The mental model — you're not asking, you're specifying

1. The reframe

The mistake at the root of most bad prompts is treating the model like a coworker.

When you ask a coworker make this better, it usually works. They have ten years of context on your codebase, your team's style, last week's incident, the constraints you're under. They fill in the gaps because they share your world.

The model shares none of it. It has whatever you typed and nothing else. The same words that mean please fix the validation logic the way we've been doing it to a coworker, to a model mean literally anything plausible.

That's the gap. The fix isn't "ask better" — it's stop asking. Start specifying.

2. The same task, two ways

Same task: cleaning up a messy function. Two ways to write the prompt.

Hey, can you refactor this and make it cleaner? Thanks!

Refactor the function below with these goals, in priority order:

1. Reduce cyclomatic complexity to <= 5 (currently 12 from the nested ifs).
2. Pull the validation block into a named helper.
3. Preserve the public signature exactly.
4. Keep the existing tests passing — don't change them.

Return only the refactored code. No commentary.

The request version is what a person would say. The spec version is what the request actually meant, unpacked. The model can't do step 1 of the spec by guessing it from cleaner. Either you say it, or the model picks for you. Pick three times in a row and three different versions come back — that's the cost.

3. Why this matters more than any prompt trick

Most prompt-engineering advice is small surface tweaks: always say "step by step," always start with a role, always use bullet points. Those help, at the margin, when the underlying prompt is already a spec.

When the underlying prompt is a request, no surface tweak can save it. Please make this code better, step by step, as a senior engineer, in bullet points is still asking the model to guess what better means. It's just guessing in bullet points now.

The reframe — I'm not asking, I'm specifying — is upstream of every prompt trick. Internalize it once and every prompt you write after gets clearer, without learning any new technique.

4. Try it

Pick the last prompt you sent that produced a disappointing result.

Three lines of spec usually beats three paragraphs of polite request. Send the spec version next time and watch the result tighten.