TL;DR
Matt Pocock's Claude Code skills library picked up 3,372 GitHub stars on a single day. Here's what's inside, how to install it, and whether the skills hold up in real projects.
Read next
Matt Pocock made his personal .claude directory public and hit 857 new GitHub stars in a day. Here's what 18 production-tested Claude Code skills look like from a prolific TypeScript educator.
5 min readThe mattpocock/skills repository hit 62k stars on GitHub's weekly trending list with 17 composable Claude Code agent skills that tackle the four failure modes killing real engineering teams.
5 min readA single CLAUDE.md file inspired by Andrej Karpathy's observations about LLM coding mistakes gained 24,000 GitHub stars in one week. Here is what it teaches and why it resonated.
5 min readGitHub's daily trending list on May 13, 2026 has one clear standout: mattpocock/skills, a Shell repository titled "Skills for Real Engineers. Straight from my .claude directory." It picked up 3,372 stars in a single day and appears on both the daily and weekly trending lists - a combination that is genuinely rare and usually signals that something clicked at scale, not just a single share that briefly spiked.
Matt Pocock is well known in the TypeScript world for ts-reset and Total TypeScript. Over the last year he has been building aihero.dev, an AI-native development learning platform with roughly 60,000 newsletter subscribers. The skills repo is not a marketing asset - it is a direct export of the .claude directory he runs on real projects, structured so any developer can install it in minutes.
The timing matters. Developers have been accumulating .claude/ directories, writing ad-hoc skills, and passing around repo links for months. A curated, opinionated set from someone who has clearly thought through the failure modes lands differently than another generic agent framework promising to solve everything.
The README opens with an honest critique of existing approaches. Frameworks like GSD, BMAD, and Spec-Kit solve AI coding by owning the process end to end. The trade-off is that when the process has a bug, you have limited surface area to debug it, and the framework's opinions overwrite your own.
mattpocock/skills takes the opposite position: small, composable skills that target the four specific failure modes behind most AI coding breakdowns.
Misalignment - The agent starts building before you and it share a mental model of the problem. The /grill-me skill runs an exhaustive planning interview before any code gets written. The /grill-with-docs variant adds domain model validation against your existing documentation, so the agent's understanding is grounded in what your project actually does rather than what it inferred.
Verbosity - AI coding sessions accumulate noise. The solution here is a CONTEXT.md file that establishes shared domain language for the project. Once that exists, the agent stops asking you to re-explain the same concepts session after session.
Non-functional code - Rather than hoping the agent writes tests, /tdd enforces a strict red-green-refactor loop. When something is already broken, /diagnose walks through a structured sequence: reproduce, minimize, hypothesize, instrument, fix, verify. That sequence forces methodical debugging instead of speculative patching.
Architecture degradation - Long-lived AI-assisted codebases tend to accumulate complexity silently. /improve-codebase-architecture and /zoom-out are designed to surface that drift before it compounds into something expensive to untangle.
Beyond those four, the library ships practical utilities. /to-prd synthesizes a messy conversation thread into a clean product requirements document. /to-issues turns a plan into GitHub or Linear issues. /prototype spins up throwaway code for design validation. /caveman compresses agent communication by roughly 75% for token-heavy sessions. /handoff compacts conversation history into a format a new agent session can pick up cleanly.
There is also git-guardrails-claude-code, a skill that blocks a set of destructive git commands at the agent level - useful on any repo where you want the agent to think before running reset --hard or push --force.
Get the weekly deep dive
Tutorials on Claude Code, AI agents, and dev tools - delivered free every week.
From the archive
May 12, 2026 • 8 min read
May 12, 2026 • 8 min read
May 12, 2026 • 9 min read
May 11, 2026 • 7 min read
The entire library installs with one command:
npx skills@latest add mattpocock/skills
After that, run /setup-matt-pocock-skills inside your coding agent. The setup flow asks you to select which skills you want active, pick your issue tracker (GitHub, Linear, or local markdown files), configure triage labels, and set a path for documentation storage. The whole thing takes about five minutes.
Once configured, you invoke skills directly in Claude Code. To start a session with proper alignment:
/grill-me
To kick off test-driven work on a specific feature:
/tdd
To compact a long conversation before handing off to a new session:
/handoff
The skills load on demand. You do not need all of them active at once, and the install flow makes it easy to add or remove individual skills later without touching anything else.
