mattpocock/skills

These are the author's personal agent skills for doing real engineering with coding agents rather than "vibe coding". They are small, editable Markdown skill files, intended to work with any model, and are grouped into engineering skills (planning, specs, tickets, TDD, bug diagnosis, architecture review, code review) and general productivity skills. Because they target agents writing and maintaining code, they sit in developer tooling rather than in a general-purpose AI framework.

What changes in your workflow

Rather than one tool owning the process, the skills are orchestration pieces you invoke by name. A typical flow: /grill-with-docs interviews you about the change and sharpens project terminology into GLOSSARY.md and ADRs; /to-tickets breaks a spec into tracer-bullet tickets with blocking edges; /implement builds each ticket, driving /tdd and finishing with /code-review. /implement-spec instead works a whole spec on one integration branch using subagents across the ready frontier of the task graph. Two invocation classes keep this manageable: user-invoked skills (you type them; they orchestrate) and model-invoked skills (the agent can reach for them when a task fits), with user-invoked skills allowed to call model-invoked ones but never each other (README reference).

The mechanisms that matter

• Shared language over verbosity. /grill-with-docs and /domain-modeling push agreed terms into a glossary and record hard decisions in ADRs. The author argues this also makes naming consistent and reduces token spend on reasoning (grill-with-docs, domain-modeling).
• Feedback loops as the speed limit. /tdd enforces red-green-refactor in small vertical slices, and /diagnosing-bugs runs a gated loop that starts by building a failing reproduction before hypothesising and fixing (tdd, diagnosing-bugs).
• Design treated as ongoing work. /improve-codebase-architecture surveys a codebase and returns deepening candidates as a report, then grills you on the one you pick. The author explicitly calls it a survey, not a rescue, and says it will not untangle a genuinely messy codebase for you (improve-codebase-architecture).

Setup and scope

Install through one plugin route per agent: Claude Code and Codex plugins update themselves; GitHub Copilot (CLI and VS Code) needs a one-time settings entry and updates daily; Gemini CLI is manual and re-run to update; other agents or editable files use the skills CLI, where you must include setup-matt-pocock-skills when offered. Installing both the plugin and the copied files gives you every skill twice (installation section). Then run /setup-matt-pocock-skills once per repository and answer what issue tracker, triage labels and docs location to use — several skills depend on that configuration. The repository is MIT-licensed. The README documents no test suite, benchmark or evaluation of outcomes; the evidence is the author's account of his own daily practice.