| name | bootstrap |
| description | Generate feature plan with codebase-aware context. Use when starting a new feature in an existing codebase to skip full spec workflow. |
| allowed-tools | Read, Write, Edit, Glob, Grep, Bash, AskUserQuestion |
Bootstrap
Generate a feature plan with codebase-aware context. Scans your codebase, understands existing patterns, and creates a targeted feature specification and execution plan.
When to Use
- Starting a new feature in an existing codebase
- You have a clear idea of what to build
- You want to skip the full product/technical spec workflow
- You ran
/adopt from the orchestrator and saved context
Workflow
Copy this checklist and track progress:
Bootstrap Progress:
- [ ] Step 1: Gather initial description
- [ ] Step 2: Orient to project root
- [ ] Step 3: Run /setup automatically
- [ ] Step 4: Codebase scan
- [ ] Step 5: Check for unfinished plans
- [ ] Step 6: Ask clarifying questions (if needed)
- [ ] Step 7: Derive and confirm feature name
- [ ] Step 8: Create feature directory and documents
- [ ] Step 9: Update CLAUDE.md
- [ ] Step 10: Clean up and report
Step 1: Gather Initial Description
Check for existing context:
cat .claude/bootstrap-context.md 2>/dev/null
If exists, offer to use it. Otherwise, ask:
What would you like to build?
Store as USER_DESCRIPTION.
Step 2: Orient to Project Root
- Check for git repository:
git rev-parse --show-toplevel
- If not git, search upward for AGENTS.md
- Store result as
PROJECT_ROOT
Step 3: Run /setup Automatically
Ensure the toolkit is installed in the project:
-
Determine toolkit location:
- If running from the toolkit repo (e.g.,
GENERATOR_PROMPT.md exists in CWD), use that path
- Otherwise resolve from
.claude/toolkit-version.json if present
- If still unknown, try common paths (e.g.,
~/Projects/ai_coding_project_base)
-
Run /setup PROJECT_ROOT (idempotent — safe to run multiple times)
- If toolkit already installed and current, setup will report "Already up to date"
- If toolkit needs updating, setup will perform incremental sync
- If toolkit not installed, setup will perform full installation
-
Continue after setup completes (no user interaction needed for idempotent runs)
Step 4: Codebase Scan
See CODEBASE_SCAN.md for detailed scanning procedures.
Compile findings as CODEBASE_CONTEXT with:
- Language/Framework detected
- Patterns to follow
- Related existing code
- Testing approach
Step 5: Check for Unfinished Execution Plans
find PROJECT_ROOT -name "EXECUTION_PLAN.md" -type f
For each found plan:
- If
.claude/phase-state.json exists, use it for completion status
- If no state file, parse the plan for completion:
- Count
- [x] (completed) vs - [ ] (incomplete) criteria
- A plan is incomplete if any criteria are unchecked
If unfinished plans exist, report:
UNFINISHED EXECUTION PLANS DETECTED
===================================
- EXECUTION_PLAN.md: Phase 2 of 4 (5/12 tasks complete)
- features/analytics/EXECUTION_PLAN.md: Phase 1 of 2 (3/8 tasks complete)
Ask with AskUserQuestion:
Question: "There are unfinished execution plans. How should I proceed?"
Options:
- "Continue anyway" — Create new feature plan alongside existing work
- "Add TODO reminder" — Add reminder to TODOS.md and continue
- "Cancel" — Stop bootstrap to focus on existing work
If "Add TODO reminder":
echo "- [ ] [backlog] Complete unfinished execution plans before starting new work" >> PROJECT_ROOT/TODOS.md
If "Cancel": Stop and report existing plans that need attention.
Step 6: Ask Clarifying Questions
Only ask what's necessary. Do NOT ask about:
- Things implied by USER_DESCRIPTION
- Implementation details you can decide from patterns
- Nice-to-haves for later
Step 7: Derive and Confirm Feature Name
- Extract key concepts from description
- Create kebab-case name (2-4 words)
- Confirm with user
Store as FEATURE_NAME.
Step 8: Create Feature Directory and Documents
mkdir -p PROJECT_ROOT/features/FEATURE_NAME
Create:
FEATURE_SPEC.md — Problem, solution, scope, acceptance criteria
EXECUTION_PLAN.md — Phases, tasks, files to modify
See TEMPLATES.md for document templates.
Step 9: Update CLAUDE.md
If CLAUDE.md doesn't reference AGENTS.md, add:
@AGENTS.md
Step 9b: Cross-Model Review (Automatic)
After creating documents, run cross-model review if Codex CLI is available:
- Check if Codex CLI is installed:
codex --version
- If available, run
/codex-review on the generated documents
- Present any findings to the user before proceeding
Review invocation:
/codex-review --mode research --research "{detected technologies from codebase scan}" PROJECT_ROOT/features/FEATURE_NAME/EXECUTION_PLAN.md
If Codex finds issues:
- Show critical issues and recommendations
- Ask user: "Address findings before proceeding?" (Yes/No)
- If Yes: Apply suggested fixes
- If No: Continue with noted issues
If Codex unavailable: Skip silently and proceed to Step 10.
Step 10: Clean Up and Report
If bootstrap-context.md was used, offer to delete it.
FEATURE PLAN CREATED
====================
Feature: {FEATURE_NAME}
Location: PROJECT_ROOT/features/FEATURE_NAME/
Files created:
- FEATURE_SPEC.md
- EXECUTION_PLAN.md — {N} phases, {M} tasks
Cross-Model Review: PASSED | PASSED WITH NOTES | SKIPPED
Next steps:
1. Review: cat PROJECT_ROOT/features/FEATURE_NAME/FEATURE_SPEC.md
2. Start: cd PROJECT_ROOT/features/FEATURE_NAME && /fresh-start && /phase-start 1
Example
Input: "Add dark mode toggle to settings page with localStorage persistence"
Codebase scan finds: React app, Settings.tsx, existing Toggle component, CSS modules
Output: Feature directory at features/dark-mode/ with FEATURE_SPEC.md and EXECUTION_PLAN.md referencing existing patterns and files.