The full arc from idea to shipped code.
- Claude Code: install as a plugin and run
/arc:*commands in Claude. - Codex: this repo includes
.agents/skillsso the sameskills/*/SKILL.mdworkflows can run directly in Codex (no Claude plugin install required).
Arc provides 22 skills covering the complete development lifecycle:
ENTRY /arc:go - Main entry point, routes to right workflow
↓
WHY /arc:vision - High-level goals (500-700 words)
↓
WHAT /arc:ideate - From idea to design doc
↓
DO /arc:implement - Plan + execute with TDD
/arc:design - UI/UX design with wireframes
/arc:build - Quick build (no formal plan)
/arc:testing - Test strategy and execution
/arc:letsgo - Production readiness checklist
/arc:legal - Generate privacy policy and terms
CROSS-CUTTING
/arc:review - Review a plan for feasibility
/arc:audit - Comprehensive codebase audit (includes hygiene)
/arc:document - Feature documentation
/arc:suggest - Opinionated next-step recommendations (+ discovery mode)
/arc:naming - Generate and validate project names
/arc:deps - Dependency management and updates
/arc:responsive - Mobile responsive audit & fix
/arc:seo - Deep SEO audit for web projects
/arc:tidy - Clean up completed plans
/arc:prune-agents - Kill orphaned subagent processes
TOOLS /arc:commit - Smart commit + push with auto-splitting
/arc:rules - Apply coding standards to project
- Review is woven throughout, not bolted on at the end — Each design section gets micro-reviewed before moving on
- Reviewers advise, the user decides — Suggestions are presented as questions, not mandates
- One question at a time — Never overwhelm with multiple questions
- YAGNI where appropriate — Simplifications suggested, but user has final say
- TDD mandatory — Tests first, implementation second
- Frontend-design integrated — Bold aesthetic direction, not generic AI slop
claude plugins install arc@howells
Codex discovers skills from ~/.agents/skills (and from .agents/skills inside repositories).
Recommended (install once, use anywhere):
Tell Codex:
Fetch and follow instructions from https://raw.githubusercontent.com/howells/arc/refs/heads/main/.codex/INSTALL.md
Follow .codex/INSTALL.md, then in any project you can invoke Arc skills directly.
If you open this repo itself in Codex, it includes .agents/skills/* symlinks so Codex can discover the skills without a global install.
Invoke skills explicitly (recommended):
- In CLI/IDE: run
/skillsor type$to pick a skill - In the Codex app: type
$<skill-name>in chat
$start
$ideate add user authentication with magic links
Codex loads the selected skill’s SKILL.md and follows its workflow.
- These skills are stored in
skills/<name>/SKILL.mdfor Claude Code;.agents/skills/<name>is a symlink to the same folder so Codex can discover them. - Some skills reference Claude-specific tooling (e.g.
TaskList,mcp__claude-in-chrome__*). In Codex, use the closest equivalent:- terminal exploration instead of
Taskblocks - Playwright (or user-provided screenshots) instead of Claude-in-Chrome MCP
- terminal exploration instead of
Arc uses these plugins for enhanced functionality:
| Plugin | Used by |
|---|---|
| Figma | /arc:ideate, /arc:implement, /arc:design |
| Context7 | /arc:implement |
| Chrome | figma-implement agent |
# Official plugins
/plugin install figma@claude-plugins-official
/plugin install context7@claude-plugins-official
# Chrome extension: https://chromewebstore.google.com/detail/claude-in-chrome/
Arc works without these, but relevant features will be limited.
Note: Arc maintains an activity log (.arc/log.md, gitignored) for knowledge persistence across sessions. Every skill auto-appends a brief entry on completion.
These plugins provide additional review capabilities:
| Plugin | Skill | Used by |
|---|---|---|
| agent-skills | vercel-react-best-practices |
/arc:implement, /arc:letsgo |
vercel-composition-patterns |
/arc:implement, /arc:design |
|
vercel-react-native-skills |
/arc:implement, /arc:letsgo, /arc:responsive |
|
| web-interface-guidelines | web-design-guidelines |
/arc:design, /arc:implement, /arc:responsive |
# Vercel Labs plugins (optional)
/plugin marketplace add vercel-labs/agent-skills
/plugin install agent-skills@vercel-labs-agent-skills
/plugin marketplace add vercel-labs/web-interface-guidelines
/plugin install web-interface-guidelines@vercel-labs-web-interface-guidelines
When installed, Arc commands will automatically use these skills for React/Next.js performance reviews and UI compliance checks.
cd your-project
claudeThis starts an interactive Claude Code session in your terminal.
Commands start with /. Type the command and press Enter:
/arc:ideate add user authentication with magic links
Claude will ask clarifying questions, explore your codebase, and create a design document.
