nothing until you type it. Model-invoked keeps a description in context so the agent can discover it.
  • Trigger-rich descriptionSay what it is up front, list distinct “use when…” branches, cut synonym spam.
  • Checkable stepsEach step ends on something the agent can verify, or it will declare victory early.
  • Progressive disclosureKeep the ladder short in SKILL.md. Park long reference in linked files.
  • Repo truthPoint at real files (CONTEXT.md, ADRs, issue trackers). Do not invent process in chat memory.

Start by copying pain you already feel

Copy one Pocock skill that matches a real friction (grilling or TDD for most people). Run it on a real change. Rewrite the copy in your project’s language. That loop teaches more than ten generic prompt-engineering posts.

5 A loop worth forcing this week

  1. GrillAlign on the real goal
  2. SpecWrite what “done” means
  3. TicketsBreak work into checks
  4. ImplementTDD + review
  • Install the set and run setup once.
  • Force one full loop on a non-trivial change.
  • If the agent still drifts, improve shared language before you buy another course.
  • On living codebases, /improve-codebase-architecture every few days keeps the mud from winning.

6 Where this fits a broader AI roadmap

You still need to know how to use LLMs: APIs, prompting, RAG, agents, evaluation. Agent skills do not replace that stack. They sit on top. Once you can call a model, skills decide whether it behaves like a careful engineer or a slot machine.

For full-stack folks, the path that has paid off for me is simple: keep shipping product code, add agent skills for alignment and feedback, then deepen into RAG or eval when the product actually needs them. Build first. Do not train a model for sport.

Back