Project Creator is a Claude Code meta-project that creates other Claude Code projects through reverse prompting. The key insight is that **requirements and design knowledge live in the user's head** — the LLM's job is to draw it out systematically, not generate it.
Installs into .claude/skills of the current project.
Are you the author of 2677 Project Setup Plan Dc91e61a?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/tools-only-2677-project-setup-plan-dc91e61a)
# Project Setup Plan: Project Creator
## Executive Summary
Project Creator is a Claude Code meta-project that creates other Claude Code projects through reverse prompting. The key insight is that **requirements and design knowledge live in the user's head** — the LLM's job is to draw it out systematically, not generate it.
**Core Function:** Transform ambiguous project ideas into well-configured Claude Code projects with appropriate CLAUDE.md, README.md, commands, and initial structure.
**Usage Pattern:**
- Point Cowork task or Claude Code session at Project Creator directory
- Sub-projects created in git-ignored subdirectories, organized by client
- Shared directory means both Cowork and Claude Code see full context
- Sub-project eventually graduates to independent Claude Code project
---
## Phase 1: Understanding the Domain
### 1.1 The Three Phases of Reverse Prompting
From `llm-workflow-methodology.md`:
| Phase | Focus | What Happens |
|-------|-------|--------------|
| **Seeding** | Capture | LLM asks questions to draw out raw material — requirements, constraints, context |
| **Cultivation** | Develop | LLM provides prompts to develop and connect fragments, resolve ambiguity |
| **Shaping** | Generate | LLM structures answers into artifacts (CLAUDE.md, README, commands, etc.) |
**This ticket covers Seeding phase commands only.** Cultivation and Shaping commands will come later.
### 1.2 The Parent-Child Pattern
From `llm-workflow-methodology.md` (Meta-Project Patterns):
```
project-creator/
├── CLAUDE.md # Parent configuration
├── .gitignore # Contains: projects/
└── projects/ # Git-ignored; sub-projects live here
├── [client-a]/
│ ├── [project-1]/ # Independent git repo
│ └── [project-2]/
└── [client-b]/
└── [project-3]/
```
**Key principles:**
1. **Git isolation** — `projects/` folder is git-ignored; sub-projects have independent repos
2. **Client grouping** — Projects organized by client
3. **CLAUDE.md at root** — Parent ignores child CLAUDE.md files when working at parent level
4. **Context separation** — When working IN a sub-project, only sub-project context applies
### 1.3 Reference Implementation Patterns
From `case-study-writing-companion.md`:
| Writing Companion | Project Creator Equivalent |
|-------------------|---------------------------|
| `/capture` — guided conversation to extract life material | `/intake` — guided conversation to extract project requirements |
| `/seed` — generate prompts from captured material | (Cultivation phase — future) |
| `/harvest` — end-of-session capture | `/checkpoint` — session state capture |
| Anchor files with "quality extraction" | Project context files with requirement extraction |
The Writing Companion evolved commands through use (v1 → v4). Project Creator should expect similar evolution.
---
## Phase 2: Proposed Architecture
### 2.1 Directory Structure
```
project-creator/
├── CLAUDE.md # Project Creator configuration
├── README.md # Workflow guide for humans
├── methodology.md # Copy of llm-workflow-methodology.md
├── tracking/
│ ├── current-project.md # Format: client/project-name
│ ├── projects-log.md # What projects have been created, their status
│ └── patterns-discovered.md # Learnings that should become templates
├── templates/ # Project archetypes (emerges over time)
├── .gitignore # Contains: projects/
├── .claude/
│ └── commands/
│ ├── project.md # Set/show/create current project context
│ ├── intake.md # Seeding: new project via reverse prompting
│ ├── onboard.md # Seeding: existing project analysis + gap filling
│ ├── process.md # Seeding: analyze external inputs
│ ├── gaps.md # Seeding: assess what's captured vs. needed
│ └── checkpoint.md # Seeding: session capture
└── projects/ # Git-ignored; organized by client
└── [client]/
└── [project]/ # Independent git repo
```
### 2.2 Commands (Seeding Phase)
| Command | Purpose | Modeled After |
|---------|---------|---------------|
| `/project` | Set/show/create current project context | Git branch pattern |
| `/intake` | New project via reverse prompting | Writing Companion `/capture` |
| `/onboard` | Analyze existing project + fill gaps | Methodology "onboard vs intake" pattern |
| `/process` | Analyze external inputs (transcripts, docs) | Writing Companion input processing |
| `/gaps` | Assess what's captured vs. what's needed | Quality check before proceeding |
| `/checkpoint` | Session state capture | Writing Companion `/harvest` |
**All commands:** Accept optional `[client/project]` parameter; use current if not specified.
### 2.3 Skills
**None for v1.** The reverse prompting methodology lives in CLAUDE.md and methodology.md. Skills can emerge through use if we find reusable patterns that benefit from skill encapsulation.
### 2.4 Sub-agents
**None for v1.** Reverse prompting is inherently conversational and sequential. Sub-agents would be useful for parallelizable tasks, but that's not the nature of this work.
