Exploring patterns for long-running, autonomous AI development sessions.
This repository documents strategies and workflows for enabling AI agents to work independently on complex software projects for extended periods without constant human intervention.
Use formal design documents before implementation:
- Design documents (
memory/design/) define what and why - Milestone plans (
memory/milestones/) break work into testable chunks - Design stacking - new designs build on previous ones (001 → 002 → 003)
See: commands/design_system.md, commands/plan_milestones.md
Use tagged, structured logging throughout the codebase:
- Tags like
[WEBSOCKET],[AGENT],[API]enable filtering - Log to files (
logs/app.log) to avoid context pollution - Log decision points, state changes, and API calls
- Reduce verbosity after each milestone
Write tests first, run them constantly:
- Unit tests - individual functions/components
- Integration tests - API endpoints, database, external services
- E2E tests - full user flows via Playwright
- Run tests after every small change, not just at the end
Leverage Model Context Protocol servers for enhanced capabilities:
- Context7 - fetch up-to-date library documentation before integration
- Playwright - browser automation for E2E testing and visual verification
Iterative development cycle:
Design → Milestones → Implement → (New) Design → (New) Milestones → ...
Each iteration brings you closer to the long-term goal.
Voice chat application with multiple AI agent backends.
Prompt:
I want you to make a React + Express (typescript) app that will let me voice chat with either OpenAI (either voice-to-voice, or STT-LLM-TTS) or Eleven Labs. I want to be able to hot swap at any time.
@commands/build_project.md
Reusable workflow templates in commands/:
build_project.md- Main orchestration guidedesign_system.md- Design document templateplan_milestones.md- Milestone planning guide