AGENT SKILL

askmatt

The one command to remember after start-project: describe what you want and it finds and runs the right skills, flows, agents and plugins - including starting build-ticket, harden or ship when that's what you asked for. Works in Cowork and in the Code tab. Matt Pocock's ask-matt router extended to everything available (this plugin's skills and agents, other installed plugins, connectors and MCP servers, and ones worth installing). Loads the project's truth, looks at the live app when anything visible is involved, routes to the right flow, runs it with approval gates and hands changes to the Code tab. Use when the user says "askmatt", "/askmatt", "ask matt", "/iterate", "iterate", "hvad skal jeg bruge til", "jeg vil ændre", "det ser forkert ud", "tilføj en funktion", "review det vi har bygget", "tjek preview", or shares a screenshot of their app.

License 许可
MIT
最近更新
2026-10-07
Stars
★ 0
Forks
0

01 前言 Front Matter

name
askmatt
description
The one command to remember after start-project: describe what you want and it finds and runs the right skills, flows, agents and plugins - including starting build-ticket, harden or ship when that's what you asked for. Works in Cowork and in the Code tab. Matt Pocock's ask-matt router extended to everything available (this plugin's skills and agents, other installed plugins, connectors and MCP servers, and ones worth installing). Loads the project's truth, looks at the live app when anything visible is involved, routes to the right flow, runs it with approval gates and hands changes to the Code tab. Use when the user says "askmatt", "/askmatt", "ask matt", "/iterate", "iterate", "hvad skal jeg bruge til", "jeg vil ændre", "det ser forkert ud", "tilføj en funktion", "review det vi har bygget", "tjek preview", or shares a screenshot of their app.

02 资源文件 2

03 SKILL.md 正文

askmatt (ask-matt, for the whole toolbox)

The user remembers two commands: /start-project for a new idea and /askmatt for everything else. The others (/build-ticket, /harden, /ship, /grill-design, /audit-loop) are shortcuts; when the user's words ask for them, start them yourself and say so in one line:

  • "byg", "byg ticket #n", "tag næste" (Code tab) → product-studio:build-ticket (in Cowork: write the Code-tab prompt instead)
  • "er PR'en klar", "test PR'en", "tjek PR'en", "gennemgå PR'en", "må det merges" → product-studio:harden (one verdict for one change)
  • "gennemgå alt", "audit alt", "test alt", "kør et loop", "kvalitetstavle", "hver uge" → product-studio:audit-loop (every area of the product, round by round)
  • "grill mig på designet", "hjælp mig med at vælge designet" → product-studio:grill-design
  • a named UI site (shadcn, Motion, Kokonut, Bklit, 21st.dev, Magic UI, Aceternity), "find en komponent", "hvordan kan vi animere", "en anden graf" → product-studio:ui-sources
  • "lancér", "go live", "afslut projektet" → product-studio:ship
  • a brand-new product idea → product-studio:start-project

You don't remember every tool, so look it up: every run starts from a fresh inventory and the catalog. A flow is a path through the tools. Most work runs along one main flow; on-ramps merge onto it; the rest is standalone or a vocabulary layer underneath. Danish to the user, English in files. Say which route you picked in one line so the user can override it.

When the user asks what askmatt does, or on the first run in a project, show references/workflow-map.md. On every run, name the route you picked and the skills/agents it will use in one line.

Surfaces: Cowork and the Code tab

The same skill runs in both; it adapts, it doesn't change its judgement.

  • Cowork (planning, browser, connectors): routes, decides with the user, updates the docs, and hands work to the Code tab through build-handoff. No product code.
  • Code tab (Claude Code in the repo, with this plugin installed there): the same routing, but after "godkendt" it builds right here: product-studio:build-ticket per ticket (Superpowers: worktree → writing-plans → subagent-driven-development/executing-plans with test-driven-development and verification → PR via finishing-a-development-branch), then /harden for an independent verdict. The builder never judges its own work: the verdict comes from harden's inspector.
  • No arguments (/askmatt alone), on either surface: give a status briefing first, in ≤10 Danish lines, from the project truth below: what the product is (one line from PRODUCT.md), what's built, what's in progress (open PRs/tickets), the last PLAN-REVIEW-LOG.md verdict, known risks, and the 3 most valuable next steps. Then ask which one.

