Accio

Accio

Summoning Charm. You do not remember 64 skills. Name the problem, not the skill.

Fired bare, with nothing attached, ask what is being worked on. Reprinting this map back is a no-op: the reader is already looking at it. Name the one or two that fit and say what to type.

Two markers below:

• (type it) the skill is disable-model-invocation: true. Nothing reaches it but the user typing its name, so naming it is the whole deliverable.
• (plugin) it comes from an enabled plugin, not this collection.

Every route on this page was verified against disk. If you add a skill, add it here; if you remove one, remove it here. A router that names something that does not exist is worse than no router, because it is confidently wrong. scripts/check-routes.py measures whether that has rotted.

Arriving somewhere new

• /onboard-repo: a repo Claude has no context on. Recon, mutual interrogation, a feedback loop it actually executes, a capped CLAUDE.md, a project wiki, and an interview-gated .claude/ harness. Runs once per repo.
• /onboard-light: the same flow with the harness phase pre-answered "skip". For a repo whose .claude/ is not yours to write: a client repo, a restricted one.
• /read-wiki: before touching anything in a repo that has a wiki. Strictly read-only.
• /caveman-explore: cold-start orientation when you need file:line answers and want the reading kept out of the main context.

Idea to shipped

The route most work travels.

1. /grill-with-docs (type it) sharpens the idea by interview and leaves a paper trail in CONTEXT.md and ADRs. Use it whenever there is a working directory to leave that trail in. /grill-me (type it) is the same interview with no repo under it. Both run /grilling, which you reach directly only when you want the interview with no wrapper.
   • Lighter alternative: /spec-interview (type it) when the shape is roughly known and you just want it pinned down fast, one question at a time. It hands off to writing-plans (plugin, superpowers), not to /to-spec or /to-tickets.
   • brainstorming (plugin, superpowers) covers the same ground from the other direction and fires on its own. If it has already run, do not re-interview from scratch.
   • /domain-modeling: fires on its own when codebase terminology comes up. Writes and maintains CONTEXT.md and the ADRs the interview leaves behind.
2. Question needs a runnable answer? Detour through /prototype: throwaway code that settles one design question (does this state model feel right, what should this UI look like). The prototype is kept as a primary source, not destroyed.
3. Multi-session build?
   • Yes → /to-spec (type it), then /to-tickets (type it) for tracer-bullet tickets with blocking edges, then /implement (type it) per ticket, /clearing between each.
   • No → /implement right here.

Keep steps 1 to 3 in one unbroken context window. Do not compact or clear until after /to-tickets, so grilling, spec and tickets all build on the same thinking. Each /implement then starts fresh from its ticket.

/setup-matt-pocock-skills (type it) is the precondition: run once per repo before the first flow, to configure the issue tracker and label vocabulary the ticket skills assume.

Too big to hold in one session

/wayfinder (type it). When the way to the destination is not visible yet, it charts a map of decision tickets and resolves them one at a time, producing decisions rather than deliverables. When the fog clears it hands off to /to-spec; it does not build. Save it for genuinely foggy work, never a well-scoped feature.

Something is broken

• systematic-debugging (plugin, superpowers): the first reach. Refuses to theorise before it has a tight feedback loop that goes red on this bug.
• /surgical-patch: once the cause is known, fix at the narrowest responsible layer with a regression proof and the surrounding behaviour preserved.
• /resolving-merge-conflicts: an in-progress merge or rebase, resolved by intent hunk by hunk. Never runs --abort.

Bugs and requests piling up

/triage (type it), for issues you did not create: bug reports, incoming requests, raw arrivals. Tickets that /to-tickets produced are already agent-ready, so do not triage them.

Changing code safely

• /safe-refactor: restructure while preserving behaviour, verification bracketing every structural edit.
• /migration: reversible, compatibility-safe transitions (schema, data, API, config, dependency).
• /ponytail: forces the laziest solution that works. Reach for it on any coding task where over-engineering is the risk. Levels: lite, full, ultra.
• /modern-python: before or while editing Python. Lists the modernization rules (Ruff's UP, FURB and friends) the project's target Python version allows, reads them, applies them, and at the verified level checks the edit with Ruff and applies the safe fixes. Fires on its own on any Python edit; /modern-python setup runs the per-project interview (greenfield or existing, enforcement level, target version, Ruff source, profile) that writes .claude/modern-python.md.

Reviewing

• /ponytail-review (type it): the current diff, over-engineering only. What to delete.
• /ponytail-audit (type it): the same lens over the whole repo.
• /smells (type it): the twelve Fowler smells with concrete fixes.
• /code-review (plugin) and requesting-code-review / receiving-code-review (plugin, superpowers) for general review of a branch or PR.
• /codebase-design: not a review, the shared vocabulary (module, interface, depth, seam, adapter) for designing a module's shape. /improve-codebase-architecture (type it) is the survey that finds candidates; this is the bench you design the chosen one on.

Deciding

• /mental-models: pressure-test a real decision through named frameworks, ending in a recommendation.
• /the-llm-council: five advisors attack it from different angles, peer-review anonymously, a chairman delivers a verdict. Advisory.
• /scope-creep-check: fires proactively when a request has quietly grown. Names the drift, shows the cost, asks proceed / park / drop.
• /goal-creator: turn rough intent into a /goal-ready session goal with a verifiable outcome and a scope guard.

