What it does

A collection of prompt-based “skills” (think: structured system instructions) for Claude Code, Codex, and similar coding agents. Install via npx skills@latest add mattpocock/skills, pick your poison, and run commands like /grill-me or /tdd inside your agent chat. The setup scaffolds per-repo config for issue trackers, triage labels, and doc directories.

The interesting bit

The whole thing is built around a single insight: agents fail the same ways junior devs do—misalignment, verbosity, missing feedback loops, and entropy. So Pocock weaponized software engineering fundamentals into repeatable agent rituals. The /grill-with-docs skill is the standout: it cross-examines your plan and forces a shared domain language into a CONTEXT.md, which then compresses agent output and keeps naming consistent across sessions.

Key highlights

• /grill-me and /grill-with-docs — structured interrogation before any code gets written
• /tdd — red-green-refactor loop enforced in the agent conversation
• /diagnose — formalized debugging protocol: reproduce → minimise → hypothesise → instrument → fix
• /improve-codebase-architecture — periodic “rescue” pass against codebase rot, informed by your CONTEXT.md and ADRs
• /caveman — ultra-compressed agent responses, claimed ~75% token reduction
• Model-agnostic; works with “any model” per the README

Caveats

• Requires buying into the full ritual stack: CONTEXT.md, ADRs, triage labels, vertical slices. The README warns to run setup before using dependent skills, so this is not a grab-one-and-go situation.
• The “decades of engineering experience” claim is present but unsourced; the skills themselves are opinionated distillations of Pragmatic Programmer/DDD/XP canon.
• Token savings from /caveman are stated but not benchmarked.

Verdict

Worth a look if you’re already using Claude Code or Codex daily and tired of rewriting the same prompts. Skip it if you want fire-and-forget AI assistance without the upfront process tax.