Project Setup Plan: Project Creator

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:

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:

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)

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.