The gap between a useless AI answer and a usable one is almost never the model, it is the brief. AI Central's How to Write Great Prompts sorts every request into three tiers. A bad prompt is a one-line wish. A good prompt makes the model produce several drafts and puts you in the judging seat. A great prompt is a six-part brief, a role, a task, a format, constraints, a stop condition, and a dump of your own context.
Reviewed on August 7, 2026.
Why most prompts fail before the model reads them
AI Central works through seven ordinary jobs, a LinkedIn post, a newsletter, an SEO plan, getting mentioned inside ChatGPT answers, market research, choosing what to work on next, and setting strategy. Each appears three times over, once badly, once well, once properly. The bad versions are the recognisable ones, and not one of the seven runs longer than eight words.
Write a viral LinkedIn post about entrepreneurship. Give me an SEO strategy for my website. Research my market. These are not badly worded, they are empty, and AI Central opens with a one-line diagnosis of why people write them.
Most people use AI like a search engine
The alternative, in AI Central's framing, is the second half of the same thought.
Founders use it like a co-founder
That distinction is doing real work. A search query asks for something that already exists. A brief asks for something only you can specify. A seven-word prompt fails because every word in it could have been guessed by a stranger, so the model answers for the average of everyone who ever asked, which is exactly the generic output people then blame the tool for.
The middle tier, let the model draft and keep the judgement for yourself
The second tier in AI Central's ladder is the one most people can adopt this afternoon, and all seven examples build it the same way. Ask for several drafts rather than one. Pick the strongest. Say out loud why it is the strongest. Then have the model rewrite using that reason. The LinkedIn version reads like this.
Write 5 LinkedIn posts aimed at founders building real businesses. I'll pick the strongest one, explain why it worked, then you'll rewrite 5 more using that feedback
All seven good prompts contain that same loop, worded for the task. In the market research version the user promises to point out what is useful and what is not. In the buying-behaviour version, to challenge whatever feels theoretical. The verb changes, the mechanism does not.
It works because it moves the hard part to where you are competent. Most people cannot describe their own taste in advance, which is why the first tier fails. Almost everyone can recognise the better draft when five sit in front of them. Explaining the choice converts a private preference into a rule the model can apply to the next five, and it costs two sentences.
The six slots inside a great prompt
The top tier is not a longer sentence, it is a structured brief. All seven of AI Central's worked examples use the same six slots in the same order, and each removes a specific failure mode.
- Role, who the model is answering as. It sets the vocabulary, the assumed experience, and the things a person in that seat would not bother saying.
- Task, the job itself, usually with your own material pasted in and a two-step instruction, analyse this first, then produce that.
- Format, the shape of the output, so you are not handed an essay when you needed a list.
- Constraints, the boundaries, which in practice are mostly prohibitions.
- Stop when, the definition of done.
- Context, a dump of your situation, which every single example ends with.
The role lines are the tell. All seven cast the model as a founder or founder-operator, never as an assistant and never as a generic expert, and the sharpest of them is written as a negation. AI Central's market research brief opens this way.
Act as a founder trying to make a call, not a consultant explaining the market
That line pre-empts the most common disappointment in AI output. A model asked to explain a market will explain it thoroughly, for pages, and leave you no closer to a decision. Naming the seat also names the output the seat produces.
The stop condition is the slot almost nobody writes
Every one of AI Central's seven great prompts ends its instruction block with a stop condition, and this is the least imitated part of the whole structure. One asks that a clear, executable plan exists. Another asks that a 30-day action list is complete. The strategy brief sets the bar in physical terms.
The strategy fits on one page and guides decisions
Two things happen when you write that line. The model stops optimising for volume, because more output no longer counts as progress, and starts checking its draft against a target. You also get an acceptance test. If what comes back does not meet the stop condition, that is not a matter of taste to agonise over, it is a re-run, and you know which part failed.
Constraints are mostly there to say no
Five of the seven constraint blocks in AI Central's examples contain at least one line that begins with the word no. No generic motivation. No filler. No theory. No speculation. The prioritisation brief is the bluntest of them.
No mindset or motivation advice
A method note on that count. It comes from reading AI Central's seven worked examples and counting the constraint lines by hand, so it describes these seven briefs and nothing wider. The pattern is consistent enough to be deliberate. Positive constraints are weak, because a model already believes it is being useful and specific. Prohibitions are strong, because they name the default behaviour you have seen a hundred times and rule it out before it appears.
The same instinct shows up inside the task slots. AI Central's strategy brief begins by stating what it does not want at all.
I don't want a strategy deck
Context is the slot doing the heavy lifting
All seven of AI Central's great prompts close with a placeholder for a context dump, and two of them also paste in the user's own proven work, five LinkedIn posts that performed well, three newsletters that readers replied to. That is the difference between asking for content and asking for more of your content.
Your context is the only input a model cannot generate for itself. Everything else in the brief is steering. The context is the raw material, and a perfectly structured prompt without it still returns an averaged answer, just a better organised one.
What to do with this
The structure scales down cleanly, so match the effort to where you are.
- If you are starting out, do not rebuild your prompts. Take the last vague request you sent and add two lines, a role and a stop condition. That alone moves most outputs from unusable to editable.
- If you already use AI daily, make the middle tier a habit. Ask for three to five options, choose the best, and write two sentences on why before asking for the rewrite. The explanation is the part that compounds.
- If you are building repeatable work, keep a written context dump about your business and a short set of role lines you trust. The six slots become a template you fill in, and a great prompt takes about ninety seconds to assemble.
- Whatever your level, write the stop condition first. If you cannot describe what finished looks like, the model has no chance of producing it, and you learned something useful about the task before spending a token on it.
Do I have to use all six parts every time?
No. AI Central's own middle tier is proof that a two-turn loop with no formal structure already beats a one-line request. The six slots earn their keep on work that matters, a strategy you will act on, a plan you will hand to someone. For a quick rewrite, a role and a stop condition are usually enough.
What actually goes in the context dump?
Whatever the model cannot infer. In AI Central's examples that means your own work that performed well, what you sell and to whom, the revenue band of your customers, the constraint you operate under such as a small team or limited time, and the decision you are trying to make. Long and messy is fine, nobody else reads it.
Does this only work for founders?
The seven examples are founder tasks and every role line casts a founder, so the framing is deliberate. The slots are neutral. Swap the role for the seat that fits your work, a hiring manager, a clinician, a teacher, and the structure behaves identically. Nothing in it depends on a particular tool or job title.
How do I know if my prompt was good enough?
Read the output against your stop condition, not against your mood. If the condition said a 30-day action list is complete and you got four paragraphs of background, the brief failed at the format and stop slots, and you fix those two rather than starting again. A prompt without a stop condition cannot be debugged this way, which is the practical argument for writing one.