I score websites for a living. I look at 60 signals across 5 dimensions and tell a business owner what is working and what is not. A prompt is the exact same problem from the other direction: you are giving an AI a set of signals and asking it to produce a specific output. The better the signals, the better the output. The fewer gaps the AI has to fill in with its own guesses, the closer what it produces will be to what you wanted.
Most prompts people send me for review are missing at least two of the five elements below. Often they are missing three. And then the person wonders why the output was generic or wrong. It was generic because the input was vague. Here is how to stop doing that.
The five parts every prompt needs
Role. Who is the AI being right now? A good role is a specific professional identity: "You are a senior bookkeeper specializing in small construction businesses." A bad role is a platitude: "You are a helpful assistant." Specificity in the role gives the AI a voice, a knowledge boundary, and a set of assumptions to operate from.
Context. What does the AI need to know to do this task well? This is the part most people skip. It includes what your business does, who your customer is, what just happened, what has been tried before, and anything the AI could not possibly infer. If a new employee walked into your office and you gave them this task, what would you need to brief them on? That is your context block.
Task. State the task as an active, specific directive. Not "help me with X." Not "I am thinking about X." Say: "Draft X." "Summarize Y." "List Z." A task is a verb plus a specific object.
Output format. What shape should the result take? A paragraph? A bulleted list? A table? Three options? JSON? If you do not specify, the AI will pick for you — and you will get inconsistent shapes across runs.
Constraints. What are the guardrails? Length limits. Tone requirements. Things it must mention. Things it must never mention. Words you ban. Output it should avoid. This is where you prevent the AI from going off-brand.
Bad prompt vs. good prompt
Here is a prompt I see all the time, and the same request rebuilt with the five-part anatomy.
The second prompt will produce a useful draft every time. The first one produces a coin flip.
Why specificity wins
Here is the principle: vague in, vague out. When a prompt has gaps, the AI fills them with its best guess about what a generic version of your request looks like. The more gaps, the more generic the output. The fewer gaps, the closer the output is to what you would have written if you had the time.
Think of it the way I think about websites. When I score a site, I look at how clearly the business describes itself, whether the services are explicit, whether the authority signals are verifiable, whether the entity is unambiguously established. A site with strong signals gets recommended by AI. A site with weak signals gets skipped. Prompts work the same way. The stronger your signals to the AI about what you want, the stronger the output.
Before you hit send on any prompt you plan to reuse, run it against these five: role (is it specific?) · context (does the AI know what it needs to know?) · task (is it an active verb on a specific object?) · output format (is the shape locked?) · constraints (are the guardrails in place?). Missing one is a gap. Missing two is a coin flip.
The next step
Once your prompts have the right anatomy, the question becomes whether they work consistently. A prompt that produces a great result once is not the same as a prompt you can deploy fifty times. That is what testing is for — and that is the topic of my next piece, How to Test Whether Your Prompt Actually Works.