Tony CantaveHow I Write Prompts That Give Consistent Results

A lot of people judge an AI model by the first answer it gives them. If the answer is vague or off target, they decide the tool is not very good. In my work as a freelance AI evaluator, scoring model output against rubrics and writing Python and JavaScript samples for model training, I have learned that the prompt often matters as much as the model. A clear prompt gets a useful answer far more often, and it gets that answer again the next time you ask.

Here is the process I use when I need output I can rely on, whether I am drafting copy for Local Flyer Ads or writing a small automation script.

1. Start with the outcome, not the topic

"Write about Google Business Profiles" is a topic. "Write a 300 word checklist for a Connecticut barbershop owner who has never logged in to their Google Business Profile" is an outcome. The second version tells the model who the reader is, what they need and how long it should be.

Before I type anything, I ask myself one question: if a person handed this back to me, how would I know it was right? Whatever answers that question goes into the prompt.

2. Give the context the model cannot guess

Models do not know your business, your customers or your constraints unless you say so. I include the facts that change the answer. For a marketing piece, that might be the price, the audience and the town. For code, it is the language version, the input format and what the output should look like.

A good habit is to paste a small sample of the real data. If I want a Python function that cleans a CSV export, I include five representative rows, including one messy row. The model writes much better code when it can see the actual problem.

3. State the constraints plainly

Constraints are where most prompts fall short. If you need plain language, no jargon, a certain reading level, a word limit or a specific format, say it directly. I write constraints as short, separate sentences near the end of the prompt so they are easy to scan. A typical set looks like this.

Keep it under 400 words. Use short paragraphs. Return valid JSON with the keys name, category and hours. Do not invent statistics.

That last line matters. When a model is not told otherwise, it will sometimes fill gaps with confident sounding numbers. Telling it to avoid made up figures, and to say when it does not know something, cuts down on that.

4. Show an example of good output

One example is often worth a paragraph of instructions. If I want product descriptions in a certain voice, I paste one I already like and say "match this tone and length." If I want a function in a certain style, I show a short function written the way I want it.

Be careful with examples, though. Models copy them closely. If every example is about a barbershop, the output may drift toward barbershops even when you ask about a laundromat. Use varied examples when you can.

5. Ask for structure you can check

Free form text is hard to verify. Structured output is easy. When I need the result to feed into another step, I ask for a table, a numbered list or JSON. Then I can check it with a few lines of Python instead of reading every word.

This is the same idea behind rubric based evaluation. A rubric breaks a vague question like "is this good?" into specific checks: did it follow the instructions, is the code correct, is the format right, is anything missing. You can write your prompt so the output is easy to grade against those same checks.

6. Test the prompt more than once

A prompt that works once is not finished. I run it several times, and I try it with a different input or two. If the answers vary a lot, the prompt is leaving too much open. Usually the fix is adding a missing piece of context or tightening a constraint.

I keep prompts that work in a plain text file, with a note about what each one is for and what input it expects. Over time that file becomes a small library I can reuse instead of starting from scratch every time.

7. Review before you use it

Even a great prompt does not remove the need for review. I read every piece of output before it goes anywhere public, and I run every piece of generated code before I trust it. AI is a fast first draft, not a final answer.

The short version

Say what a good result looks like. Give the context that only you know. List your constraints. Show an example. Ask for output you can check. Test it more than once. Review everything.

None of this requires special tools. It is the same discipline I learned in logistics and manufacturing: define the standard, make the process repeatable, and check the result. Prompts are just another process.

If you run a small business in Connecticut and want to talk about using AI tools practically, or about putting your business on a screen inside a busy local shop, reach out at sales@localflyerads.com.