Files
Maintainerr/README_AGENTS.md
enoch85 8ab6fc10f1 docs: unify agent instructions into a single index, add handoff notes (#2980)
- AGENTS.md is the single documentation index; the Claude, Copilot, Cursor,
  and Codex entrypoints route to it and name the standing rules directly so
  they can't be missed.
- Split standing rules (read every session) from task-specific docs (read on
  demand); scope release-review's Copilot applyTo to release artifacts so it
  no longer loads on every interaction.
- Add project-notes.instructions.md (non-obvious project knowledge and
  conventions for handoff) and README_AGENTS.md (the wiring map).
- Move dev mocks + DB seed to tools/dev/ (fix seed-db repo-root resolution);
  add the seeded-DB + Playwright step to the release-review checklist.
2026-05-25 22:24:21 +02:00

2.9 KiB

Agent instruction wiring

How this repo's AI coding agents (Claude, GitHub Copilot, Cursor, Codex) load their instructions. Each agent auto-loads a different entrypoint, but they all converge on AGENTS.md as the single documentation index, and each entrypoint also names the two standing rules directly so they can't be missed.

This file is the single source for the wiring; AGENTS.md links here.

CLAUDE
  auto-loads → .claude/rules/implementation.md
       ├─→ AGENTS.md ........................ (doc index)
       │      ├─→ implementation.instructions.md   [standing]
       │      ├─→ project-notes.instructions.md    [standing]
       │      ├─→ release-review.instructions.md    [task-specific]
       │      └─→ ARCHITECTURE.md                   [task-specific]
       └─→ implementation.instructions.md  +  project-notes.instructions.md
              (named directly → read before any code, can't be missed)
  also: SessionStart hook injects AGENTS.md (belt-and-suspenders)

COPILOT
  auto-loads → .github/copilot-instructions.md
       ├─→ AGENTS.md → (same index as above)
       └─→ implementation.instructions.md  +  project-notes.instructions.md  (named)
  also auto-applies via applyTo:"**" → implementation.instructions.md, project-notes.instructions.md
       release-review → applyTo scoped to CHANGELOGs/release workflows (not every file)

CURSOR
  auto-loads → .cursor/rules/project.mdc  (alwaysApply: true)
       ├─→ AGENTS.md → (same index as above)
       └─→ implementation.instructions.md  +  project-notes.instructions.md  (named)

CODEX
  auto-loads → AGENTS.md  (the index itself)
       ├─→ implementation.instructions.md   [standing]
       ├─→ project-notes.instructions.md    [standing]
       ├─→ release-review.instructions.md    [task-specific]
       └─→ ARCHITECTURE.md                   [task-specific]


all four entrypoints ──────────────→ AGENTS.md  (single doc index)
project-notes.instructions.md ──→ ARCHITECTURE.md, AGENTS.md   ✓
implementation.instructions.md ─→ ARCHITECTURE.md              ✓

Rules of the structure (keep it working)

  • AGENTS.md is the single index. Add any new doc to its "Documentation map".
  • Standing rules (read before any code): implementation.instructions.md and project-notes.instructions.mdapplyTo: "**" and named in every entrypoint.
  • Task-specific (read on demand, not every session): release-review.instructions.md (Copilot applyTo scoped to release artifacts) and ARCHITECTURE.md.
  • Each agent entrypoint is a thin router to AGENTS.md + the two standing rules. When you change the wiring, update all four entrypoints together: .claude/rules/implementation.md, .github/copilot-instructions.md, .cursor/rules/project.mdc, and AGENTS.md.