The 12 Most Common Prompt Mistakes and Their Fixes

How do you find the prompt mistake from the answer?

The answer's symptom often points to the location of the gap in the prompt. If the text is generic, the context and audience are missing. If facts were invented, there's no source and no unknown-information rule.

If the structure changes every time, the output format wasn't written. If the answer is long but useless, no success criterion was defined.

OpenAI's current model guidance recommends writing the goal, context, constraints, required evidence, success criteria and output format concretely. The same document stresses reducing repeated instructions and measuring changes on real tasks. The 12 mistakes below show how those principles get broken in practice.

4 prompt mistakes about goal and context

1. Writing the task but dropping the goal

Weak: "Summarise this report." The model doesn't know whether the summary is for a decision, for the archive, or for a client presentation.

Fix: "Summarise the report for the manager's budget decision tomorrow. Keep only revenue, costs, variance from plan and the risks requiring a decision." The goal also determines what to cut.

2. Leaving the context to the model's guesswork

Weak: "Write a plan to increase sales." No product, market, channels, current results or constraints.

Fix: give the product's price, the audience, the last 90 days' channels, the current conversion point and the usable resources. After "why?" must come "where are we now?"

3. Not labelling the input data

Pasting a long document under the prompt and writing "analyse this" is not enough. It's unknown which document the model should treat as primary, which as supplementary, and which date range as reliable.

Fix: name the inputs: "Source A is the contract, Source B is the price list. In a conflict, the contract prevails. Use only rows dated 2026." For long text, give section boundaries too.

4. Describing the audience as "everyone"

"Write simply, so everyone understands" is not a concrete instruction. A 15-year-old student, a small business owner and a CFO don't expect the same vocabulary, examples and detail.

Fix: write the reader's role, their knowledge of the topic, the decision they'll make after the answer, and the terms they don't know. A role-assignment prompt should work here as a viewpoint and responsibility, not a title.

4 prompt mistakes about instructions, format and evidence

5. Giving contradictory instructions

Pairs like "write in detail, but keep it very short" and "be creative, use only the given sentences" give the model no rule for choosing. The answer can honour one requirement and break the other.

Fix: write the priority: "Maximum 180 words. If completeness collides with the word limit, keep only the three main risks." If the contradiction can't be removed, show which rule wins.

6. Inflating the prompt with repetition and decoration

Repeating the same instruction in three forms doesn't make it three times as important. The words "perfect, world-class, incredibly professional" don't replace acceptance criteria and bury the important rule in the text.

Fix: write each instruction once, delete synonym repetition, and describe the style through observable behaviour. Instead of "write professionally," "start with the conclusion, tie every claim to evidence, use no ad slogans" is clearer.

7. Expecting a specific style without examples

"Write in my style" is an empty instruction if the model has no reliable sample of that style. The same problem appears in classification, JSON output and brand terminology.

Fix: give two or three short, correct examples; write which aspect of the example should be imitated. Anthropic's current prompting guide explains using relevant examples alongside clear instructions to steer the output's format and behaviour.

8. Closing the output format with the word "tidy"

"Give a tidy report" means paragraphs to one user and a table to another. If the result will be written into a system, that ambiguity costs even more.

Fix: write the section order, column names, field types, maximum length and empty-value behaviour. In the 10 rules of prompt writing, format is a part checked separately from the goal.

4 prompt mistakes about sources, safety and testing

9. Asking for current facts without sources

Prices, laws, product features, statistics and platform rules change. Saying "write with the latest information" gives the model no live source and no date.

Fix: write the trusted sources, the data date and the citation requirement. If there's no source access, require it to state that openly and to leave out unverified numbers.

10. Not defining behaviour for unknown information

When the model fills an empty field with an invention, the problem isn't only "hallucination"; the prompt also never said what to do with the unknown. And "don't make mistakes" is not an executable rule.

Fix: add behaviour like "if it's not in the source write 'not found', don't present a guess as fact, ask a question if it matters for the decision." In chain of thought too, long reasoning doesn't make an unknown fact true.

11. Copying confidential data into the prompt as-is

Customer lists, identity data, contracts, medical notes and access keys should not be passed in full "so the model understands better." Prompt quality doesn't cancel data minimisation.

Fix: keep only the necessary fields, anonymise the data and use an environment your organisation has approved. NIST's Generative AI Profile emphasises mapping, measuring and managing risk across the lifecycle.

12. Declaring the prompt done after one good answer

A prompt can work on a normal sample and fall apart on incomplete and risky input. Liking one answer is not evaluation.

Fix: build a fixed test set from typical, borderline, incomplete, contradictory and risky samples. OpenAI's evaluation guide recommends data representing the real usage distribution, concrete criteria, human review and continuous testing after changes.

How do you fix a weak prompt in 5 minutes?

Before rewriting the prompt from scratch, fill in the eight lines below. If a line doesn't affect the result, delete it. A short prompt isn't bad; a vague prompt is.

Goal: [which decision or work is the result for?]
Audience: [who will read it, what do they know?]
Context: [the current situation and constraints]
Input: [the data to use and its precedence order]
Task: [what exactly must the model do?]
Sources and unknowns: [what to rely on, how to mark what isn't found?]
Output: [format, order, length and mandatory fields]
Acceptance criteria: [when is it done, who checks?]

Then test the old and new variants of the same task on at least 10 identical inputs. Compare the results by factual errors, mandatory fields, format, editing time and stopping on risky input. Prompt engineering is this testing discipline more than it is word choice.

If you're producing articles with AI, add one more gate: original observation, author responsibility and reader benefit. Google Search's guidance on generative AI content doesn't prohibit AI use in itself, but says creating many pages without value to users can violate the spam policy. That's why the prompt "write an SEO article" isn't enough without sources, experience, an editor and a purpose.