Skills For Real Engineers: AI Agent Skills Review

What Are These AI Engineering Skills?

Skills For Real Engineers is Matt Pocock's repo of roughly 25 agent skills — Markdown SKILL.md files with instructions for Claude Code and Codex, stated to work with any model since they're plain LLM agent prompts, not model-specific code. Each skill is a slash command, split into ones you type yourself and ones the agent reaches for on its own. It's MIT-licensed, and built around developer agent workflows rather than replacing them.

Solving Common AI Agent Frustrations

The README names four specific failure modes it was built to fix: agents that build the wrong thing because no one aligned on intent first, agents that pile on jargon because they never learned the project's vocabulary, code that silently doesn't work because there's no feedback loop, and codebases that rot fast because agents write bad code faster than any team can review it.

Core Engineering and Productivity Skills

• ✓/grill-me and /grill-with-docs run a grilling session — the agent interviews you until every branch of the design is resolved; /grill-with-docs also writes a CONTEXT.md domain-modeling glossary and ADRs as it goes.
• ✓/tdd drives a red-green-refactor loop: the agent writes a failing test before touching implementation code, one vertical slice at a time — test-driven development enforced by the prompt itself.
• ✓/diagnosing-bugs gates debugging strategies into fixed phases — minimise, hypothesise, instrument, fix, regression-test — instead of letting the agent guess.
• ✓/improve-codebase-architecture scans a codebase for software-architecture deepening opportunities and renders them as an HTML report you pick through; Pocock calls it 'a survey, not a rescue.'
• ✓/code-review runs two parallel sub-agent passes — one checking coding standards against a Fowler smell baseline, one checking the diff against the original spec — so neither review pollutes the other.
• ✓/ask-matt is a router: type it when you're not sure which user-invoked skill fits your situation.
• ✓/triage moves issues through a state machine of roles, using whatever labels you set during setup.

Getting Started: Installing the Skills

Two install paths, deliberately different in what they trade off. The Claude Code plugin — `claude plugins install mattpocock-skills`, or `/plugin install mattpocock-skills` from inside a session — pulls the whole set as a managed, read-only bundle already listed in Claude Code's official marketplace; it updates automatically when Pocock ships changes, but you can't edit it. `npx skills@latest add mattpocock/skills` (the skills.sh installer) writes the skills into your repo as plain files you own, works with Codex and other agents too, and lets you pick individual skills — just make sure `setup-matt-pocock-skills` is one of them. Pull upstream changes later with `npx skills update`. Installing both leaves you with every skill twice, so pick one. Either way, run `/setup-matt-pocock-skills` once per repo: it asks which issue tracker you use (GitHub, Linear, or local files), what labels you apply when triaging, and where to save generated docs.

Applying the Skills in Your Workflow

Run /grill-me before a non-code decision, or /grill-with-docs before a coding change — either way, the agent interviews you until every branch of the plan is resolved before it lets you start. Once you're building, /tdd takes over: it writes a failing test, then the implementation, one slice at a time, and won't move on until the current slice is green. Hit a bug you can't explain? /diagnosing-bugs walks a fixed sequence — minimise, hypothesise, instrument, fix, add a regression test — instead of trial and error. /ask-matt exists for the moment you forget which command does what.

Considering Other AI Development Approaches

Frequently Asked Questions

Best use cases

• •Kicking off a new feature by running /grill-me or /grill-with-docs first, so the agent asks its questions before it writes any code.
• •Onboarding an agent to an unfamiliar codebase by building a CONTEXT.md glossary with /grill-with-docs, so later sessions spend fewer tokens re-explaining jargon.
• •Driving red-green-refactor on AI-generated features or fixes with /tdd, one vertical slice at a time.
• •Running /improve-codebase-architecture every few days on a codebase that's aging fast under heavy agent-driven commits.
• •Turning a long planning conversation into tracked tickets with /to-spec or /to-tickets before an agent starts implementing.

Who should try it — and who should skip

Reach for Skills For Real Engineers if you already run Claude Code or Codex daily and keep hitting the same walls: misaligned output, jargon-heavy verbosity, code that silently doesn't work, or a codebase getting harder to touch every sprint. Skip it if you're not running an AI coding agent yet, or if you want a framework that owns the whole process for you — Pocock built this specifically not to do that.

Related repositories