Agent Skills For Real Engineers
Matt Pocock skills: small composable agent workflows, not vibe coding. Install, user vs model skills, core flow.
Vibe coding ships demos. Real engineering ships systems that survive change. mattpocock/skills ↗ — 275k stars, 23k forks, MIT — is a set of small, composable agent skills for real work: grilling, domain modeling, TDD, bug diagnosis, architecture, triage.
Not GSD, BMAD, or Spec-Kit. Those own the process and take your control. These skills are sharp single-purpose tools. You compose them. They work with any model.
Why These Skills Exist: 4 Failure Modes#
| # | Failure | Fix skill |
|---|---|---|
| 1 | Agent did not build what you wanted | /grill-me, /grill-with-docs |
| 2 | Agent too verbose, jargon soup | shared language in GLOSSARY.md + ADRs |
| 3 | Code does not work | /tdd, /diagnosing-bugs |
| 4 | Ball of mud, entropy at agent speed | /to-spec, /improve-codebase-architecture |
ELI5: handing an agent a task with no questions is like sending a builder to “build a house” with no blueprint. Grilling is the blueprint interview. Shared language is agreeing “wall” means brick, not curtain. TDD is checking each brick before the next. Architecture review is fixing the foundation before adding floors.
Install in 30 Seconds#
Two ways in, two philosophies. Pick one — both leaves every skill twice.
# Claude Code: managed read-only bundle, auto-updates
claude plugins install mattpocock-skills
# or inside session:
/plugin install mattpocock-skills
# Codex and other agents: editable copy you own
npx skills@latest add mattpocock/skills
# select skills, include setup-matt-pocock-skillsbashThen run once per repo:
/setup-matt-pocock-skills
# asks: issue tracker (GitHub / Linear / local files)
# asks: triage labels for /triage
# asks: docs location for generated docstext[!WARNING] Installing both plugin and skills.sh copy duplicates every skill. Subscribe (plugin) or fork (skills.sh), not both.
[!TIP] Tinkerers: use
npx skills@latest addeven on Claude Code. Files land in your repo, nothing updates behind your back. Pull later withnpx skills update.
User-Invoked vs Model-Invoked#
The one axis that matters:
- User-invoked (
/grill-me,/implement,/triage…) — only you trigger by typing. Job: orchestrate. Never calls another user-invoked skill. - Model-invoked (
/tdd,/code-review,/diagnosing-bugs…) — you or the agent triggers when the task fits. Holds reusable discipline.
A user skill may invoke model skills, never another user skill.
Engineering Skills Map#
User-invoked:
/ask-matt— router: which skill fits your situation/grill-with-docs— interview + sharpen terms, updateGLOSSARY.md+ ADRs inline/triage— move issues through a state machine of triage roles/improve-codebase-architecture— scan for deepening opportunities, HTML report, grill your pick/to-spec— synthesize conversation into spec on tracker, no interview/to-tickets— break plan into tracer-bullet tickets with blocking edges/implement— build from spec/tickets, TDD at seams, close with/code-review/implement-spec— whole spec on one branch, subagents across ready frontier/wayfinder— huge work as decision-ticket map, resolve one at a time/retro— environment improvements after session, most severe first
Model-invoked:
/tdd,/diagnosing-bugs,/research,/domain-modeling,/codebase-design,/code-review,/pr,/prototype,/wizard
Productivity Skills Map#
User-invoked: /grill-me, /handoff, /teach, /to-questionnaire, /wait-what
Model-invoked: /grilling (interview primitive behind grill-me, grill-with-docs, triage, wayfinder), /writing-for-agents
Practical Example: Feature Slice#
npx skills@latest add mattpocock/skills
# select: setup-matt-pocock-skills, grill-with-docs, to-spec, to-tickets, implementbash# 1. per repo, once
/setup-matt-pocock-skills
# 2. align before code
/grill-with-docs
# agent interviews you, sharpens terms into GLOSSARY.md, records ADRs
# 3. spec then tickets
/to-spec
/to-tickets
# 4. build with discipline
/implement
# drives /tdd at seams, closes with /code-review before committextShared-language payoff, real example from author’s repo: BEFORE “a lesson inside a section given a spot in the filesystem” → AFTER “materialization cascade”. One term replaces twenty words, every session, plus consistent names for variables, functions, files.
Cross-Agent Notes#
Built multi-agent: Claude Code, Codex, Cursor, Copilot, Windsurf, Gemini, Cline via skills.sh. Two nuances: OpenCode is not on the installer list — install into .claude/skills or .agents/skills which it reads, or copy folders. And user-invoked flags are Claude Code convention — agents ignoring the flag expose them as model-invocable instead.
When to Use, When to Skip#
| Use it if | Skip it if |
|---|---|
| You want discipline-preserving workflows | You want autonomous framework owning process |
Tool loads SKILL.md | Tool cannot load skills |
| Team values TDD, ADRs, issue hygiene | You want one-shot generation |
| You adapt small skills to repo | You need hosted opinionated PM product |
Conclusion#
Small, composable, hackable. Start with /setup-matt-pocock-skills, then /grill-with-docs for every feature. Part 2 covers the grill → TDD → debug loop in depth.