Skip to content

CLAUDE.md

Latest commit

 

History

History
60 lines (42 loc) · 3.74 KB

CLAUDE.md

File metadata and controls

60 lines (42 loc) · 3.74 KB

CLAUDE.md

This project is managed by @hivehub/rulebook. The authoritative rules come from the imports below. Claude Code loads all of them automatically at session start (see Anthropic memory docs).

Critical rules (highest precedence — apply on every turn)

  1. Read AGENTS.md and AGENTS.override.md before making changes. These contain project-specific conventions that override generic guidance.
  2. Never revert or discard uncommitted work — fix forward. Treat the working tree as sacred; investigate before destructive operations.
  3. Edit files sequentially, not in parallel. When a task touches 3+ files, decompose into 1–2 file sub-tasks.
  4. Run check/type-check before test — diagnostic-first. Cheap diagnostics catch issues that expensive test suites miss or take longer to surface.
  5. If a fix fails twice, escalate — stop, research, or open a team. Do not retry the same approach a third time.
  6. Prefer MCP tools (mcp__rulebook__* and project-specific MCP servers) over shell commands when the equivalent tool exists.
  7. Capture learnings: at the end of significant work, save patterns and anti-patterns to .rulebook/knowledge/ and insights to .rulebook/learnings/.
  8. Never archive a task without docs updated, tests written, and tests passing — the task tail enforces this structurally.

Persistent memory

This project uses the Rulebook MCP server for persistent memory across sessions.

  • Start of session: rulebook_memory_search for relevant prior context.
  • During work: rulebook_memory_save for decisions, bugs, discoveries, user preferences.
  • End of session: rulebook_session_end to write a session summary.

Memory is auto-captured for tool interactions (task create/update/archive, skill enable/disable). Manual saves are required for everything else worth remembering.

Knowledge base

Before implementing anything non-trivial:

  • rulebook_knowledge_list — check existing patterns and anti-patterns.
  • rulebook_learn_list — review past learnings.
  • rulebook_decision_list — review architectural decisions.

After implementing, capture at least one entry per task:

  • rulebook_knowledge_add for reusable patterns or anti-patterns to avoid.
  • rulebook_learn_capture for implementation insights that don't belong in code comments.
  • rulebook_decision_create for significant architectural choices.

Task workflow

MANDATORY: ALWAYS use the Rulebook MCP tools for task management. Never create task directories or files manually — use rulebook_task_create, rulebook_task_update, rulebook_task_archive, rulebook_task_list, rulebook_task_show, rulebook_task_validate. These tools enforce naming conventions, mandatory tail items, phase structure, and metadata that manual file creation skips.

  1. rulebook_task_list to see pending work.
  2. rulebook_task_create to create new tasks — never mkdir + Write manually.
  3. Pick the first unchecked item from the lowest-numbered phase — never reorder.
  4. Read the task's proposal.md and tasks.md before touching code.
  5. Implement step by step. Run lint + type-check after each significant change.
  6. rulebook_task_update to change task status as you progress.
  7. Mark items [x] in tasks.md as you finish them.
  8. The mandatory tail (docs + tests + verify) is not optional — rulebook_task_archive will refuse to close the task otherwise.