Arc commands chain together. After /arc:ideate creates a design:
- Claude asks if you want to continue to
/arc:implement(plan and build) - Implementation creates its own plan, then executes with TDD
You can also jump in at any point if you already have docs.
- Open your project in Codex.
- Ensure
.agents/skillsis present (see "Using In Codex"). - Run skills in chat, e.g.
$startor$ideate ....
# Design a new feature (full flow)
/arc:ideate add a notification system
# Get suggestions for what to work on
/arc:suggest
# Ship to production
/arc:letsgo- One question at a time — Arc asks focused questions, not overwhelming lists
- You're in control — Suggestions are questions, not mandates. Say no if you disagree.
- TDD by default — Implementation writes tests first, then code
- Documents are created — Plans go in
docs/plans/, features indocs/features/
The main entry point is /arc:ideate, which flows through to implementation:
/arc:ideate → /arc:implement
Each step asks if you want to continue. You can also enter at any point:
- Have a design doc already? Start at
/arc:implement - Have an implementation plan?
/arc:implementwill use it
| Command | When to use | Output |
|---|---|---|
/arc:go |
Main entry point, routes to workflow | Context-aware guidance |
/arc:vision |
Starting a new project | docs/vision.md |
/arc:ideate |
From idea to design doc | docs/plans/YYYY-MM-DD-<feature>-design.md |
/arc:implement |
Plan + execute with TDD | Code changes |
/arc:design |
UI/UX work | Wireframes + code |
/arc:build |
Quick implementation | Code changes |
/arc:testing |
Test strategy | Test files |
/arc:letsgo |
Ship to production | Deployment |
/arc:legal |
Generate legal pages | Privacy policy, ToS, cookies |
/arc:review |
Review a plan for feasibility | Updated plan file |
/arc:audit |
Comprehensive codebase audit | docs/audits/YYYY-MM-DD-*.md |
/arc:document |
Document features | docs/features/<feature>.md |
/arc:suggest |
What to work on next (+ discovery mode) | Recommendations |
/arc:naming |
Generate project names | Name candidates |
/arc:commit |
Commit and push changes | Git commits |
/arc:rules |
Apply coding standards | .ruler/ directory |
/arc:tidy |
Clean up completed plans | Archived/deleted plans |
/arc:deps |
Dependency management | Updated dependencies |
/arc:responsive |
Mobile responsive audit & fix | Responsive code changes |
/arc:seo |
Deep SEO audit for web projects | docs/audits/YYYY-MM-DD-seo.md |
/arc:prune-agents |
Kill orphaned subagent processes | Clean process state |
Arc includes 22 specialized agents:
| Category | Agents |
|---|---|
| Research | docs-researcher, git-history-analyzer, duplicate-detector, naming, feature-scout |
| Review | architecture-engineer, simplicity-engineer, daniel-product-engineer, data-engineer, designer, lee-nextjs-engineer, hygiene-engineer, performance-engineer, security-engineer, senior-engineer, seo-engineer, accessibility-engineer, organization-engineer, test-quality-engineer |
| Build | implementer, fixer, debugger, ui-builder, figma-builder, design-specifier, unit-test-writer, integration-test-writer, e2e-test-writer, test-runner, e2e-runner, spec-reviewer, code-reviewer |
| Design | figma-implement |
| Workflow | spec-flow-analyzer, e2e-test-runner, docs-writer |
Implementation methodologies in disciplines/:
- test-driven-development — Red-green-refactor cycle
- systematic-debugging — Methodical bug investigation
- verification-before-completion — Prove it works before claiming done
- using-git-worktrees — Isolated development branches
- finishing-a-development-branch — Cleanup after work complete
- subagent-driven-development — Parallel agent execution
- dispatching-parallel-agents — Efficient multi-agent coordination
- receiving-code-review — Handling review feedback
Commands work together:
/arc:suggestreads Linear issues (if configured), in-session tasks (TaskList), codebase, and/arc:vision(priority cascade with opt-in discovery mode)/arc:ideateflows to/arc:implement(which creates plans internally)/arc:buildsuggests/arc:ideateif scope is too large/arc:letsgoruns/arc:testingand/arc:audit --hygieneas part of quality checks- Claude Code uses TaskList for in-session task tracking; Linear MCP for persistent issue tracking
For complex projects, Arc integrates with Linear via MCP for issue tracking:
// .mcp.json
{
"mcpServers": {
"linear": {
"command": "npx",
"args": ["-y", "@anthropic/linear-mcp"]
}
}
}When Linear MCP is available, /arc:suggest queries active issues and /arc:audit can create issues from findings.
Arc builds on patterns and disciplines from:
- superpowers — Implementation disciplines (TDD, debugging, verification)
- compound-engineering — Agent patterns and workflows
MIT