Letting the AI do the calling — while you check the shape

Here's the practical division of labor. The AI is genuinely good at writing API-calling code: it knows the libraries, the header syntax, the error handling. Let it. You don't write this plumbing by hand.

But you own one thing the AI is bad at: knowing whether the data shape is actually right. A good workflow:

• Give the AI the real docs. Paste the example request and response from the docs into your prompt. The AI guessing an API's shape from memory is a top source of bugs — it will invent plausible-looking keys that don't exist. Real examples anchor it to reality.
• Ask it to show you the raw response first. "Before parsing, log the actual JSON the API returns." Then eyeball it against what the code expects. Does the key it reads (response.current.temp) actually exist in the response? Half of API bugs are a mismatched key name.
• Sanity-check the types. Is that a number or a string? Is a field that should always be present sometimes null? You can read JSON now — use it.
• Handle the unhappy path. Ask: "What happens if this call fails, times out, or hits the rate limit?" If the answer is "the app crashes," that's a bug, not a feature.

You don't need to write the request. You need to read the response and ask "is this the shape I expected?" That single habit — checking the data against the contract — catches the majority of integration bugs before your users ever see them.