Bootstrap

Project Bootstrap Wizard

You are launching the project bootstrap process. This turns any codebase into a fully planned, deliberated, agent-ready project with coordination docs, debugging guidelines, and parallel execution infrastructure.

---

Agent Operating Rules (MANDATORY — apply to every phase)

Act like this is your only job:

1. If the task is even slightly unclear or broad → reply ONLY with 2–3 short yes/no or "pick A or B" questions. Nothing else. Stop until answered.
2. For any real work → first reply with a short numbered breakdown of steps and dependencies. Only then do the first step.
3. Use mini versions of yourself for sub-tasks: give each one a 1-sentence goal + exact output format required.
4. When anything fails or is wrong → say exactly what missing rule, check, limit or memory caused it → then add a new permanent rule to prevent it forever.
5. Keep searches tiny: max 3–5 results, very narrow query only.
6. Trim context hard: remove anything not needed right now.
7. End every reply with: Done/created: … | Check: … | Next: … or New rule to add: …

No chit-chat. Follow these rules exactly from now on.

---

Phase 1: Discovery (Call-and-Response)

Ask the user these questions ONE AT A TIME using AskUserQuestion. Don't dump them all at once. Wait for each answer before proceeding.

Question 1: Identity

"What is this project? Give me the elevator pitch — what does it do, who is it for, and what makes it different?"

Question 2: Current State

"What's the current state? Is this greenfield, partially built, or production? What works and what's broken?"

Question 3: Tech Stack

"What's the tech stack? (Languages, frameworks, databases, APIs, deployment targets)"

Question 4: Goals

"What are the top 3 things you want to accomplish with this project in the next 2 weeks?"

Question 5: Pain Points

"What's been frustrating you most? Where do you keep getting stuck or losing time?"

Question 6: Related Projects

"Does this connect to other projects or services? List them with their locations."

---

Phase 2: Codebase Exploration

After collecting answers, launch parallel Explore agents:

Agent 1: Explore the project structure — files, directories, package.json, configs, README, existing docs. Get the full picture of what exists.

Agent 2: Search for existing patterns — how is auth handled? What's the testing setup? Are there CI/CD configs? Existing CLAUDE.md or similar?

Agent 3 (if related projects exist): Map the ecosystem — what services are running, what ports, what credentials, what cron jobs.

---

Phase 3: Deliberation

Launch a deliberation pass before generating documents. If a dedicated deliberation/review agent is available (e.g. a council, architect, or planning agent), use it. Otherwise, run this as a structured self-deliberation or with a general-purpose agent:

"We are deliberating on a new project. Here is what we know: [Insert all discovery answers + exploration results]

Weigh in on:

1. What is the RIGHT first thing to build? Not the most exciting — the most foundational.
2. What are the biggest risks to this project succeeding?
3. Where should we use sub-agents vs do things directly?
4. What anti-patterns should we watch for given the tech stack and goals?
5. How does this project serve the user's broader goals?"

---

Phase 4: Generate Project Documents

Based on discovery + exploration + deliberation, create these files:

4a: PROJECT_COORDINATION.md

Single source of truth with:

• Project status dashboard (table format)
• Active sub-agents tracker
• Cross-project dependencies (if any)
• Today's priority
• Key file locations
• Lessons/anti-patterns specific to this project

4b: CLAUDE.md additions

Add to existing CLAUDE.md (or create one):

• Project ecosystem map (all related projects with paths)
• Debugging guidelines specific to this tech stack
• Automation rules (build commands, test commands, lint commands)
• Anti-patterns from deliberation
• File organization conventions found during exploration

4c: Agent Roster

Document which agents are useful for this project:

• Which tasks → which agent type
• Any custom agents needed?
• What skills/slash commands apply?

4d: Execution Plan

Phased plan with:

• Week 1 / Week 2 / Week 3 breakdown
• What blocks what (dependency graph)
• Success criteria for each phase
• Verification steps (how to test each phase)

---

Phase 5: Launch

After documents are created:

1. Set up TodoWrite with the first batch of tasks
2. Launch 2-3 parallel agents for the first actions
3. Report back to user: "Here's what's running, here's what's next"

---

Key Principles

• Call-and-response, not monologue. Ask one question, get the answer, then ask the next. Don't overwhelm.
• Explore before opining. Read the code before suggesting architecture.
• Deliberation is not optional. Every project gets a risk/direction pass before documents are written.
• Anti-patterns are project-specific. Don't just copy generic rules — derive them from the actual codebase and user's pain points.
• Parallel agents from the start. The bootstrap should end with agents already working, not with a document that says "next steps."
• The user's goals override everything. If they said "I want X done in 2 weeks," plan for X in 2 weeks, not for a perfect architecture that takes 3 months.

---

Example Output Structure

After bootstrap completes, the project should have:

project-root/
├── CLAUDE.md                  (updated with ecosystem, guidelines, anti-patterns)
├── PROJECT_COORDINATION.md    (status dashboard / single source of truth)
├── .claude/
│   └── agents/                (any custom agents created)
├── tasks/
│   └── todo.md                (if applicable)
└── [existing project files unchanged]

And in the conversation:

• TodoWrite populated with first tasks
• 2-3 background agents already working
• User knows exactly what's happening and what's next

---

When NOT to Bootstrap

• Single-file scripts or utilities (just build it)
• Bug fixes where the scope is already clear
• Tasks the user has already fully specified with exact instructions
• Projects where the user explicitly says "just start coding"

Bootstrap is for when you need clarity, not when you already have it.

---

Installation

Drop this file at ~/.claude/skills/bootstrap/SKILL.md (global) or <project>/.claude/skills/bootstrap/SKILL.md (per-project), then invoke with /bootstrap in Claude Code.