I Tried the Claude Code Skills Repo That Got 77K Stars — Here Is What Works and What Does Not

#claude #ai #programming #api

I kept running into the same problem with coding agents: I would describe a task, the agent would build something, and it was not what I meant. Not broken — just off.

The fix turned out to be surprisingly low-tech. Matt Pocock published a repo of "skills" — small instruction files that go in your .claude directory and change how the agent approaches work. The repo exploded: 77,000+ stars, 6,700+ forks, #1 on GitHub Trending.

I installed it. Here is what I found.

Setup (About 60 Seconds)

npx skills@latest add mattpocock/skills

Pick the skills you want. Select /setup-matt-pocock-skills — it is the bootstrap.

Inside your agent:

/setup-matt-pocock-skills

It asks your issue tracker (GitHub / Linear / local files), triage labels, and docs folder. Done.

Works with Claude Code, Codex, Cursor, or anything that reads .claude/ directories.

The 4 Skills I Actually Use

There are 28 skills. I use 4 regularly.

`/grill-with-docs`

This is the best one. Before you start coding, the agent asks you detailed questions about what you are building. Edge cases, constraints, why you are doing it this way.

The output is a CONTEXT.md file — a shared vocabulary that the agent reads in every future session. One of my projects had a 15-word phrase that got replaced with "materialization cascade." Every session after that was shorter and more accurate.

`/tdd`

Forces test-driven development: tests first, implementation second, verification third. Works great on isolated functions. Gets annoying on complex UI where the tests are hard to specify upfront.

`/diagnose`

Structured debugging instead of the agent guessing. Most useful when the error message does not point to the real problem.

`/caveman`

Extreme concision. The agent says as little as possible and just executes. Perfect for experienced devs who already know what they want.

Everything Else (Quick Map)

Alignment: /grill-me (non-code version), /to-prd, /to-issues
Code quality: /prototype, /triage, /zoom-out
Workflow: /handoff, /write-a-skill, /review (WIP), /setup-pre-commit

Honest Limits

Skills are instructions, not plugins. They do not make the model smarter — they make the conversation more structured.
CONTEXT.md drifts. You need to update it as the project evolves, or re-run the grill.
Agent-agnostic = lowest common denominator. Skills work everywhere but cannot use agent-specific features like Claude Code's /goal.
28 skills is too many to learn at once. Start with the four above.

When It Gets Bigger

Once a team runs multiple agents on the same codebase, the next problem is not prompt quality — it is routing. Who uses which model, how much does each workflow cost, and where do the API keys live.

Skills handle the agent behavior side. For the model routing side, connecting Claude Code CLI to a gateway makes the access path concrete. The EvoLink Claude Code CLI guide documents that setup if your team is at that stage.