Batch requests ("ret det her, det her og det her, og kig designet, leaks og visualiseringerne igennem")

Split the message into numbered items and classify each with the routes below in one short table (item → route → surface). Order them: blockers and broken things first, then changes, then reviews. Review items ("kig det igennem", "grundig gennemgang", "leaks", "visualiseringer") route to /harden (scoped to the area if one is named); change items go through the main flow. Confirm the table with the user once, then work through it, reporting per item.

0. Load and look (every run)

  1. Project truth, cheaply and in this order, only what exists: AGENTS.md, docs/log.md (last entries), CONTEXT.md, PRODUCT.md, docs/tech.md, docs/toolchain.md, DESIGN.md, ADR titles, graphify-out/GRAPH_REPORT.md. Then recent commits, open PRs and issues (GitHub connector, or gh if signed in). Delegate code reading to the code-explorer agent; don't read source wholesale. No such docs? Offer start-project (new product) or build-handoff steps 1-3 (existing code without docs).
  2. Toolbox inventory. The skills listed in this session, ListPlugins/ListSkills/ListConnectors when present, and the catalog references/catalog.md (canonical tool per job, alternates, agents, plugins, MCPs). Anything the route needs that isn't available → section 5.
  3. Look before talking. If anything visible is involved, run a short product-studio:visual-qa pass first (PR preview, else production). Compare a user screenshot with the live page and DESIGN.md. Talk about what you saw.

1. The main flow: idea → ship

  1. Classify the change (Superpowers brainstorming's rules; say it aloud; when in doubt, the heavier path; the ratchet only goes up): spike (we need to see or try something before deciding) → prototype, throwaway; bounded (an existing flow, one seam, no schema/RLS/auth change) → steps 1-3 light, then one ticket labelled size:bounded; architectural (new area, schema, several seams) → the full flow with to-spec + to-tickets. Approving the scope is not approving a plan that doesn't exist yet.
  2. Sharpen by interview. product-studio:grilling + product-studio:domain-modeling (Matt's grill-with-docs pair): stateful, keeps CONTEXT.md and ADRs. No repo? grill-me (same interview, stateless).
  3. Branch: can every question be settled in conversation? If one needs a runnable answer (state, business logic, a UI you have to see), detour through product-studio:prototype: logic → a clickable HTML artifact; UI → structurally different variants on one page with a switcher (an artifact while planning; real ?variant= routes only in the Code tab). Bridge with handoff only if the prototype lives in another directory or tool.
  4. Update the truth the change touches: PRODUCT.md (define-product), docs/tech.md + ADR (define-tech, with a supabase-security-reviewer pass on any data or permission change), DESIGN.md (define-design → impeccable), docs/toolchain.md (define-toolchain, whenever a service, plugin or MCP is added or removed).
  5. Branch: multi-session build? Yes → to-spec → to-tickets (tracer bullets with blocking edges) via product-studio:build-handoff. No → one ticket via build-handoff. Either way the Code tab runs build-ticket per ticket (worktree, plan + approval, test-first build with builder-side reviews, PR with its rulings), clearing context between tickets.
  6. Close the loop. PR preview → /harden (independent verdict; light changes may use visual-qa + review agents) → FIX-FIRST: fix tickets back to step 4, at most 2 rounds → SHIP: merge go-ahead to the user → docs/log.md.

Context hygiene. Keep steps 1-4 in one unbroken context so the interview, spec and tickets share the same thinking. Near the smart-zone limit (~150k tokens), compact at the nearest phase boundary, never mid-phase. Push self-contained work (research, reviews, audits, scouting) to subagents.

