What makes an AI prompt good?

A good prompt names the deliverable, supplies the context the model cannot infer, marks the inputs that change between runs as variables, states what the output must avoid, and tells the model how to choose between options. Prompts that only describe a topic leave all five of those decisions to the model, which is why the output reads like a generic template.

How long should an AI prompt be?

Length is the wrong measure — structure is the right one. A three-line prompt that names the job, the audience and the output format will beat a page of prose that names none of them. Add words only when they remove a decision the model would otherwise guess at.

Do prompt variables work in ChatGPT and Claude?

Variables like {{product}} are a convention for you, not a feature of any model. Every model treats them as literal text, so you replace them before running the prompt. That is exactly what makes them portable: the same template works in ChatGPT, Claude, Gemini or anything else, because nothing model-specific is baked in.

Why does the same prompt give different results each time?

Because language models sample rather than look up answers, and because an under-specified prompt leaves room for that sampling to matter. You cannot remove the randomness, but you can shrink its effect: fix the output format, cap the number of items, and give an explicit ranking rule so the model is choosing within a boundary rather than inventing one.

Should I include examples in my prompt?

Include one when the format is hard to describe but easy to show — a table layout, a specific tone, an unusual structure. One good example is usually worth more than another paragraph of description. Skip examples when you have already specified the format precisely, since they add length without removing ambiguity.

How to write better AI prompts that work more than once

A good AI prompt is reusable. It works today, next week, and next quarter — with new inputs. Here are the seven steps that turn a one-shot request into a prompt template your team can keep using.

1. Start with the job

Write the actual job, not a vague request. “Help me with content” is not a job. “Generate 10 ranked content ideas for our LinkedIn audience this month” is a job. Start with a sentence that names the deliverable.

2. Add context

Name the audience, the situation, the goal, and the constraint. Without context the model picks a plausible-sounding default. With context it stays on-topic for your business.

3. Add variables

Variables are the inputs that change between runs. Wrap them in {{double_braces}} so the same prompt can be reused with a new product, audience, or offer next time.

Product: {{product}}
Audience: {{audience}}
Main objection: {{main_objection}}

4. Add constraints

Say what the output must avoid. Word limits, banned phrases, off-limits topics, format restrictions. Constraints cut the model’s default fluff and force specificity.

5. Add decision criteria

Tell the model how to rank options. “Pick the headline that most directly neutralizes the main objection” is a decision rule. Without one, the model returns a list and lets you do the work.

6. Define the output format

Specify the structure: number of items, sections, fields, table format. If the output format is undefined, every run looks different and the prompt is not reusable.

7. Save the reusable version

Once the prompt works, save it. Give it a name, lock the parts that should never change, and leave the variables open. Next week, run the same prompt with a new product, audience, or offer. Browse reusable AI prompt templates for examples you can copy.

Want a faster path?

Use the free bad prompt checker to grade your prompt against all seven criteria and turn it into a reusable prompt template automatically. For more reading, see bad prompt examples and prompt refinement.

Grade my prompt → Browse reusable templates →

Common questions

What makes an AI prompt good?: A good prompt names the deliverable, supplies the context the model cannot infer, marks the inputs that change between runs as variables, states what the output must avoid, and tells the model how to choose between options. Prompts that only describe a topic leave all five of those decisions to the model, which is why the output reads like a generic template.
How long should an AI prompt be?: Length is the wrong measure — structure is the right one. A three-line prompt that names the job, the audience and the output format will beat a page of prose that names none of them. Add words only when they remove a decision the model would otherwise guess at.
Do prompt variables work in ChatGPT and Claude?: Variables like {{product}} are a convention for you, not a feature of any model. Every model treats them as literal text, so you replace them before running the prompt. That is exactly what makes them portable: the same template works in ChatGPT, Claude, Gemini or anything else, because nothing model-specific is baked in.
Why does the same prompt give different results each time?: Because language models sample rather than look up answers, and because an under-specified prompt leaves room for that sampling to matter. You cannot remove the randomness, but you can shrink its effect: fix the output format, cap the number of items, and give an explicit ranking rule so the model is choosing within a boundary rather than inventing one.
Should I include examples in my prompt?: Include one when the format is hard to describe but easy to show — a table layout, a specific tone, an unusual structure. One good example is usually worth more than another paragraph of description. Skip examples when you have already specified the format precisely, since they add length without removing ambiguity.