Rwlsetup

If AskUserQuestion is unavailable, ask conversationally.

After gathering answers, confirm the full picture with the user before proceeding.

Step 2: Create prd.json

Create prd.json in the project root using the template in templates/prd.json.

Break the core features into granular, independently-completable tasks. Each task must have:

• A unique sequential id (starting at 1)
• A clear title and description
• Specific, testable acceptance criteria
• "done": false

Order tasks by priority:

1. Architectural decisions and core abstractions
2. Integration points between modules
3. Unknown unknowns and spike work
4. Standard features and implementation
5. Polish, cleanup, and quick wins (README, formatting)

Present the draft PRD to the user for review. Adjust if they request changes.

Step 3: Create progress.txt

Create progress.txt in the project root:

# <project name> — Progress Log

Replace <project name> with the actual project name.

Step 4: Create PROMPT.md

Create PROMPT.md in the project root using the template in templates/prompt.md.

Replace the following placeholders with actual values:

Derive the commands from the tech stack chosen in Step 1. Examples:

If the stack has no build step, replace {{BUILD_CMD}} with a comment explaining why (e.g. # Python — no build step). The agent will still gate commits on the remaining checks.

Step 5: Initialize Git (if needed)

If the project directory is not already a git repository:

git init
git add -A
git commit -m "chore: initial project setup with PRD and prompt"

Step 6: Confirm Setup

List the files created:

• prd.json — product requirements with task breakdown
• progress.txt — append-only progress log
• PROMPT.md — iteration prompt for the agent

Tell the user to start with a single iteration to make sure everything works:

./ralph-loop.sh 1

After that first run, they should review the commit(s), check progress.txt, and verify the task was marked done in prd.json. Once they're satisfied, they can let it rip:

./ralph-loop.sh 20