What happens when one of the most respected TypeScript educators in the world decides to open-source the exact AI agent workflow he uses to build production software? The answer is 176,432 GitHub stars in under six months. Matt Pocock, creator of Total TypeScript and one of the most influential voices in developer tooling, published his personal .agents directory as an open-source repository in February 2026. It crossed 176,000 stars by mid-July, making it one of the fastest-growing developer tool repositories on the platform this year. The repo is called skills, and it is not another vibe coding framework. It is a collection of small, composable, and rigorously engineered agent skills that Pocock uses daily to ship production software. The philosophy is stated plainly in the README: Skills for real engineers, not vibe coding.
The repository is MIT-licensed with 15,114 forks. It ships as a Shell-based skill collection that works with Claude Code, OpenAI Codex CLI, Cursor, Hermes, and Kiro. There are two installation methods: a skills.sh-based installer that copies the skills directly into your project for modification, and a native Claude Code plugin that delivers them as a managed, always-current bundle. The install command takes 30 seconds: npx skills@latest add mattpocock/skills. After that, a single setup skill configures your preferred issue tracker, triage labels, and documentation layout, and the entire library is ready to use.
What Makes These Skills Different
Pocock identified four common failure modes in AI-assisted development and built a skill to fix each one. The first is misalignment. The agent does not do what you meant because you and the agent never aligned on the actual goal. The fix is the /grill-me and /grill-with-docs skills, which are relentless interviewing sessions that force the user to think through every branch of a decision before the agent writes a line of code. /grill-with-docs goes further by building a shared domain language and writing it into a CONTEXT.md file that every subsequent agent session inherits, dramatically reducing verbosity and token waste across sessions.
The second failure mode is verbosity. AI agents default to over-explaining because they lack a shared vocabulary with the project. Pocock addresses this by encoding domain language as a first-class artifact. Each project that uses the skills builds a CONTEXT.md file that defines the project's unique terminology. The agent references this file automatically, using 10 words where it previously used 50. This is not a stylistic preference. It directly reduces token consumption, improves code consistency, and makes the codebase more navigable for both humans and agents.
The third failure mode is broken output. The agent produces code that looks right but does not actually run. The fix is a /tdd skill that enforces a red-green-refactor loop. The agent writes a failing test first, then makes it pass, ensuring every change has a verifiable feedback signal. A companion /diagnosing-bugs skill wraps the standard debugging loop of reproduce, minimize, hypothesize, instrument, fix, and regression-test into a single slash command. Both skills are model-invoked, meaning the coding agent can reach for them automatically when the task fits.
The fourth and most important failure mode is what Pocock calls the ball of mud. AI agents accelerate software entropy because they generate code faster than developers can refactor it. The fix is /improve-codebase-architecture, a skill that scans the codebase for deepening design opportunities, presents them as a visual HTML report, and then grills through whichever improvement the developer selects. Pocock recommends running it every few days. The companion /code-review skill runs two-axis reviews of every diff: one axis checks coding standards and Fowler smell baselines, the other verifies faithful spec implementation. Both run as parallel sub-agents so neither pollutes the other's context window.
How Installation and Daily Usage Works
The quickstart is genuinely fast. Run npx skills@latest add mattpocock/skills in any repository, select the skills you want, and pick which coding agents to enable them on. The /setup-matt-pocock-skills command asks three questions: which issue tracker to use (GitHub, Linear, or local files), what triage labels to apply, and where to store documentation files. After that, the workflow is driven by slash commands that chain into each other in natural engineering sequence.
A typical day with these skills looks like this. Start with /triage to move issues through a state machine of review roles. Run /grill-with-docs before starting any new feature to align intent and update the domain model. Use /to-spec to turn the conversation into a structured spec posted to the issue tracker. Use /to-tickets to break that spec into tracer-bullet tickets with declared blocking edges. Run /implement to build the work, which invokes /tdd at pre-agreed seams and closes out with /code-review before committing. For large features that span multiple sessions, /wayfinder plans the work as a shared map of investigation tickets on the tracker. Every skill inherits context from the one before it because everything relies on the same CONTEXT.md and ADR files in the project root.
The skills also ship as a native Claude Code plugin. Inside Claude Code, run /plugin marketplace add mattpocock/skills followed by /plugin install mattpocock-skills@mattpocock. The plugin version is a managed bundle that updates automatically when Pocock ships changes, while the skills.sh version copies editable files into the project for developers who want to hack on them directly. Both installation paths converge on the same setup step and the same daily workflow.
Why 176,000 Developers Flocked to This Repo
The scale of adoption tells a story about the state of AI coding tools in mid-2026. The developer community is past the hype phase. Vibe coding frameworks that promise zero-config magic are losing credibility as developers discover that production software still requires engineering discipline. Pocock's skills repo resonated because it does the opposite of what most agent frameworks do. It gives control back to the developer rather than taking it away.
Frameworks like GSD, BMAD, and Spec-Kit attempt to own the entire development process. They define how work gets done, and debugging the process itself becomes a challenge when something goes wrong. Pocock's skills are intentionally small and adaptable. Each skill solves one problem well. They compose naturally because they share a common context file system and a consistent interface. The /ask-matt skill acts as a router, asking the user what they want to accomplish and recommending the right skill for the job, so beginners do not need to memorize 17 slash commands before getting value.
The comparison to gstack, Garry Tan's 23-agent workflow with 122K stars, is instructive. Both repos hit viral growth by solving the same fundamental problem: AI agents need structure to produce quality output. But they solve it differently. gstack encodes a full sprint cycle as chainable Claude Code agents inside a single session. Pocock's skills are smaller, model-agnostic, and designed to be copied and customized per project. A developer can adopt just the /grill-me skill without committing to the full stack. That modularity is why skills may have broader long-term adoption than any monolithic workflow: you start with one skill, see results immediately, and add more as the need arises.
Who This Is For
This repository is for developers who are already using AI coding agents and want to level up from basic prompt engineering to structured, repeatable engineering practices. It is specifically designed for production engineering, not prototypes or throwaway projects. The skills encode decades of software engineering wisdom from Pocock's career into repeatable prompts that any agent can execute. The /code-review skill references Martin Fowler's refactoring catalog. The /tdd skill follows Kent Beck's red-green-refactor discipline. The /domain-modeling skill draws from Eric Evans on domain-driven design. These are not ChatGPT wrapper prompts. They are executable engineering textbooks.
For solo founders using AI agents to build products, this repo provides something that no other open-source tooling offers: a curated engineering workflow from someone who has taught TypeScript to over a million developers and has been refining his AI agent practice for two years. The skills.sh installer works across every major coding agent, so there is no vendor lock-in. The Claude Code plugin offers a managed experience for developers who want to subscribe and follow along as the skills evolve. Either way, the price of entry is a single npx command and 30 seconds of setup time.