Skill autopilot

<SUBAGENT-STOP> If you were dispatched as a subagent to execute a specific task, skip this skill. </SUBAGENT-STOP>

Skill Autopilot

The Iron Law

EVERY user message gets skill-matched BEFORE you respond, write code, or ask questions.

This is not a suggestion. If a skill exists that could help — even 1% chance — invoke it.

Decision Flowchart

digraph autopilot {
  "Message received" [shape=doublecircle];
  "Classify intent" [shape=box];
  "Match skills" [shape=box];
  "Duplicates?" [shape=diamond];
  "Resolve precedence" [shape=box];
  "Multi-skill chain needed?" [shape=diamond];
  "Build chain (process→domain→quality)" [shape=box];
  "Any match?" [shape=diamond];
  "Search online" [shape=box];
  "Invoke skill(s)" [shape=box];
  "Follow skill instructions" [shape=doublecircle];

  "Message received" -> "Classify intent";
  "Classify intent" -> "Match skills";
  "Match skills" -> "Duplicates?";
  "Duplicates?" -> "Resolve precedence" [label="yes"];
  "Duplicates?" -> "Multi-skill chain needed?" [label="no"];
  "Resolve precedence" -> "Multi-skill chain needed?";
  "Multi-skill chain needed?" -> "Build chain (process→domain→quality)" [label="yes"];
  "Multi-skill chain needed?" -> "Any match?" [label="no"];
  "Build chain (process→domain→quality)" -> "Invoke skill(s)";
  "Any match?" -> "Invoke skill(s)" [label="yes"];
  "Any match?" -> "Search online" [label="no"];
  "Search online" -> "Invoke skill(s)";
  "Invoke skill(s)" -> "Follow skill instructions";
}

Phase 1: Classify Intent

Read the message. What is the user doing? Match against these intent signals:

Signal	Intent
build, create, design, page, component, UI, layout, style	Create UI
fix, bug, error, broken, crash, not working, fails	Debug
plan, architect, complex, multi-step, "how should I"	Plan
review, audit, polish, check, accessibility, a11y	QA/Review
animate, motion, transition, hover, micro-interaction	Animation
optimize, slow, performance, bundle, render, memo	Performance
refactor, simplify, clean up, best practices	Code Quality
test, playwright, e2e, verify, screenshot	Testing
commit, push, PR, deploy, merge, ship	Ship
pdf, docx, pptx, xlsx, word, excel, slides, spreadsheet	Documents
image, logo, banner, art, poster, generate	Media
SEO, keywords, copy, headline, marketing, convert	Marketing
Claude API, Anthropic SDK, MCP, agent SDK	AI/MCP
twitter, reddit, product hunt, web search, scrape	Research
schema, migration, postgres, neon, SQL, database	Database
n8n, workflow, webhook, automation	Automation
SwiftUI, iOS, React Native, Flutter, mobile	Mobile
figma, design handoff, implement design	Figma
find skill, install skill, "is there a skill for"	Discovery
sentry, logging, tracing, metrics, monitoring	Monitoring
remotion, video	Video
domain, registrar, .com, .ai	Domains
shadcn, component registry	Component Lib

Multiple intents can co-exist. "Build a landing page and deploy it" = Create UI + Ship.

Phase 2: Resolve Duplicates

Many skills overlap. Always use the winner. Full precedence table: references/precedence.md

Quick rules:

• Design skills → impeccable:* wins over standalone and document-skills:*
• Process skills → superpowers:* wins over standalone
• Document skills → standalone wins over document-skills:* and example-skills:*
• React/Next → react-best-practices (standalone) = vercel-react-best-practices (pick one, not both)
• Never invoke both document-skills:X and example-skills:X — they're identical

Phase 3: Build Skill Chain

Skills have a strict invocation order. Process → Domain → Quality.

Phase	Skills	When
1. Process	superpowers:brainstorming, superpowers:systematic-debugging, superpowers:writing-plans	New features, bugs, complex tasks
2. Domain	impeccable:frontend-design, design-system, react-best-practices, pdf, pptx, etc.	The actual work
3. Quality	impeccable:audit, impeccable:polish, superpowers:verification-before-completion	After implementation
4. Ship	superpowers:requesting-code-review, commit-commands:commit-push-pr, vercel:deploy	Wrapping up

Common chains — see references/chains.md for the full recipe book.

Critical gate: Process skills come FIRST. Do NOT invoke impeccable:frontend-design without superpowers:brainstorming first for new features. Do NOT invoke code fixes without superpowers:systematic-debugging first for bugs.

Phase 4: No Match → Discover

If nothing in your installed inventory matches:

1. Run npx skills find "<query>" with 2-3 relevant keywords
2. Present results with install command: npx skills add <pkg> -g -y
3. If npx skills isn't available, try: gh search repos "claude-code skill <topic>"
4. If nothing found online, proceed with general capabilities

See references/discovery.md for the full workflow.

Phase 5: Invoke

1. Announce: "Using [skill] for [purpose]"
2. Invoke: Call Skill tool with the winning skill name
3. Follow: Execute the skill's instructions exactly — skills are authoritative
4. Chain: If the skill's output names another skill, invoke that too

Red Flags — You're Bypassing Autopilot

Your Thought	Reality
"Too simple for a skill"	Simple tasks have the most unexamined assumptions. Check.
"I'll check skills after"	Skills are checked BEFORE. Always.
"Let me just write the code"	Process skills (brainstorming, debugging) come before code.
"I already know this"	Skills have checklists you'll miss. Use them.
"No skill matches exactly"	Partial match > no match. Also search online.
"The user didn't ask for a skill"	Auto-invoked. Users never need to ask.
"I'll handle design myself"	impeccable skills exist. Use them.
"It's just a quick fix"	systematic-debugging exists for a reason.
"I know what that skill says"	Skills evolve. Invoke the current version.

Session Lifecycle

Moment	Action
Start	Is this a new feature? → superpowers:brainstorming. Is there a plan? → superpowers:executing-plans.
During	Match domain skills to each sub-task as it arises.
Before claiming done	superpowers:verification-before-completion — ALWAYS. No exceptions.
Before commit/PR	superpowers:requesting-code-review → commit-commands:commit-push-pr

Always-On Context Skills

These fire automatically when their context is detected, regardless of intent:

• design-system — any project with design tokens + any visual change
• react-best-practices or next-best-practices — any React/Next.js code touch