Skip to main content

AI Developer Guide

Instructions for AI coding assistants working on this project.


Why This System​

AI coding assistants are powerful but need structure to be effective. This system has three layers:

  1. The Cage — The AI runs inside a devcontainer. It can only see the project directory. Your machine, SSH keys, and other projects are protected.
  2. The Plan — The AI creates a plan before writing code. You review it. This prevents hallucinations, scope drift, and wasted work.
  3. The Tests — Validation catches mistakes. The AI runs checks after each phase and self-corrects before you review.

Documents​

DocumentPurposeWhen to Read
WORKFLOW.mdEnd-to-end flow from idea to implementationWhen starting new work
PLANS.mdPlan structure, investigation guidance, templatesWhen creating or implementing a plan
DEVCONTAINER.mdHow to work inside the devcontainerWhen running commands or installing tools
GIT.mdGit safety rules and platform operationsWhen doing git operations
COORDINATION.mdFleet coordination: where gh applies and where urb doesWhen working with other agents
SECURITY.mdBefore writing anything sensitive into a published siteBefore publishing
VERIFICATION.mdHow to prove a change worksBefore declaring work done
WORKTREE.mdGit worktreesWhen working on parallel branches

Start Here​

When starting a new session, read files in this order:

  1. Read all project-*.md files — project-specific setup, tools, commands, conventions
  2. Read all template-*.md files (if any) — tech stack from installed templates
  3. Read WORKFLOW.md when starting new work
  4. Read PLANS.md when creating or implementing a plan
  5. Reference DEVCONTAINER.md, GIT.md, COORDINATION.md, VERIFICATION.md as needed

File Naming Convention​

PrefixMeaningPortable?Created by
(none)Universal workflow docsYes — copy to any projectCopied from template
project-*Project-specific setup and conventionsNoProject maintainer
template-*Tech stack from installed templateNodev-template command
plans/Implementation plansNoAI + maintainer

Plans Folder​

Implementation plans are stored in plans/:

plans/
├── backlog/ # Approved plans waiting for implementation
├── active/ # Currently being worked on (max 1-2 at a time)
└── completed/ # Done - kept for reference

File Types​

TypeWhen to use
PLAN-*.mdSolution is clear, ready to implement
INVESTIGATE-*.mdNeeds research first, approach unclear

Quick Reference​

When user says "I want to add X" or "Fix Y":​

  1. Create INVESTIGATE-*.md or PLAN-*.md in plans/backlog/
  2. Ask user to review the plan
  3. Wait for approval before implementing

When user approves a plan:​

  1. Ask: "Do you want to work on a feature branch? (recommended)"
  2. Create branch if yes
  3. Move plan to plans/active/
  4. Implement phase by phase
  5. Ask user to confirm after each phase

When implementation is complete:​

  1. Move plan to plans/completed/
  2. Create Pull Request if on feature branch

Project-Specific Instructions​

See CLAUDE.md in the repo root for always-loaded rules (workflow enforcement, container rule).

See project-*.md files in this directory for project-specific setup, tools, and conventions.