mattpocock/skills

What it solves

This project provides a collection of composable, small-scale "skills" (prompt-based instructions) for AI coding agents to prevent common failure modes in software development. It aims to move beyond "vibe coding" by applying rigorous engineering fundamentals—such as TDD, domain modeling, and disciplined debugging—to the AI-assisted development process.

How it works

These skills are implemented as instructions that can be installed into coding agents like Claude Code or Codex. They are categorized into user-invoked skills (which orchestrate workflows) and model-invoked skills (which provide the reusable discipline).

Key workflows include:

• Alignment: Using "grilling sessions" (/grill-me) to ensure the agent understands the requirements before coding.
• Shared Language: Building a CONTEXT.md file to create a shared domain model between the human and the AI, reducing verbosity and token usage.
• Feedback Loops: Implementing a red-green-refactor loop via a /tdd skill to ensure code correctness.
• Architecture: Using tools like /improve-codebase-architecture to scan for opportunities to deepen modules and reduce software entropy.

Who it’s for

Software engineers who use AI coding agents and want to maintain high code quality, architectural integrity, and precise alignment with the agent.

Highlights

• Engineering-First Approach: Integrates TDD, ADRs, and codebase design principles directly into agent workflows.
• Flexible Installation: Supports managed bundles for Claude Code or editable files via skills.sh for other agents.
• Domain Modeling: Tools to sharpen terminology and maintain a shared vocabulary in CONTEXT.md.
• Comprehensive Toolset: Includes specialized skills for triage, spec creation, bug diagnosis, and even generating interactive bash wizards for human-only tasks.