Verified against ChatGPT · 2026-06-30
Turn an API description into a working Custom GPT Action in one pass
A setup prompt that generates the three things a Custom GPT with an Action actually needs — system instructions, an OpenAPI schema, and conversation starters — from a plain-English description of the API you want it to call.
The prompt
Ready to copy — highlighted parts are example details you can swap.
You are helping me configure a Custom GPT in GPT Builder that calls an external API through an Action. Produce three labeled, copy-paste-ready outputs.
API DESCRIPTION
A REST API for our internal order-tracking system, returns order status by order ID over HTTPS
ENDPOINTS TO EXPOSE
GET /orders/{id} returns status, carrier, and estimated delivery date
AUTHENTICATION METHOD
API key sent as a header named X-API-Key
1. GPT INSTRUCTIONS — the system prompt for the Instructions field on the Configure tab: what this GPT does, when to call the Action versus answer from its own knowledge, and exactly what to say if the Action call fails or returns an error.
2. ACTION SCHEMA — a valid OpenAPI 3.1 JSON schema covering only the endpoints listed above, with an operationId and a one-line description per endpoint precise enough that the model can decide correctly when to call it.
3. CONVERSATION STARTERS — three example user messages that should trigger this GPT to call the Action, phrased the way a real user would actually type them.
CONSTRAINTS
- The schema must validate as OpenAPI 3.1. No placeholder types or invented fields.
- If API key sent as a header named X-API-Key needs an API key or bearer token, note in the instructions that it belongs in the Action's authentication settings, never hardcoded into the schema or the instructions text.
- The GPT instructions must explicitly tell the model to say the lookup failed rather than invent a plausible-looking result if the Action call errors or times out.Customize the highlighted detailsoptional — the prompt above already works
Why this works
A GPT Action is only as reliable as its OpenAPI schema and the instructions that tell the model when to reach for it: operationId is effectively the function name the model calls internally, and the per-endpoint description is what it reads to decide this is the right tool for this request, so vague or missing descriptions produce a GPT that either never calls the Action or calls the wrong one. Generating the instructions, the schema, and the conversation starters together in one pass keeps them consistent with each other, since a conversation starter that does not actually map to something the schema and instructions agree on is a common cause of a Custom GPT that looks configured but silently fails on its own example prompts. Separating authentication into a named constraint matters because the real auth configuration for an Action lives outside the schema, in its own settings pane, and a model asked to just write the schema will sometimes bake a placeholder key directly into it if not told explicitly where that value actually belongs. The explicit instruction to report a failed call rather than invent a result is the most load-bearing line here: tool-calling models are documented to paper over a failed or errored function call with a fluent, plausible-sounding answer unless directly told that doing so is worse than admitting the lookup did not work.
What you get back
Instructions: "Call get_order_status whenever the user asks about an order, a shipment, or a delivery date. Never estimate a status from memory. If the call returns an error or times out, tell the user the lookup failed and ask them to confirm the order ID, do not guess a plausible status." Schema excerpt: {"paths": {"/orders/{id}": {"get": {"operationId": "getOrderStatus", "description": "Returns current status, carrier, and estimated delivery date for one order by ID.", "parameters": [{"name": "id", "in": "path", "required": true, "schema": {"type": "string"}}]}}}}
Verified against
ChatGPT GPT Builder (GPT-5.1) · 2026-06-30
Changelog
- 2026-06-30 — Initial publish, verified against ChatGPT GPT Builder on GPT-5.1.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
