Thirty-nine skills in one repo, and every guide does the same thing: list all of them. That leaves you no closer to knowing which Matt Pocock skills deserve a slot in your setup. This ranking covers the seven worth your time, ordered by what people report. Each one fixes a single problem, and each entry says who should skip it.
To build it, we scouted the internet for the skills real humans actually rate. That meant reading Hacker News threads, a YouTube review, a developer blog and GitHub issues. Bad reports count too.
Need a skill for a job none of these cover? Create your own Claude skill from a few real examples, and download it as a ZIP.
The Rundown
- Claude Builds The Wrong Thing: Grill Me. You agree the plan before anything gets built.
- Every Session Starts From Zero: Grill With Docs. It records project terms and decisions as you answer.
- The Job Is Too Big For One Session: Wayfinder. Turn one huge idea into decisions, one at a time.
- Your Tests Pass And Prove Nothing: TDD. Make it write a failing test before any code.
- Claude Guesses At Fixes: Diagnosing Bugs. Make it reproduce the bug before hunting for the cause.
- Your Session Fills Up Mid Job: Handoff. Pack the work into a summary for a fresh session.
- The Codebase Keeps Getting Harder To Change: Improve Codebase Architecture. Pick one refactor from its report and grill it.
The Best Matt Pocock Skills, Ranked
Grill Me
Price | Free, MIT license |
Install | Plugin, or npx skills add with grilling |
Platforms | Claude Code, Codex, GitHub Copilot, Gemini CLI, others via skills.sh |
Best for | Agreeing the plan before anything gets built |
The problem. You agree a plan with Claude, then it builds something else.
What it does. Type /grill-me at the start of your prompt, then describe your plan. That line calls a second skill named grilling, which maps the plan as a tree of decisions. Each round brings numbered questions with a recommended answer, worded so “yes” accepts it. It looks up facts itself and leaves the decisions to you. It will not act until you confirm you both agree. One install trap: grill-me alone fails, because it needs grilling. A GitHub issue on the repo reports it. Install both.
What people found
- A developer on Hacker News says a planning session without it brings maybe 2 questions from Claude. With it: 22.
- Theo, on their channel, ran the grilling text on a project called Lakebed. The first round opened with who the user is and who pays. Their take: “These are actually good questions.” They also admit to “disagreeing with the recommended answers pretty often”, and say that helped.
- A developer on Hacker News says it “felt like a step change in the output quality” with 10 or more questions on complex features.
- A user in GitHub issue 1145 gave specific answers against the recommendation. They say “It documents what it recommended anyway.” A second user reports the same on a hosted MCP server.
- A developer on Hacker News likes the idea of clarifying questions but finds the public skill “very mumbo-jumbo”.
Skip it if the change fits in one line, because a round of questions will only slow you down. If you are coding inside a repo, jump to the next entry. Pocock lists /grill-me “for non-code uses”.
Grill With Docs
Price | Free, MIT license |
Install | Plugin, or npx skills add |
Platforms | Claude Code, Codex, GitHub Copilot, Gemini CLI, others via skills.sh |
Best for | Coding in a real repo with lasting decisions |
The problem. Claude asks again what “customer” means in your app. Every new session starts from zero, so you re-explain the same terms and the same decisions.
What it does. Type /grill-with-docs. It runs the grilling interview plus a domain-modeling skill in one go. As you answer, it writes a glossary of your project terms and decision records into docs/adr/, so later sessions read them instead of asking.
What people found
- Colin Eberhardt, CTO at Scott Logic, rebuilt a karting app called KartLog. A roughly 5-minute talk brought “some quite sensible questions”. It produced a context file of about 45 lines and one decision record. Their verdict: “So far, so good.”
- A developer on Hacker News, who had never written Swift, spent almost 6 hours answering questions on a Swift app. The output was two files and almost 20 GitHub issues, then almost 25 hours of Codex work they called impressive.
- A user in GitHub issue 1303 replied “lgtm” to the closing summary. The agent started editing code they had not asked for. They reproduced it in Claude Code.
- A user in GitHub issue 299 says decision records get written for work not built yet, and other skills treat them as in force.
Skip it if you build throwaway code, where a glossary nobody reads just adds files. Eberhardt also left one question open: “I wonder how it would cope with something less familiar?”
Wayfinder
Price | Free, MIT license |
Install | Plugin, or npx skills add |
Platforms | Claude Code, Codex, GitHub Copilot, Gemini CLI, others via skills.sh |
Best for | Planning work that spans many sessions |
The problem. You have an idea too big for one session. You can see the shape of it, but the decisions inside are a blur.
What it does. Type /wayfinder plus the loose idea. It charts the job as a map issue on your tracker, then files child tickets that are decisions to make, not build tasks. It works them one at a time until nothing is left to decide. It plans by default and does not build. One setup step first: run /setup-matt-pocock-skills so it knows which tracker to file into.
What people found
- A developer on Hacker News building a product from scratch calls it “highly recommended”. It helps them “make sure the path through the fog is clear enough”.
- A developer on Hacker News used it for the first stage of a 2D pixel game, in plain JS and HTML. Opus 5.5 thought for 30 minutes, and they asked for a few changes.
- A developer on Hacker News calls wayfinder maps “a Godsend”.
- A commenter on GitHub issue 578 ran it on a large greenfield repo. It produced implementation-ready tickets despite missing decisions, and the build later invented several idempotency systems the docs prohibited.
- A developer on Hacker News paired it with grilling for a shopping list app. They call the result “absolutely ridiculous”, with Claude still expanding the scope.
Skip it if the job fits in one session. The tracker setup is wasted on something small.
