Agent skills for developers who hate vibe coding

What it does

Matt Pocock open-sourced his personal .claude directory as a set of slash-command skills for Claude Code, Codex, and similar agents. Each skill is a structured technique designed to fix a specific failure mode—misalignment, verbosity, missing tests, or architectural rot. They are model-agnostic, small, and meant to be hacked on.

The interesting bit

The standout idea is /grill-with-docs, which conducts a pre-flight interrogation about your plan and simultaneously distills a shared domain language into a CONTEXT.md file. The README notes the payoff is hard to describe—mostly because the value is in the boring part: forcing an agent to understand your jargon before it names your variables.

Key highlights

• /grill-me and /grill-with-docs run a “grilling session” to close the gap between what you want and what the agent thinks you want
• /tdd enforces red-green-refactor loops so the agent writes failing tests before fixing them
• /diagnose wraps a disciplined debugging loop: reproduce → minimise → hypothesise → instrument → fix
• /improve-codebase-architecture and /zoom-out try to combat the entropy that agents accelerate when they write code too fast
• /caveman compresses agent communication to cut token usage by roughly three-quarters while keeping technical accuracy

Verdict

Worth a look if you want granular control over how an agent builds software; skip if you prefer all-in-one frameworks like GSD or Spec-Kit that own the entire process for you.