How to Prompt AI for Code That Actually Works

Models are genuinely good at code now. They still fail constantly — not because they can't write the function, but because they're guessing at everything you didn't tell them. Good coding prompts remove the guessing.

Context is the whole game

The number one cause of bad AI code is missing context. The model invents a language version, a framework, a data shape — and it invents wrong. Front-load the facts:

• Language and version (“TypeScript 5.4, strict mode”)
• Framework and version (“Next.js 14 App Router, not Pages”)
• Constraints (“no new dependencies,” “must run in a Deno edge function”)
• The actual surrounding code, not a paraphrase of it

Define the contract: inputs, outputs, edge cases

Describe the function like you'd describe it to a contractor who bills by the surprise. What goes in, what comes out, and what happens at the edges:

Naming the edge cases up front does two things: you get correct code the first time, and you force yourself to actually decide what should happen — which is usually where the real bug was hiding anyway.

Ask for assumptions and tests, not just code

A model will happily hand you confident code built on a wrong assumption. Make it surface the assumptions before you trust the output:

The tests aren't just for coverage — they're how you find out whether the model understood the task. If its test cases don't match what you meant, the code won't either, and you caught it in ten seconds instead of in production.

Debug by pasting the error, not “it broke”

“It doesn't work” forces the model to guess what went wrong, and it'll often rewrite the whole thing to dodge a one-line fix. Give it what a good bug report gives a teammate:

• The exact error message and stack trace
• What you expected vs. what happened
• The relevant code and the input that triggered it
• What you've already tried

Have a second model review the code

The model that wrote the code is the worst reviewer of it — it shares its own blind spots and tends to defend its choices. A different model reads the same code cold and catches the off-by-one, the unhandled null, the SQL injection, the race condition.

Running “write” and “review” across two different models is one of the highest-value habits in AI-assisted coding. It's the whole idea behind a draft → critique → fact-check flow — AskOnce runs it across models so the reviewer isn't the author.

Keep it on a short leash

The failure mode of AI coding is asking for too much at once: “build the whole feature.” You get 200 lines you didn't read, one of which is quietly wrong. Work in small, verifiable steps:

• One function or one file at a time
• Run it before asking for the next piece
• When it goes sideways, start a fresh chat instead of fighting a polluted context

You're the one who has to maintain the code. Prompt so that you understand every line before it ships — the goal is leverage, not a black box.