Skip to content

ARCHITECTURE.md

Latest commit

 

History

History
37 lines (25 loc) · 1.04 KB

ARCHITECTURE.md

File metadata and controls

37 lines (25 loc) · 1.04 KB

Architecture

This template has three layers.

1. Instruction Layer

  • AGENTS.md: shared operating rules
  • CLAUDE.md: Claude entry file
  • .claude/agents/: optional reusable Claude subagents

This layer defines how agents should behave inside the repository.

2. Workflow Layer

  • PROJECT_FORM.md: user-owned brief
  • PROJECT_STATE.md: agent-owned execution state
  • RUN_PROMPT.md: repeated session prompt
  • work/: generic work management
  • experiments/: experiment management
  • scripts/workctl.py and scripts/expctl.py: workflow automation
  • Taskfile.yml: repeated command entrypoints

This layer keeps long-running work resumable across sessions.

3. Domain Layer

  • configs/, results/, logs/, reports/, publish/, notes/

This layer holds real project artifacts.

Design Principles

  • Keep project intent in files, not in chat history.
  • Separate user intent from execution state.
  • Separate generic work from experiments.
  • Keep entry files short and obvious.
  • Prefer resumable progress over one-shot prompting.