Why my prompt doesn't work — 3 causes and how to fix each one

You wrote what seemed like a fine prompt, and back came waffle, the wrong thing, or a total miss. First thought: "the model is dumb." Almost always it's not the model but three specific things, and each fixes in a minute. Let's go by symptom — from the most common to the rarest.

Symptom: the answer is generic and "about nothing"

You ask, and you get averaged-out textbook text — as if the model answered in general, not to you.

Cause #1, the most common: the prompt is too vague. You didn't set a role, a format, or context, so the model guesses — and guesses toward the average. "Write a post about a café" it can do a thousand ways, and it'll pick the blandest one.

How to check: reread your prompt and ask — would a real person understand exactly what I want? No who for, in what form, and with what constraints — there's your cause.

How to fix: add a role, context, and a format. Not "write a post," but "you're a social-media editor; a café by the metro; audience — office folks; 3 paragraphs plus a question at the end; no word 'unique.'" These are the basics from the guide how to write a good prompt — 80% of "it doesn't work" is cured right here.

Symptom: the model does only part of it, or gets confused

You gave a detailed prompt, and it ignores some instructions or does the opposite.

Cause #2: the prompt is overloaded or self-contradictory. Too many tasks at once — or one instruction fights another ("be brief" and right after "describe each point in detail"). The model doesn't choose wisely, it stumbles.

How to check: count how many different tasks are in one prompt. More than two or three — overload. And look for direct contradictions in the requirements.

How to fix: break it into steps. Do one thing first, check it, then the next — instead of "do it all at once." And remove conflicting requirements. If the model stubbornly ignores a specific instruction — pull it out separately and to the front; more in the breakdown why the model ignores instructions.

Symptom: a long prompt, and the model loses the point

The prompt is big, with all the context, but the model seems to miss the single most important requirement.

Cause #3: the important bit is buried in the middle. Models have a known bias: they hold the start and the end of a long text better, and "sag" in the middle (it's literally called "lost in the middle"). A key instruction wedged into paragraph 7 drops out easily.

How to check: find the single most important requirement in your prompt. Is it somewhere in the middle of a wall of text? There it is.

How to fix: move the key thing to the start or the very end of the prompt. Trim the extra context — the shorter it is, the less gets lost. And don't stretch the prompt to the edge of the context window: the closer to the limit, the worse the model holds details.

Bonus: maybe you're asking for something the model doesn't know

Sometimes the prompt isn't the issue. Ask for fresh facts (yesterday's news, exact current prices) or data the model doesn't have, and it'll make it up confidently and wrong. A prompt won't help here: you need either access to the source, or search, or to paste the facts straight into the request. First separate "bad prompt" from "the model can't know this."

Q: Is it the prompt or a weak model?

In 9 cases out of 10 — the prompt. Before blaming the model, run the three checks above: add specifics, unload it, move the key thing to the front. If even with a clear, short, unambiguous prompt the answer is bad — then it's worth trying a stronger model.

Q: Why does the same prompt give good then bad?

Models have a dash of randomness (the "temperature" handles it), so answers drift a little from run to run. If the spread is large, that's a sign the prompt isn't tight enough: the sharper the spec and format, the steadier the result. A good prompt narrows the corridor the model can swerve in.