Skip to content

ChatGPT Adapter

Authoritative source: skill/adapters/chatgpt.md

Rendered preview

ChatGPT Adapter

  1. Read the consuming project's AGENTS.md and local instructions.
  2. Verify .agent/skills.lock.json against the shared checkout.
  3. Read skill/START-HERE.md, then classify through skill/workflow-router.md.
  4. Load only the selected recurring context; do not preload the full skill/ tree.
  5. Keep mutable plans, reports, and TODOs in the consuming project's .agent/ directory.
  6. Treat the shared checkout as read-only unless the task explicitly changes the skill pack.
  7. Consult skill/registry/reuse-catalogue.json only when a reuse trigger applies.

ChatGPT is reference-only: provide the project files and shared-checkout location in the conversation or connected workspace. 8. For whole-project discussion, use project planning in discovery mode; treat nothing as approved until the user confirms it. 9. For approved multi-iteration milestones, use execution planning and produce stable slices plus one first step. 10. When repository writes are unavailable, return each approved record in a labelled Markdown block with its target .agent/ path.

Complete source

# ChatGPT Adapter

1. Read the consuming project's `AGENTS.md` and local instructions.
2. Verify `.agent/skills.lock.json` against the shared checkout.
3. Read `skill/START-HERE.md`, then classify through `skill/workflow-router.md`.
4. Load only the selected recurring context; do not preload the full `skill/` tree.
5. Keep mutable plans, reports, and TODOs in the consuming project's `.agent/` directory.
6. Treat the shared checkout as read-only unless the task explicitly changes the skill pack.
7. Consult `skill/registry/reuse-catalogue.json` only when a reuse trigger applies.

ChatGPT is reference-only: provide the project files and shared-checkout location in the conversation or connected workspace.
8. For whole-project discussion, use project planning in discovery mode; treat nothing as approved until the user confirms it.
9. For approved multi-iteration milestones, use execution planning and produce stable slices plus one first step.
10. When repository writes are unavailable, return each approved record in a labelled Markdown block with its target `.agent/` path.