Write Apps SDK tool names, descriptions, parameter docs, and annotations so ChatGPT reliably selects the correct tool

Documented steps

1. Name tools with a domain-action pairing (e.g. `calendar.create_event`) so the model can infer scope and function from the name alone.
2. Open each tool description with 'Use this when...' to state the intended scenario explicitly.
3. State disallowed cases directly in the description (e.g. 'Do not use for reminders') to suppress mis-selection.
4. Give every parameter a description and examples, and use enums for constrained values.
5. Set the annotations: `readOnlyHint: true` for retrieval-only tools, `destructiveHint: true` for tools that delete or overwrite data, and `openWorldHint` true/false depending on whether the tool reaches outside the user's account; add `idempotentHint: true` when repeat calls with identical arguments have no extra effect.
6. Optionally set `_meta["openai/toolInvocation/invoking"]` and `_meta["openai/toolInvocation/invoked"]` status strings (max 64 characters each).
7. Iterate like product copy: change one metadata field at a time and re-test against a golden prompt set covering direct, indirect, and negative prompts; optimize precision (no wrong-tool selection) before recall.
8. Official docs: https://developers.openai.com/apps-sdk/guides/optimize-metadata