TDD
Price | Free, MIT license |
Install | Plugin, or npx skills add |
Platforms | Claude Code, Codex, GitHub Copilot, Gemini CLI, others via skills.sh |
Best for | Features where a wrong result costs you |
The problem. Your tests go green and you still do not trust them. The agent wrote the assertion the same way it wrote the logic, so a wrong answer sails through.
What it does. Ask for test-first work, or say red-green-refactor, and the TDD skill starts. Before any test, it writes down the seams, meaning the public boundaries to test at, and waits for you to confirm them. Then one failing test, one minimal fix, repeat. It also bans tests that recompute the expected value the same way the code does.
What people found
- A developer on Hacker News says grill-me and tdd “massively helped”, but the code is still “hard to follow and bloated”.
- A user in GitHub issue 1210 ran it on a bug fix. A compile error against a method that did not exist counted as the failing test. The bug never got reproduced, and no regression guard was written.
- Colin Eberhardt watched the red and green loop run under /implement and was delighted. Then the agent wrote CRUD tests for about 15 lines of adapter code that mocked Supabase. Their finding: “the tests do more to validate the mocks”.
- A user in GitHub issue 479 says /implement almost never runs its tdd and code review steps.
Skip it if a wrong result costs you nothing. Even Pocock has mixed feelings on TDD with agents, per The Pragmatic Engineer, and now asks agents for proof the code works, with or without it.
Diagnosing Bugs
Price | Free, MIT license |
Install | Plugin, or npx skills add |
Platforms | Claude Code, Codex, GitHub Copilot, Gemini CLI, others via skills.sh |
Best for | Hard bugs and performance regressions |
The problem. Something breaks and Claude jumps straight to a fix. It edits three files on a guess, and nobody has reproduced the bug yet.
What it does. It is model-invoked, so there is no command to remember. Say something like “diagnose this: POST /orders returns 500 on large carts”, or just report that something is broken, and it starts. Its first phase builds a feedback loop that goes red on that bug. That might be a curl script, a failing test, or a bisection harness. Only then does it chase the cause.
What people found
- Theo says it fired on its own over the last few days and “has helped me a ton already”. Their agents find root causes and explain what is wrong more effectively.
- A user in GitHub issue 578 ran it for four days on gpt-5.6-sol. It fired on quick questions and built a mock reproduction first, causing long reply delays.
- A commenter on the same issue, also on gpt-5.6-sol, hit it on simple CSS bugs. It launched a headless browser and a full red and green run. Their workaround was to make the skill user-invoked only.
Skip it if your bugs are small and visual. Both bad reports show the wait it adds. Pocock replied on that issue that the skill could start lighter and escalate. The issue was still open on 2 November 2026.
Handoff
Price | Free, MIT license |
Install | Plugin, or npx skills add |
Platforms | Claude Code, Codex, GitHub Copilot, Gemini CLI, others via skills.sh |
Best for | Moving work into a fresh session |
The problem. The session is nearly full and the job is only half done. You start a fresh one, and the new agent has to learn the whole task again from scratch.
What it does. Type /handoff plus a line on what the next session is for. It writes a summary to your OS temp folder and names the skills the next agent should call. It points to specs and issues in place of pasting them, and it redacts secrets.
What people found
- A user on GitHub issue 622 says it “works well in Claude Code CLI”. In Codex though, the /tmp file had been cleared before the new session started.
- A user on GitHub issue 306 watched a handoff pass on “not built yet” without checking the code. The next session rebuilt a feature that already existed. A second user there hit the same failure, then moved the save path to a durable folder on macOS.
Skip it if your jobs rarely fill a context window. And if you work in Codex, move the save path somewhere durable before you rely on it. Whatever the agent, read the document before you trust it.
Improve Codebase Architecture
Price | Free, MIT license |
Install | Plugin, or npx skills add |
Platforms | Claude Code, Codex, GitHub Copilot, Gemini CLI, others via skills.sh |
Best for | Older repos with tangled modules |
The problem. You move one piece of the code and two unrelated screens break. Every change now costs more than the last, and the reason is buried in the module layout.
What it does. Type /improve-codebase-architecture. It reads recent commit history for hot spots, sends a sub-agent through the code, and writes an HTML report of refactor candidates. Each candidate gets before and after diagrams plus a strength badge, such as Strong or Speculative. Pick one and it grills you on that pick.
What people found
- A user in GitHub issue 923 watched a refactor pass with types, tests and review all green. It still changed the UI by dropping a conditional-render branch. Every gate read code, and none looked at the rendered app.
- A user in GitHub issue 651 got the report as unstyled raw HTML. It loads Tailwind and Mermaid from CDNs, and a second user hit the same on a home network running Pi-hole.
Skip it if your repo is young and small, since there is little there to find. And skip it if you cannot look at the rendered app yourself after each refactor. The reports found on this skill so far are bug reports, not success stories, which is why it sits last in this ranking.
Which Matt Pocock Skill Should You Install First
Install Grill Me first, and install it with the grilling skill it calls. It is the smallest step in the pack, and it fixes the costliest mistake: building the wrong thing. Run npx skills@latest add mattpocock/skills, and when the installer asks which skills to take, pick grill-me, grilling and setup-matt-pocock-skills. Then run /grill-me on your next feature.
Once you are working in a real repo, move to Grill With Docs.
Hold off on the /wizard skill for now. Colin Eberhardt at Scott Logic reports its script sent them to pages that no longer existed, and “the shell script has no escape hatch”.
And do not install all 39 on day one. Theo says to read the files and take the parts you want. If the agent is not set up on your machine yet, install Claude Code on Mac first. Want more lists built this way? See the best anti-AI-slop Claude skills and the best Claude animation skills.
