Skills For Real Engineers
Small editable skills for coding agents: user-invoked ones orchestrate, model-invoked ones carry the discipline.
The problem it solves
Matt Pocock opensourced the skills he actually uses for day-to-day engineering work from his .agents directory, for coding agents like Claude Code and Codex. A typical round: type /grill-me and the agent grills the requirements through before acting; /implement drives /tdd at the agreed seams and passes through /code-review before committing. The README calls out that approaches like GSD, BMAD and Spec-Kit "take over" the process — you can't change what you want to change, and bugs in the process itself are hard to diagnose; this set goes the other way: small, freely modifiable, composable, model-agnostic, having the agent do one thing at a time with human feedback points between steps.
It answers four common failure modes:
- The agent didn't understand the requirement: the grilling step forces the agent to ask the questions through before acting, corresponding to
/grill-meand/grill-with-docs - The agent waffles too much:
/grill-with-docsalso captures the project glossaryCONTEXT.mdand ADRs, establishing a shared language - The code doesn't run: it fills in the test, type and browser feedback loops, and
/tddenforces a red-green-refactor cycle - The codebase turns into a mud ball:
/improve-codebase-architectureperiodically scans for "deep module" candidates
How to use it
The two install paths work on different philosophies — pick one; installing both makes every skill appear twice:
# Path one: Claude Code official marketplace plugin, read-only whole package, auto-updates with the author's releases
claude plugins install mattpocock-skills
# Path two: the skills.sh installer, copies the skill files into your repo, freely modifiable
npx skills@latest add mattpocock/skillsPath one is a "subscription", for when you just want to use it; path two is a "fork", and other agents like Codex take that route — you can pick skills when installing, and updates need a manual npx skills update. After installing, run /setup-matt-pocock-skills once per repo; it asks which issue tracker you use (GitHub, Linear or local files), what labels triage applies, and where docs are stored. If you don't know where to start, ask /ask-matt, which acts as a router for the set.
Inputs and outputs
Every skill falls into one of two layers, by "who can trigger it":
- user-invoked: runs only when you type the command, handling orchestration, e.g.
/grill-me,/triage,/to-spec,/wayfinder,/implement; it can call model-invoked skills but not another user-invoked one - model-invoked: also callable manually, and the agent reaches for it when it meets a fitting task; it carries reusable discipline, e.g.
/tdd,/code-review,/diagnosing-bugs,/research
Inputs are the current conversation, plan or codebase; outputs are specs, tickets, HTML architecture reports or code changes. /grilling is the underlying interrogation primitive that grill-me, triage and wayfinder all build on; /code-review reviews a diff in parallel along two dimensions, Standards and Spec.
Who it's for
Engineers already running real projects inside coding agents like Claude Code and Codex who don't want to hand their whole engineering process over to some framework. Codebases where the requirements are hard to pin down and that you want to push through a test—implement—review loop see the most direct benefit.
Boundaries
The plugin is a read-only hosted whole package whose updates follow the author's releases and whose content you can't change; what skills.sh installs are ordinary files you own, which won't quietly update themselves behind your back. MIT licensed, and the README's own words are "hack around with them, make them your own". The set doesn't replace your project's own tests, type checking and browser feedback; /improve-codebase-architecture is a survey, not a rescue — it can surface real candidates in an old codebase, but it won't untangle the mud ball for you.