- The app helps the user manage a small local item list directly inside ChatGPT.
| Job | User outcome | Without this app, ChatGPT cannot... |
|---|---|---|
| See current local state | User knows what items exist now | Read live app state |
| Add one local item | User changes app state from chat or widget | Persist the action in app state |
| Clear local sample items | User resets local state | Perform the reset action |
| Value type | App capability | Why it matters in ChatGPT |
|---|---|---|
| Know | Reads current local item state | ChatGPT gets current app context instead of guessing |
| Do | Adds or clears local items | ChatGPT can act on the user's behalf |
| Show | Renders item state in the widget | The user sees state and recovery actions faster than plain text |
- Core capabilities: get current state, add one item, clear local sample items.
- Explicit non-capabilities: production data, accounts, payments, external writes.
- Product features intentionally not ported: navigation, dashboards, settings, multi-page UI.
- Example prompt: Show me what is in this app.
- App behavior: call get_app_state and show the current widget state.
- Maximum clarifying questions before useful output: 0.
- Example prompt: Add an item named "first task".
- App behavior: call add_item with the title already provided.
- Information the app must not ask for again: item title.
- First useful response must explain the app role in one line and deliver value immediately.
- First useful response: I can manage this local ChatGPT app's item list; here is the current state.
- Tool names are concrete: get_app_state, add_item, clear_items.
- Parameters are minimal and structured: add_item only accepts title.
- Outputs include stable IDs or handles: every item has an id.
- Outputs avoid hiding important facts only in free-form text: items are returned in structuredContent.
- Sensitive data is minimized: no sensitive data is required.
- Outputs that another tool/app can reuse: item ids, titles, completion state.
- Control returns to the conversation after each tool call: yes.
- Long tunnel-like flows avoided: yes; each action is atomic.
| Evaluation item | Target | Evidence |
|---|---|---|
| Direct prompts | Correct tool selected | golden-prompts.md |
| Indirect prompts | Useful progress starts quickly | golden-prompts.md |
| Negative prompts | App is not invoked or refuses safely | golden-prompts.md |
| App answer vs ChatGPT alone | App adds clear know/do/show value | verification-report.md |
- No clear know/do/show value.
- App clones a product UI instead of exposing focused capabilities.
- Tool names or parameters are ambiguous.
- First interaction requires unnecessary onboarding.
- Outputs cannot be chained or reused.