Five Prompt Patterns That Quietly Outperform Everything Else

After a year of A/B-testing prompts against each other, five patterns keep winning. None of them are exotic. All of them get used 10× less than they should.

1. Role priming

Open with who the model should be, not what it should do. "You are a senior copy chief at a B2B SaaS company" outperforms a generic "write me copy" by a noticeable margin because it pulls in a whole cluster of vocabulary, taste and rules.

Template: "You are a [role] at [context]. Your job is to [task]. You care about [3 specific values]."

2. Constraints-first

Lead with the limits before the task. Word count, audience, tone, things to avoid. Models obey constraints better when they appear before the request, not after.

Template: "Constraints: under 120 words, no jargon, no em-dashes, second person. Now write [task]."

3. Worked example (one-shot)

A single high-quality example beats five paragraphs of instruction. Show one input → output pair, then ask for the next one.

Template: "Here is an example. Input: [X]. Output: [Y]. Now do the same for: [Z]."

4. Chain-of-thought, but bounded

"Think step by step" is overrated. "Think step by step, but only write the final answer" is what you actually want for most tasks — you get the reasoning quality without the wall of text.

5. Rate then revise

Have the model produce a draft, score it 1-10 on three named criteria, then rewrite to fix the lowest score. One extra round trip, dramatically better output.

Template: "Draft an answer. Then rate it 1-10 on clarity, specificity and originality. Identify the lowest score. Rewrite to fix it."

Mix and match

The strongest prompts stack two or three of these. Role + constraints + worked example is the workhorse combination — try it next time and see.