Prompts & Writing
The Anatomy of a Prompt That Actually Works
By Jim Vernon, Editor, AI Intelligence International · Published 10 January 2026 · Reviewed against our editorial standards · About the author
Prompting advice tends to be either magic phrases or vague encouragement to be specific. Neither helps when the output is subtly wrong and you cannot say why.
A working prompt has six identifiable components. Missing components produce predictable failures, which means diagnosis is possible: read the bad output, identify which component was absent, and add it.
Key takeaways
- The six components: Role and audience, task, input material, constraints, output format, and success criteria.
- Why grounding matters more than phrasing: A model given nothing to work with will produce the average of everything it has seen, which reads as competent and says nothing.
- Constraints do the heavy lifting: Negative constraints are underused.
- Format specification: Say exactly what you want back: a table with named columns, five bullets of under fifteen words, a JSON object with defined keys, a two-paragraph answer with no headings.
The six components
Role and audience, task, input material, constraints, output format, and success criteria. Most failed prompts contain only the task.
Role sets vocabulary and depth. Audience sets what can be assumed. Task states the verb. Input material grounds the work in your reality rather than generic knowledge.
Constraints say what to avoid: length, jargon, claims you cannot support, structures you dislike. Format says exactly what the response should look like. Success criteria tell the model how to judge its own draft before returning it.
Why grounding matters more than phrasing
A model given nothing to work with will produce the average of everything it has seen, which reads as competent and says nothing. Supplying real material — your notes, your data, a competitor's page, last quarter's report — changes the output more than any wording change.
The practical rule: if you did not paste anything in, expect generic output and do not blame the model.
Length of input matters less than specificity. Three concrete facts beat three pages of background.
Constraints do the heavy lifting
Negative constraints are underused. 'Do not use the words leverage, seamless or unlock', 'do not open with a rhetorical question', 'never claim a benefit not present in the source' each remove a whole class of bad output.
Constraints also make review faster, because you can check compliance mechanically before reading for quality.
Keep a personal constraint list and reuse it. It is the single highest-return prompting habit.
Format specification
Say exactly what you want back: a table with named columns, five bullets of under fifteen words, a JSON object with defined keys, a two-paragraph answer with no headings.
Vague format requests produce vague structure, and restructuring output by hand is where the time saving disappears.
When you need consistency across many runs, put the format in a reusable system prompt rather than retyping it.
Success criteria and self-check
Ending a prompt with 'before answering, check that every claim is supported by the material provided and remove any that is not' meaningfully reduces invented content.
Asking for the two weakest points of the draft surfaces problems you would otherwise have to find yourself.
This is also how you convert a one-shot prompt into an iterative one without extra work on your side.
A worked before and after
Before: 'Write a blog post about project management software.' The output will be five hundred words of category description that could belong to any vendor.
After: 'You are writing for operations managers at ten-to-fifty-person construction firms who currently run projects in spreadsheets. Using the three customer quotes below, write 700 words on why spreadsheet handover fails during subcontractor changes. Do not mention any product. No rhetorical questions. End with three specific checks a reader can do this week. Then list the two weakest claims in your draft.'
The second prompt takes ninety seconds to write and removes an hour of editing. That trade is the whole discipline.
Before and after: a weak prompt repaired
Weak: 'Write a professional email about the delay.' The model has no recipient, no cause, no remedy and no length, so it invents all four and produces something that reads like a template because that is exactly what it is.
Repaired: 'Write a 120-word email to a long-standing client whose delivery slipped from Tuesday to Friday because a supplier missed a shipment. Apologise once, state the new date, offer expedited shipping at our cost, no filler openings, plain British English.'
Nothing clever happened there. Role, audience, facts, constraint, tone and length were supplied, and the output stops guessing. Most disappointing prompts fail because one of those six is missing, not because the model is weak.
Order the instructions by what must not be broken
Long prompts lose their middle. Put the constraints that would make the output unusable — word count, format, prohibited claims, language variant — at the start and repeat the single most important one at the end.
Keep examples close to the instruction they illustrate rather than in a block at the bottom. Models follow the pattern nearest the request more reliably than a distant appendix.
If a prompt exceeds roughly a page, split it. One prompt that drafts and a second that checks against the constraints beats one prompt attempting both, and the second pass is cheap to run.
Debugging a prompt in three passes
First pass, remove ambiguity: replace every adjective the model has to interpret ('concise', 'engaging') with a measurable instruction ('under 150 words', 'open with the customer's problem').
Second pass, add a negative constraint for whatever went wrong last time. Explicit prohibitions work better than hoping; 'do not open with In today's fast-paced world' genuinely removes it.
Third pass, show one example of a good output. A single example moves quality further than three more paragraphs of description, because it encodes tone, structure and length simultaneously.
Frequently asked questions
Do longer prompts always work better?
No — denser prompts do. Adding background that does not constrain the output makes the response worse by diluting the instructions that matter.
Should I use role prompting?
It helps set register and depth, but it is the least important of the six components. Grounding material and constraints matter far more.
Why does the same prompt give different results?
Sampling is stochastic by design. For consistency, pin the format, supply the same source material, and reuse a system prompt rather than relying on one-off wording.
How do I stop the output sounding like AI?
Ban the vocabulary and structures you dislike explicitly, supply your own writing as a style sample, and rewrite the opening and closing by hand.
Does politeness improve output?
Not measurably. Clarity does. Please and thank you are harmless habits, but they are not what changes the result.
Should I keep prompts short or long?
As long as the constraints require and no longer. Padding dilutes the instructions that matter and makes the prompt harder to edit later.