10 prompt patterns that actually get better answers
Move beyond 'write me a paragraph about…' with patterns we use every day inside Prompt Doctor.
Most weak AI output is not a model problem. It's a prompt problem. The model is doing exactly what you asked, it just wasn't enough to ask. Inside Prompt Doctor we see the same shapes over and over, so we built a small library of patterns the team reaches for first. Here are ten of them, in roughly the order we try them.
1. Role + audience + outcome
Start by telling the model who it is, who it's writing for, and what the deliverable looks like. "You are a senior copywriter writing for time-poor founders. Give me a 90-word LinkedIn post that ends with a question." Three sentences and the output already lifts.
2. Show, don't describe
If you can paste one good example, do. Models match patterns better than they follow adjectives. "Write in this voice" is weaker than "Write in the voice of the example below" with the example pasted underneath.
3. Constraints first
Length, format, tone, what to avoid. Putting constraints up front, in a bulleted list, dramatically reduces the amount of editing you do later. "Max 120 words. No emojis. British English. Don't use the word 'unlock'."
4. Think out loud, then answer
For anything analytical, ask the model to plan before it writes. "First list the three angles you could take. Pick the strongest. Then write the post." The intermediate step is what makes the final answer feel considered.
5. The contrast prompt
Ask for two variants that disagree. "Give me a cautious version and a bold version." You'll almost always pick a hybrid, but you needed the contrast to see it.
6. The critic pass
After the model gives you a draft, send it back: "You are a sceptical editor. List the three weakest lines and rewrite them." This single follow-up is responsible for half of our quality gains.
7. Ground it in your data
Generic prompts give generic outputs. Paste a paragraph from your last newsletter, the brief, the transcript, the product spec, anything specific. Specificity in equals specificity out.
8. The format frame
Tell the model the exact structure: "H2, two-sentence intro, three bullets, closing question." Models love structure and stop hallucinating filler when you give it.
9. The named-step prompt
For multi-stage work, name the steps. "Step 1: extract the three claims. Step 2: rate each claim 1–5 for evidence. Step 3: rewrite the weakest claim." Naming steps makes the model commit to each one.
10. Save what works
The single biggest lift is keeping the prompts that worked. A folder of ten battle-tested prompts beats a clever one you'll never find again. That's the whole reason Prompt Doctor exists.
Pick two of these for your next session. You don't need all ten, you need the two that match the work in front of you.