Write an AI prompt that gives you the same answer every time

Asking for a good result does not produce one. The model has no way to know what good means for your business, so it guesses, and it guesses differently each run. Every vague word in the request is a decision you have handed to a machine that will make it a new way each time.

The model makers say this plainly. Anthropic's documentation opens with "Claude responds well to clear, explicit instructions" and offers a test: show your prompt to a colleague with minimal context and ask them to follow it. "If they'd be confused, Claude will be too." OpenAI's guide frames it as rules: "What should the model do, and what should the model never do?"

Show your prompt to a colleague who knows nothing about the task. If they would be confused, the AI is confused too. That one test catches most bad prompts before you send them.

The four parts a reliable prompt needs

1. 1

   Say what the job is, with a number

   Not "write some social posts" but "write 5 social posts." A named quantity gives you something you can check, and stops the model deciding for you. Anthropic's guidance is to "be specific about the desired output format and constraints", and a number is the simplest constraint there is.

2. 2

   Give the background it cannot guess

   Who the customer is, what you sell, what you charge, what tone you use. The model knows everything about the world and nothing about your business. This is the new-employee problem: a good hire on day one still needs the briefing.

3. 3

   Name the failure you keep getting

   This is the step people skip. If it keeps writing like a brochure, say so: do not use marketing language or unsupported claims. Naming the specific failure works far better than asking for quality, because "quality" means something different every time and "no exclamation points" does not.

4. 4

   Describe the finished thing

   What should the output contain, and in what order? Anthropic recommends numbered steps or bullets "when the order or completeness of steps matters." If you cannot describe the deliverable, the model cannot produce it consistently.

The job:
Write [number] [what you want] for [your business].
This is for [who reads it].
It matters because [what happens if it lands].

What you need to know:
We sell [offer] to [customer]. The problem we solve is [problem].
The result they want is [outcome]. Our price is [price].

Do not use: [the specific thing it keeps getting wrong].

The deliverable:
Return [number] [items], each [length].
Each must include [required parts].
Then recommend which one to use, and say why.

Ask me any questions before you start.

Save it once, so you never type it again

The template is only half the fix. The other half is not retyping it. Both major tools now have a place for exactly this. ChatGPT's Projects let you "upload reference files, and add custom instructions so ChatGPT remembers what matters and stays on-topic," and they are available on free and paid plans. Claude's Projects work the same way and are available to all users, including free accounts, up to five projects.

Put the "what you need to know" section, the part about your business, into the project instructions. Put your price list, your best past example, and your brand words into the project files. Then each new request is just the job and the deliverable, and the background is already loaded.

One more thing OpenAI's guide mentions that applies even if you never touch code: the company recommends pinning "specific model snapshots" to "ensure consistent behavior" and building tests "so you can monitor performance as you iterate". For a business owner that translates to: when a tool updates its model, re-run your saved prompt once and check the result before trusting it again.

The whole method fits on an index card: the job with a number, the background, the failure to avoid, the deliverable. Fill it in once for the task you do most, save it where the tool will remember it, and the different-answer-every-time problem mostly stops.