---
## Phase 3: Key Files to Create
### Configuration Files
1. **CLAUDE.md** — Master configuration
- Define the role (collaborator, reverse prompting heavy, critical and creative)
- Describe the parent-child pattern
- Define the three phases (seeding focus for now)
- Reference methodology.md for deep patterns
2. **README.md** — Human workflow guide
- Quick start: how to begin a session
- Command reference with examples
- Directory structure explanation
- When to use `/intake` vs `/onboard`
3. **methodology.md** — Copy of `llm-workflow-methodology.md`
- Full reference for reverse prompting patterns
- Failure recovery patterns
- Meta-project patterns
### Tracking Files
4. **tracking/current-project.md** — Simple state file
- Contains: `client/project-name`
- Updated by `/project` command
5. **tracking/projects-log.md** — Project registry
- What projects have been created
- Current status (seeding, cultivation, shaping, graduated)
- Key decisions made
6. **tracking/patterns-discovered.md** — Evolution log
- Learnings from using the system
- Patterns that should become templates
- Command improvements identified
### Commands
7. **`.claude/commands/project.md`**
```
/project # Show current + list all
/project [client/project] # Set current
/project new [client/project] # Create and set as current
```
8. **`.claude/commands/intake.md`** — Reverse prompting for new projects
- Ask questions one at a time
- Draw out: purpose, users, constraints, success criteria
- Create initial project structure
- Populate context files with captured requirements
9. **`.claude/commands/onboard.md`** — Analyze existing projects
- User must `git clone` existing project first
- Analyze what exists against methodology
- Produce report: FOUND / MISSING / RECOMMENDATIONS
- Target gaps specifically via reverse prompting
10. **`.claude/commands/process.md`** — Handle external inputs
- Accept transcripts, documents, notes
- Extract structured requirements
- Update project context files
11. **`.claude/commands/gaps.md`** — Assessment
- Review captured context against methodology checklist
- Identify what's missing or unclear
- Prioritize what to capture next
12. **`.claude/commands/checkpoint.md`** — Session capture
- Summarize what was captured this session
- Update tracking files
- Identify next steps
- Prepare for potential context loss
### Infrastructure
13. **.gitignore** — Must include `projects/`
14. **templates/** — Empty directory (emerges over time)
---
## Phase 4: Design Decisions
### Resolved
1. **Project reference format:** `client/project-name`
- Matches directory structure
- Clear hierarchy
- Example: `acme-corp/sample-project`
2. **Current project storage:** Markdown file (`tracking/current-project.md`)
- Consistency with other tracking files
- Can include additional context if needed later
3. **Methodology reference:** Copy, not symlink
- Self-contained project
- Can annotate/extend for project-creator-specific notes
4. **Command parameter pattern:** Optional `[client/project]` override
- Use current if not specified
- Explicit override when needed
- Mirrors git's path behavior
### To Discover Through Use
5. **Sub-project initialization:** What files should `/project new` create?
- Minimal: just directory + .git
- Or: scaffold with empty CLAUDE.md, README.md?
- Will discover through first uses
6. **Template emergence:** When do patterns become templates?
- Track in `patterns-discovered.md`
- Formalize when pattern appears 3+ times
7. **Graduation criteria:** When is a sub-project ready to leave?
- Has CLAUDE.md, README.md, commands
- Has been validated through use
- Will define more precisely through experience
---
## Phase 5: How a Session Works
### Starting a New Project
```
# Point Claude Code at project-creator directory
cd project-creator
# Create new project context
/project new acme-corp/sample-project
# Start reverse prompting
/intake
# (Claude asks questions one at a time, captures answers)
# Process any external inputs
/process [paste transcript or path to doc]
# Check what's still needed
/gaps
# End session
/checkpoint
```
### Onboarding an Existing Project
```
# Clone existing project into projects/
git clone <repo> projects/acme-corp/existing-api
# Set as current
/project acme-corp/existing-api
# Analyze and fill gaps
/onboard
# (Claude analyzes, reports findings, asks to fill gaps)
# End session
/checkpoint
```
### Continuing Work
```
# Set current project
/project acme-corp/sample-project
# Check where we left off
/gaps
# Continue reverse prompting or process new inputs
/intake # or /process
# End session
/checkpoint
```
---
## Success Criteria
- [ ] CLAUDE.md clearly describes the pattern and role
- [ ] README.md is usable by a human starting a session
- [ ] `/project` correctly handles client/project hierarchy
- [ ] `/intake` draws out requirements through reverse prompting
- [ ] `/onboard` analyzes existing projects and fills gaps
- [ ] `/process` handles transcripts and extracts structured content
- [ ] `/gaps` identifies meaningful gaps
- [ ] `/checkpoint` captures session state usefully
- [ ] All commands respect current project context
- [ ] First test produces a seeded sub-project
---
## First Test
After initial build:
```
/project new acme-corp/sample-project
/intake
/process [transcripts]
/gaps
/checkpoint
```
This validates the full seeding workflow.