Your Prompt Isn't Working: A 5-Question Diagnosis

The owner of a vacation rental writes to the AI: "Reply to this booking request." He pastes the email from a family who wants to come in August. The reply is long, enthusiastic, promises "an unforgettable stay," and forgets to say that the rental doesn't accept dogs, even though the family has one. He asks again, adds "be more precise": the reply gets longer, still without the dog.

When a prompt disappoints, people rewrite at random. It's better to look for the cause, and it's almost always found in one of these five questions, to go through in order.

1. Does the AI have the facts it needs?

This is the most frequent cause. The owner never wrote that dogs are not allowed, that the August rate is [PRICE] per week, or that check-in is on Saturday. The AI filled the gaps with what is usually said in a hospitality reply.

Test: reread your request forgetting everything you know "by heart." What remains is what the AI really knows. Add the missing facts, or let it ask you for them.

2. Did it understand the task the way you did?

Sometimes everything is written down and the misunderstanding remains. "Reply to this request" can mean accept, decline, or propose dates. Before correcting, check what it understood:

Don't carry out my request. Restate it in two sentences: what you think you need to produce, for whom, and what you'll do if a piece of information is missing.

If the restatement is off, you know where the problem is without having read a long, useless text.

3. Did you say what the result should look like?

Length, structure, tone, what it should contain and what it shouldn't. A prompt without these details produces a text that's average in every direction: too long, too cheerful, with a closing line like "don't hesitate to contact us." Write what you want to see: "5 lines, formal tone, first sentence answers the question asked, last sentence offers a call." If you're looking for the settings, that's imposing an output format.

4. Are you asking for too many things at once?

A prompt that mixes analysis, writing, recommendation, and formatting often gets four half-successful results. You recognize it by its "and also." Split it into steps and check each one. Prompt chains explain how to link the steps.

5. Is it the conversation or the request?

Two very different situations hide here.

The conversation is polluted. After many exchanges, the AI mixes old and new instructions. If a flaw you had corrected comes back, open a fresh conversation with a rewritten prompt.

The request goes beyond what the AI can do. A very recent fact, a very local detail, the content of a file you haven't provided, a precise statistic: without a source, it fills in. No rewording will change that. Give it the document, or look for the information elsewhere. Reducing hallucinations covers this case.

The corrected rental prompt

Once the diagnosis is made, the owner's fix takes three more lines:

Reply to this booking request: [PASTE THE EMAIL].

Facts: the rental sleeps 6 people, dogs are not accepted, a week in August costs €[PRICE], check-in on Saturday from 4 p.m.
Write in a formal, polite tone, in 6 lines maximum. The first sentence answers the question asked. If a request doesn't match our rules, say so frankly and offer an alternative.

If it still disappoints, fix it in small touches, and test it on a few cases before relying on it.

To do now

Take the last prompt that disappointed you and go through the five questions. Note the one where you hesitated: that's your usual weak spot, and fixing it will improve all your future prompts.