2. On-ramps (situations that generate work, then merge onto the main flow)

Situation Route
Look & feel ("det ser billigt ud", "kedeligt", "grimt", "ligner AI", hover, cards, fonts, layout) (step-0 visual-qa findings) → impeccable critique (read-only, fine on either surface) → local problem: grill-design pass C for just those parts (rendered this-or-that, "du bestemmer" allowed), whole look: grill-design redesign mode → define-design step 6 → pass C; then pick the matching impeccable command (polish, bolder, quieter, typeset, layout, colorize, animate, delight, impeccable's own harden for UI edge cases (not our /harden), onboard); anydesign element mode on a reference and ui-ux-pro-max lookup when a new look is needed; show the proposed change as an artifact mock → "godkendt" → DESIGN.md/component-states update if the system changes → one ticket that tells the Code tab which impeccable command to run on which component. Impeccable's editing commands run only in the Code tab, after "godkendt", and the result goes through /harden.
New feature, one session main flow 1 → short design in chat → "godkendt" → one ticket
Big feature / new area main flow 1-4 with to-spec + to-tickets; code-architect agent for an implementation blueprint when the codebase is large
Huge, foggy effort wayfinder (decision tickets, one resolved per session) → to-spec when the fog clears. Never for a well-scoped feature.
"Det er langsomt" (slow page, slow query, Lighthouse/CWV red) performance-optimization (measure → fix → verify, keep or revert) + web-performance-auditor agent (real numbers only) → ticket with the measured before/after. A slowdown that started with a change is a regression → diagnosing-bugs.
Rename, remove or migrate a column, table or feature deprecation-and-migration (expand → migrate → contract, a down path, batched backfill) + define-tech (ADR) + supabase-security-reviewer → expand/contract tickets; the schema change and the code that depends on it never ship in the same deploy
Something's broken reproduce in the browser (visual-qa: console, network) → ordinary bug: systematic-debugging; hard, intermittent or regression: diagnosing-bugs (red feedback loop first) → ticket with a regression test. Post-mortem finds no seam → improve-codebase-architecture.
Continuous quality ("gennemgå alt", "audit alt", "kvalitetstavle", "loop", "hver uge") /audit-loop: measure every area with a fixed tool → independent judge → verified findings → at most 3 tickets per area → build-ticket → /harden → measure again; hard stop rules; Danish scoreboard; optional weekly run
One change's verdict ("er PR'en klar?", "leaks i den her PR?", pre-merge or pre-launch) /harden (product-studio:harden): static gates, RLS proofs, leak sweep, corner cases, E2E + design conformance, independent inspection, verdict in PLAN-REVIEW-LOG.md, fix tickets + Code-tab prompt on FIX-FIRST
Second opinion from another model ("få Codex til at kigge") claudex-route for a recommendation or one scoped handoff; claudex-loop for plan review → build → cross-provider inspection (needs both CLIs, Code tab)
Preview check before merge (light, one screen) visual-qa on the preview + parallel review agents: pr-code-reviewer, silent-failure-hunter, pr-test-analyzer, type-design-analyzer (TS-heavy diffs), supabase-security-reviewer (data diffs), and code-review (standards + spec) → one merged verdict as a PR comment
"Hvordan er det bygget / er det godt?" subagents in parallel: code-explorer (how it works), code-review against the first commit or a tag, improve-codebase-architecture, supabase-security-reviewer, toolchain-auditor (drift), impeccable audit on key screens → one plain-Danish report with the top 5 actions, each routable back onto the main flow
Bugs and requests piling up (arrived raw, not from to-tickets) triage → agent-ready issues
New capability with a new service ("vi skal have betaling", "send e-mails", "AI-opsummering") full main flow, not a shortcut: grilling on the product side (for payments: pricing, trial, VAT/moms, invoices, what happens on failed payment) → define-tech (tables, webhooks, RLS, supabase-security-reviewer pass; api-and-interface-design for idempotent webhooks and retries; security-and-hardening for SSRF, AI input and abuse) → integration-scout + define-toolchain re-audit (sandbox/test keys only, webhook endpoints vs Vercel preview protection) → to-spec + to-tickets
Just a tool ("installer X", a new MCP or plugin, no product change) integration-scout → define-toolchain re-audit → user clicks
Tool weirdness ("Code-fanen opfører sig mærkeligt efter …") toolchain-auditor on hooks, skill collisions, MCP overlaps, instruction files; a build-ticket run that looped, ignored its plan or cost too much → /diagnosing-superpowers (Code tab)
Instruction files feel stale claude-md-improver, pointed at AGENTS.md (it looks for CLAUDE.md by default; CLAUDE.md stays @AGENTS.md; Next's block and our managed block are not rewritten)
Finish the product / go live ("afslut", "gør klar til lancering") /ship (product-studio:ship)
Pre-launch check only (no release yet, just "how far are we?") /harden full scope with claude-security Scan codebase + engineering:deploy-checklist (if installed) + define-toolchain re-run + the CI gate from harden turned on
Mid-merge or rebase conflict (Code tab) resolving-merge-conflicts: resolve by intent, hunk by hunk, never --abort
A step only a human can do checklist with links and exact clicks; wizard only if the user runs a script in their own terminal

3. Codebase health (upkeep, not features)

improve-codebase-architecture surveys for deepening opportunities; picking one generates an idea for main flow step 1. codebase-design is the bench for the chosen module (deep modules, seams, adapters). code-simplifier and comment-analyzer agents after a feature lands.

