Get grilled
The agent asks you questions one at a time until every design decision is settled. Matt does this before every change.
Matt Pocock's skills are “for real engineers, straight from his .agents directory.” Each one does a single job. This page shows the order to use them through a normal workday, with example sessions for the four situations you'll hit most.
claude plugin install mattpocock-skills@claude-plugins-officialnpx skills@latest add mattpocock/skills -a cursor/setup-matt-pocock-skills once in each repo.
Every change goes round the same loop. You trigger the main skills yourself with a slash command. Helper skills join in automatically when the agent needs them.
The agent asks you questions one at a time until every design decision is settled. Matt does this before every change.
Turn the conversation into a spec, then split it into thin tracer-bullet tickets with blocking edges.
Build the tickets test-first at the seams you agreed on, and finish with a two-axis review.
A PR body with a summary, before/after evidence and a call on how risky the merge is.
Ask what slowed the session down and fix your environment, starting with the worst problem.
These are model-invoked. You don't type them. The agent loads them when the work calls for it, though you can also name them directly.
Drawn like a git-flow diagram. Time runs down, and each lane is a place work lives: your conversation, the docs, the tracker, branches and main. Dots are skill runs or files written. Arrows show what feeds what.
Scroll sideways to see all six lanes.
An example day building one feature. The times are only illustrative. The order is the part that matters.
Tell it where issues live, which triage labels you use and where docs go. Every other skill reads these answers, so you only do this once.
Describe the change in one sentence. The agent asks one question at a time, each with a recommended answer. New terms go into the glossary, and hard decisions become ADRs. For non-code plans, use /grill-me.
The whole conversation becomes a spec in your issue tracker. Then it gets cut into tracer-bullet tickets. Each ticket is a thin slice through every layer, and the blocking edges show which ones can start now.
The agent works test-first at the seams you agreed on (tdd), then reviews its own diff against your coding standards and against the spec (code-review). You review the result, not every keystroke.
Say “diagnose this”. Instead of guessing at fixes, the agent works through gates: reproduce, shrink the case, rank hypotheses, add instrumentation, fix, then lock it in with a regression test.
The body has a summary, before/after evidence and an explicit merge-danger call, so your reviewer knows how carefully to look.
/retro suggests fixes to your environment, worst problem first: a slow test command, a missing glossary term, a vague CLAUDE.md rule. If the work isn't finished, /handoff compacts the session into a document a fresh agent can continue from.
It scans for deepening opportunities, shows an HTML report, then grills you on the one you pick. Matt calls it “a survey, not a rescue.”
Pick a situation. On the left is the chain of skills. On the right is a shortened conversation showing how the session goes.
/grill-with-docs Users should be able to save a search and get told when new recipes match it./to-spec#142./to-tickets#143 Save a search end to end (UI → API → DB)#144 Match new recipes against saved searches (blocked by #143)#145 In-app alert badge (blocked by #144)#146 Rename and delete saved searches (blocked by #143)/implement #143/prapplyDiscount. Each line rounds 94.5p up to 95p, so 3 × 95 = 285 instead of round(283.5) = 284. Hypothesis 1 is confirmed./pr/wayfinder We want to move from server sessions to tokens in the web, mobile and admin services.D1 Token format and lifetimeD2 Refresh strategy (blocked by D1)D3 Rollout order across the servicesD4 How existing sessions get logged out (blocked by D3)/handoff/to-spec then /to-tickets#210 published, with 7 tickets. Three are ready now: web, admin, shared library./implement-spec #210/improve-codebase-architectureNotificationService exposes 14 methods, and callers use 3 of themPricing module with a single quote(cart, plan) function, and the 6 files behind it. The term goes into GLOSSARY.md./to-tickets/retroThese are illustrative sessions, shortened to show the shape. Real skill output is longer and depends on your repo.
Choose what's happening. If none of these fit, /ask-matt routes you to the right skill.
Get interviewed until every decision is made. New terms and decisions are written into GLOSSARY.md and ADRs as you go.
The same relentless interview, without touching repo docs. Good for talks, migrations or hiring plans.
Build a map of decision tickets, then resolve them one at a time across sessions.
Build it with tdd at the seams you agreed on, finishing with code-review. For a whole spec in parallel, use /implement-spec.
Just say “diagnose this”. Reproduce, minimise, hypothesise, instrument, fix, regression-test.
Moves issues through a state machine of triage roles, so each one ends up ready, blocked or closed.
Summary, before/after evidence and a call on how dangerous the merge is.
Re-explains the last message using your project's own vocabulary from GLOSSARY.md.
Compacts the conversation into a document another agent can pick up tomorrow.
A survey for deepening opportunities with an HTML report, then a grilling on the one you choose.
Turns a decision you can't make alone into a questionnaire for the right person.
User-invoked skills are the ones you type. They orchestrate. Model-invoked skills are tools the agent picks up when they fit.
The questions are cheap. Fixing wrong assumptions after the code exists costs far more.
When you and the agent use the same words, misunderstandings drop. /wait-what re-explains a confusing reply using those words.
When something goes wrong, /retro turns it into a permanent fix in CLAUDE.md, tests or tooling.