Matt Pocock, known for the Total TypeScript and AI Hero projects, has published the mattpocock/skills repository — 25 instruction files ("skills") for AI agents, compiled from his working .agents directory under the name "Skills for Real Engineers." The collection addresses four typical failure modes in working with agents — from misunderstanding the task to architecture degradation. Installation takes about 30 seconds: one command from the official Claude Code plugin marketplace or via npx for Codex and other agents. The repository is distributed under the MIT license.

image
image

What happened

On February 3, 2026, Matt Pocock opened the mattpocock/skills repository, gathering 25 skills from his own working .agents directory. Each skill is a plain text instruction file that connects to an AI agent as a separate command. Interview skills /grill-me and /grill-with-docs make the agent build a common project vocabulary in a CONTEXT.md file and record accepted decisions in ADRs before a task. Code correctness is maintained by /tdd with a red-green-refactor cycle and /diagnosing-bugs for complex bugs. /to-spec, /to-tickets, /implement, /wayfinder, and /improve-codebase-architecture are aimed at preventing architecture degradation, while separate skills are intended to tame agent chattiness. The set is installed with one command: claude plugins install mattpocock-skills from the official Claude Code plugin marketplace or npx skills@latest add mattpocock/skills for Codex and other agents.

Context

Pocock is known primarily as the author of the educational projects Total TypeScript and AI Hero, and his skills are not a theoretical concept but a dump of personal tools he uses daily. Skills as a format are small, editable instruction files that have grown from personal configs into a distribution layer over Claude Code and Codex. Pocock deliberately contrasts composable files with monolithic process frameworks like GSD, BMAD, and Spec-Kit, which, in his opinion, take control away from the engineer. The code review approach deserves special attention: /code-review is executed by parallel sub-agents, where standards and spec compliance are checked separately — this is a known decomposition technique that reduces the mixing of criteria. The distribution channel is also telling: the official Claude Code plugin marketplace and the skills.sh installer with the npx command show that instruction sets are becoming a standard artifact of developer tooling — in fact, a "practice layer over LLM agents" is forming.

Why this matters for the industry

For the industry, skills have become a de facto format for packaging engineering practices over LLM agents. Teams get a ready-made set of UX patterns — task setting, review, TDD, bug handling — instead of custom prompts, and the "process" part of development becomes cheaper. The claimed effect of a common vocabulary in CONTEXT.md is fewer tokens for agent reasoning. The operational risk of a pilot is minimal: these are files under the MIT license with no new models or infrastructure changes, which can be installed and removed with one command. If the format continues to standardize, independent comparisons of instruction sets and the first quantitative assessments are likely — more from the community than from the author — and skill catalogs may become a standard part of tooling and onboarding, including internal corporate sets tailored to their own standards.

Why this matters for users

If you work with Claude Code or Codex, you can try the set immediately: one-command installation takes about 30 seconds. A reasonable first step is /grill-with-docs before any task so the agent clarifies context, then /tdd for new features, /diagnosing-bugs for complex bugs, and /improve-codebase-architecture every few days on your project. Skill files are plain and editable: they can be adapted to your own process or individual techniques — common vocabulary, ADR decision recording — can be moved to your own configs. Updates are pulled with the npx skills update command, and no infrastructure changes are needed.

What is still unknown / limitations

The effectiveness of the set has not been measured: the sources contain no benchmarks, ablations, or success metrics, so claims about token savings and reduced misunderstanding remain the author's experience and interpretation, not an established result. Review by parallel sub-agents means more runs and tokens, and its impact on final code quality has not been assessed. The question of format portability between different agents and models also remains open; the specific set may become outdated over time, even if the "CONTEXT.md + composable skills" pattern proves long-lived.

Sources

Author

Look at AI, editorial team