What an API is

The fast path, with AI

What AI genuinely shortens here is reading an unfamiliar API's documentation and getting to the first correct call, the job that used to take half a day. But the version that works differs from "write me the code" in one way: instead of code, you first ask for one test call you can run and look at yourself. Writing code after seeing a real response takes ten minutes; writing it before is guesswork.

1. Take the API's documentation page and send the text of that page, not just its link. If the model cannot see the page it answers from memory, and API documentation changes quickly.
2. With the recipe below ask for a single test call that can run with no key or with a test key, and run it yourself. The first thing you see is the status code, and it tells you where you stand.
3. Once you have the real response, hand it back and say you only need these three fields. Now the model sees the real shape of the data and the code it writes matches what actually comes back.
4. Finally ask it to add the failure cases: a timeout, a 429, and a response that is not JSON at all. Those three are exactly the ones that never happen on the first day and certainly happen in the second month.

Copy-ready recipe

Here is the text of this API's documentation:

{documentation page text}

What I want to do: {for example, get the status of an order by its number}

Answer in this order:
1. Give me one test call I can run in a terminal right now, with the correct method, address and parameters. If a key is needed, say where it comes from and which part of the request it goes in.
2. Say which fields to look for in a successful response.
3. List the status codes this endpoint returns and what each one means.
4. Say what its rate limit or quota is. If the documentation does not state it, write that it does not, and do not guess.

Do not write any application code yet; only the test call.

Before you trust the output: Check two things yourself before running any call a model hands you. First, what that call does: a <code>GET</code> usually only reads, but a <code>POST</code> or <code>DELETE</code> may change something in your real account and undoing it is not up to you. Second, an API key is a secret: never put a real key in a command you write into a chat, and if you did it once, revoke that key and make a new one. And one thing that comes up every time: if the documentation says nothing about a request ceiling, the model will usually invent a plausible number. That number came from nowhere.