4. Vocabulary underneath and standalone

  • domain-modeling: when the words are the problem (an overloaded "sag" or "kunde"), or to record an ADR.
  • codebase-design: module-shape vocabulary shared by tdd and improve-codebase-architecture.
  • research: outside facts from primary sources, in the background, as a cited file in the repo; feeds main flow step 1. Context7 for library docs.
  • to-questionnaire: the answer is in someone else's head.
  • wait-what: a message didn't land; re-pitch simply with CONTEXT.md terms.
  • teach: the user wants to understand a concept properly ("hvad er RLS?") over several sessions.
  • handoff: work travels to another tool (Codex), folder or person.
  • writing-for-agents: when writing skills, AGENTS.md sections or docs agents will consume.
  • writing-skills (Superpowers, user-invoked): write or pressure-test a skill with subagents before trusting it.
  • idea-refine (user-invoked): the idea is still raw; diverge first, then grill.
  • source-driven-development (user-invoked): framework code must follow the current official docs (via Context7), cited.

5. When the toolbox is missing something (discovery)

  1. Check the catalog for the canonical pick for that job.
  2. If it isn't available here: search the org catalogs (SearchPlugins, SearchSkills, SearchMcpRegistry) and, if needed, dispatch integration-scout for the best official option with a safe config.
  3. Offer it with an install card (SuggestPluginInstall / SuggestConnectors) or exact steps (Code-tab /plugin install …, a custom connector URL). Adding a tool always routes through define-toolchain so the conflict register stays true.
  4. Meanwhile use the documented fallback; never pretend a missing tool ran.

Agents named here that aren't available as subagent types on this surface: spawn a general-purpose subagent with the body of ../../agents/<name>.md.

6. Agree, then hand off

Every route ends the same way: a short summary (what, why, how it will look, how we'll know it works), "godkendt", docs updated where the truth changed, then product-studio:build-handoff for the ticket(s) and the Code-tab prompt. Don't write product code in Cowork unless the user explicitly asks.

7. Phase boundaries

At the end of a phase, in order (first yes wins): continue if the next phase needs this reasoning or there's room; clear/start fresh if nothing here matters next; handoff only for a new tool, directory or person; subagent if the task can run AFK (reviews, research, QA); otherwise compact with an instruction naming what the next phase needs. Details: ../ask-matt/PHASE-BOUNDARIES.md.

Precondition

The repo is configured once by build-handoff (which runs setup-matt-pocock-skills). If docs/agents/issue-tracker.md is missing, run build-handoff steps 1-2 first.

04 同仓库更多 Skills

phillippasqualato
phillippasqualato/PAPcSetup

writing-skills

Use when creating new skills, editing existing skills, or verifying skills work before deployment [Superpowers; user-invoked. Use to write or pressure-test a skill, e.g. product-studio changes.]

★ 0
phillippasqualato
phillippasqualato/PAPcSetup

writing-plans

Use when you have a spec or requirements for a multi-step task, before touching code [product-studio: canonical Code-tab execution, entered via build-ticket; user-invoked otherwise]

★ 0
phillippasqualato
phillippasqualato/PAPcSetup

writing-for-agents

Writing documents for agents. Use when creating or editing skills, or modifying AGENTS.md or CLAUDE.md.

★ 0
phillippasqualato
phillippasqualato/PAPcSetup

wizard

Generate an interactive bash wizard that walks a human through steps only they can perform. Use when provisioning infrastructure, setting up credentials or CI secrets, walking an unfamiliar third-party dashboard, or running a one-off migration or cutover. Don't invoke this for steps the agent can perform itself.

★ 0
phillippasqualato
phillippasqualato/PAPcSetup

webapp-testing

Toolkit for interacting with and testing local web applications using Playwright. Supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs. [Python alternative to playwright-skill (canonical for scripted browser checks); user-invoked only.]

★ 0
phillippasqualato
phillippasqualato/PAPcSetup

web-design-guidelines

Review UI code for Web Interface Guidelines compliance. Use when asked to "review my UI", "check accessibility", "audit design", "review UX", or "check my site against best practices".

★ 0