This library is a good fit for developers who are already using Claude Code regularly and want to bring more structure to their workflow without committing to a full agentic framework.
If you are just starting with Claude Code, the skills will make more sense after you have hit the failure modes they address. Burning a session on misalignment or watching the agent silently degrade your architecture is useful context for understanding why /grill-me exists.
If you are running Claude Code on a production codebase with a team, the git-guardrails-claude-code skill and the /to-issues workflow are immediately practical. They map onto existing processes rather than replacing them. /to-prd is also useful in team settings where you want a durable artifact from a planning conversation rather than just a resolved thread.
If you are already deep into a custom .claude/ setup, treat this as a reference implementation. Each skill is small enough to read in a few minutes, and the philosophy behind them - alignment before coding, feedback loops over hope, domain language over repetition - is worth internalizing even if you adapt or replace the implementations.
Skills as a pattern have been central to coverage here since the ecosystem started consolidating. The DevDigest skills directory at skills.developersdigest.tech tracks the growing library of community skills, and mattpocock/skills is now indexed there.
For readers following the series on why skills beat prompts for coding agents, this repo is a concrete implementation of those principles. The /grill-me and /grill-with-docs skills are the alignment-first approach put into practice - not as theory, but as something you can install today.
If you are building your own skills, the write-a-skill skill in this package is worth examining. It provides a template and workflow for turning a recurring workflow into a reusable skill with consistent behavior. That patterns directly with the DevDigest hooks setup at hooks.developersdigest.tech, where pre-session alignment is the highest-leverage place to inject structure into an automated workflow.
The /handoff skill also pairs well with multi-agent setups. If you are running parallel Claude Code sessions on the same project, compacting conversation history into a clean handoff document before spawning a subagent reduces context bleed significantly.
The skill set is well-considered. The four failure modes it targets are real, and the solutions are proportionate - they add structure where structure helps without burying you in process overhead.
A few limitations to flag. The install tooling assumes npx is available and a recent Node version. The skills are written for Claude Code specifically; most will not port cleanly to other agents without adaptation. The caveman skill's 75% compression claim is from the README and has not been independently verified - token reduction depends heavily on session content and conversation patterns.
The library is also opinionated about issue tracking and documentation storage. If your team uses a system outside GitHub, Linear, and local markdown files, the setup flow may produce configuration that does not integrate with anything. Read through the setup carefully before assuming it connects to your existing workflow.
None of these are deal-breakers. The core skills - /grill-me, /tdd, /diagnose, /handoff - are solid and address pain points that show up in almost every production AI coding workflow. The star velocity suggests the developer community agrees.
Technical content at the intersection of AI and development. Building with AI agents, Claude Code, and modern dev tools - then showing you exactly how it works.
MCP server directory and ranking site. Tracks weekly downloads, GitHub stars, and build status across 5,000+ servers.
View ToolOpenAI's coding agent for terminal, cloud, IDE, GitHub, Slack, and Linear workflows. Reads repos, edits files, runs comm...
View ToolFast Rust-based formatter and linter for JavaScript and TypeScript. One tool replaces Prettier and ESLint with sub-secon...
View ToolCatch broken SKILL.md files in CI before they hit your team.
View AppSpec out AI agents, run them overnight, wake up to a verified GitHub repo.
View AppTurn a one-liner into a working Claude Code skill. From idea to installed in a minute.
View AppConfigure model, tools, MCP, skills, memory, and scoping.
Claude CodePrefix prompts with ! to run shell commands directly, bypassing Claude.
Claude CodeFile discovery via pattern matching across the repository.
Claude Codeobra/superpowers picked up 1,600+ GitHub stars in a single day by solving the real problem with coding agents - they ski...
Matt Pocock's .claude skills directory hit 74.7k GitHub stars, offering structured workflows that address the four most...
Matt Pocock released production-ready agent skills straight from his .claude directory. With 66k stars and 16k added thi...
The mattpocock/skills repository hit 62k stars on GitHub's weekly trending list with 17 composable Claude Code agent ski...
obra/superpowers is a composable, opinionated methodology for AI coding agents - structured workflows covering planning,...
A single CLAUDE.md file inspired by Andrej Karpathy's observations about LLM coding mistakes gained 24,000 GitHub stars...

New tutorials, open-source projects, and deep dives on coding agents - delivered weekly.