Harness health

The tools that measure the setup rather than the code.

• /context-budget: what is eating the context window, ranked by tokens reclaimable.
• /config-gc: garbage-collect ~/.claude (stale, orphaned, redundant), confirm each deletion.
• /skill-comply: measures whether a skill or rule is actually obeyed, by generating scenarios at three prompt-strictness levels and classifying the tool traces. Costs real quota; start with --dry-run.
• /primitive-classifier: audits a folder and says which Claude Code primitive each piece of context actually belongs in.
• /loop-design-check: before building an agent loop, and for reviewing one you suspect will spin, cheat, or run a wrong answer to completion.
• /gateguard: the design of a three-stage fact-forcing PreToolUse gate (deny, force facts, allow). A spec to build a hook from, not a runnable hook.

Knowledge

• /read-wiki: query before doing work. Read-only.
• /write-wiki: ingest a source, update pages, lint, or bootstrap a new wiki. /wiki (type it) is a backward-compatible alias.
• /remember (plugin): session handoffs and durable memory across sessions.
• /hindsight: at the end of a session, review how the work went, keep the lessons the session proved apart from the ones it only suggested, and write the durable ones into the project memory folder, updating an existing entry rather than adding a twin. Over /remember when the question is what to do differently next time, not where the work left off.

Token cost

• /caveman: terse output mode (lite / full / ultra). Cuts the expensive side of the meter.
• /cavecrew: when to delegate to the compressed subagent presets instead of a vanilla explore, so subagent results do not land in main context verbatim.
• /caveman-compress: compress a stored CLAUDE.md or memory file, with a readable backup.
• /caveman-help: the quick-reference card for the above.

Prompting

• /prompt-fixer: cleans a rough prompt into plain English and hands it back to paste and edit. No XML, does not execute.
• /prompt-optimizer: wraps the ask in XML, tuned to the target Claude model via a per-model reference (Opus 5, Sonnet 5, Fable 5.1), validates with you, then runs it. Its model references also say what to remove from an inherited prompt, which matters more than what to add.

Teaching, publishing, checking

• /explain-yogi-like-he-is-5: a Socratic session that teaches rather than summarizes.
• /watch: a YouTube transcript turned into notes, chapters, quotes, or answers.
• /field-guide-builder (type it): one deep, source-grounded, single-file HTML teaching page.
• /conference-talk-deck (type it): interviews you, then generates an offline HTML deck.
• /sdd-workshop-walkthrough (type it): an animated single-file workshop player.
• /qa-deck: QA a .pptx for visual defects and fabricated claims before it ships.
• /excalidraw: architecture diagrams as .excalidraw files from codebase analysis. PNG/SVG export needs the Playwright MCP server connected.
• /archify: system diagrams (architecture, workflow, sequence, data-flow, lifecycle) as one self-contained interactive HTML file, authored as typed JSON and validated by its Node CLI. Over /excalidraw when the reader will open it in a browser rather than edit it.
• /geo-content: content written for both classic SEO and AI answer engines.
• /no-ai-slop: edit a draft into sharper, more human writing without losing the writer's voice, or name the AI-slop patterns in it without rewriting. Also the honest answer to "does this read as AI": named patterns, not a detector score.
• /anthropic-doc-validator: check claims about Claude Code, the API or the SDKs against the official docs, with verbatim quotes. Reach for it before asserting anything version-specific.
• /research: delegate reading legwork to a background agent; it returns a cited Markdown file. Feeds the thinking, does not replace it.

Building a new skill

1. /skill-scout first: search local, marketplace, GitHub and web before writing anything. The collection has absorbed several upstream repos; the odds of duplicating are real.
2. /writing-great-skills (type it) for the vocabulary and principles. writing-skills (plugin, superpowers) covers similar ground.
3. /skill-comply afterwards, to find out whether the thing you wrote is actually followed.

Other people, and messages that missed

• /to-questionnaire (type it): the blocker is in someone else's head. Interviews you about the send, then writes them a questionnaire.
• /wizard: steps only a human can take (provisioning, credentials, a third-party dashboard). Generates an interactive bash script so the procedure stops being re-explained every time.
• /wait-what (type it): that last message did not land. Re-pitches it in Simplified Technical English using the repo's CONTEXT.md vocabulary.
• /teach (type it): learn a concept over multiple sessions, using the current directory as a stateful workspace.

Phase boundaries

At the boundary between two chunks of work you have five options, and choosing between them is the fuzziest decision here: continue, /clear, hand off via /remember (plugin), subagent, or /compact. Read PHASE-BOUNDARIES.md for the ordered tree and why continue is the one to rule out first.

Keeping this true

This page is the only index for the (type it) skills. They carry no description into the model's context, so one missing from here is unreachable by anything but someone remembering its name. That is not hypothetical: the toolbox skill this page replaced held the same job and silently drifted to missing 11 of the 24 it indexed, because nothing checked it.

So there is a check. Run it after adding, renaming or removing any skill:

python accio/scripts/check-routes.py

It fails on a dead route (this page names a skill that does not exist) and on an unreachable skill (a user-only skill this page omits). Fix SKILL.md, never the script.