The 7 prompt patterns that work everywhere

Why patterns beat prompts

People share prompts the way restaurants share recipes. Copy, paste, eat. Then the model updates and the prompt stops working.

Patterns are different. A pattern is a structure — a way of organising information for an AI model that survives whatever the model does next. Master the patterns and you can write a prompt for any model, any task, any version.

These seven are the ones that show up over and over in real work. They're what Iro AI's Prompt Lab drills.

1. The Role-Goal-Constraints frame

The most reliable opener for any non-trivial task. Four parts:

• Role. Who is the model supposed to be? ("You are a senior litigation associate.")
• Goal. What is the user trying to accomplish? ("Draft a discovery objection.")
• Context. What does the model need to know? (Facts, audience, prior decisions.)
• Constraints. What must the output do, avoid, or fit inside? ("Under 200 words. No case citations. Tone: assertive but cordial.")

Without role and constraints, you get average output. With them, you get specific.

2. Few-shot with anchors

For tasks where style matters more than reasoning — classifications, rewrites, tone-matching — give the model two or three examples that bracket the kind of output you want, then ask for the next one in the same shape.

The trick is anchoring: pick examples that span the range. Don't give three nearly-identical examples. Give a clearly-easy one, a clearly-hard one, and a middle one. The model learns the structure, not the surface.

Few-shot prompting works because large language models are strong in-context learners by design — the behavior OpenAI documented in “Language Models are Few-Shot Learners” (Brown et al., 2020).

3. Output format contract

Specify the output shape up front. Not as a wish — as a contract.

Return ONLY a JSON object with these keys: summary (string, ≤ 40 words), risks (array of objects with name and severity), and next_step (string). No prose outside the JSON.

This is how you stop the model from preambling. The format contract is also what makes a prompt programmable — you can feed the output to the next step without parsing free text.

4. Self-critique loop

Ask the model to grade its own first draft before producing the final answer:

Step 1: draft an answer. Step 2: list three specific weaknesses of the draft. Step 3: produce a second draft that fixes those weaknesses. Return only step 3.

This costs you a few seconds and almost always improves the answer. It works because the model is better at evaluating writing than at writing first drafts.

5. Decomposition

For anything complex, don't ask for the whole answer in one prompt. Ask the model to plan the answer first, then execute the plan in subsequent turns.

This is the basis of agentic workflows: break a goal into named steps, do each step, gate progress on review. It's also what makes long answers actually useful — without decomposition, the model loses the thread halfway through.

Breaking a problem into steps measurably improves accuracy on complex tasks — the core finding behind chain-of-thought prompting (Wei et al., 2022).

6. Comparison prompt

When you can't tell whether an answer is good, ask the model to produce two or three different versions and compare them on explicit criteria you supply.

This is far more reliable than asking for one "good" answer. The model is better at choosing among options than at hitting a single target.

7. Verify-then-trust

The last pattern is a discipline more than a prompt. For any factual claim the model makes — numbers, dates, names, citations — check it before you trust it.

Two cheap checks: paste the claim back and ask the model to find sources; or run the same claim through a grounded tool like Perplexity. Hallucination detection has its own post — read that next.

Putting them together

Most real prompts use three or four patterns at once. A typical work prompt might use Role-Goal-Constraints + Output format + Self-critique. A research prompt might use Decomposition + Comparison + Verify-then-trust.

The skill is knowing which patterns to combine for the task. That's not something you read your way to — you have to drill it. Iro AI's Prompt Lab is the fastest way; the ChatGPT path and Claude path apply the same patterns inside specific tools.