Prompt patterns: 8 reusable shapes that work

Eight prompt shapes that turn up in almost every working AI feature we ship. Each one has a job, a small example, and a clear note on when not to use it. Steal them, adapt them, keep them in a notes app for the next time a model won’t do what you want.

1. Role + audience framing

Tell the model who it is and who it’s talking to. Models adopt the register, vocabulary and assumptions of the role you give them.

You are a senior plumber writing a quote email to a homeowner who is
not technical. Use plain English. Avoid jargon. Quote in pounds.

• Use when: the same content needs different tone for different audiences (homeowner vs trade, executive vs engineer).
• Avoid when:the role is bizarre (“you are a Nobel-winning chef”). Models flatter the role and accuracy can drop.

2. Few-shot examples

Show the model 2-5 input/output pairs. It will copy the shape. The cheapest way to teach a new task without fine-tuning.

Convert each enquiry into a one-line summary.

Enquiry: "Boiler keeps cutting out, pilot light is fine"
Summary: Boiler intermittent shutdown - pilot OK

Enquiry: "Tap drips overnight, kitchen sink"
Summary: Kitchen mixer drip - overnight

Enquiry: "Cold water tank making banging noises"
Summary:

• Use when: the format matters and you have a few real examples handy.
• Avoid when: your examples contradict each other. The model averages them.

3. Chain-of-thought (think before answering)

Ask the model to show its working. Trades short answers for more accurate ones, especially on multi-step problems.

Before giving the final answer, list the steps you'll take. Then
work through each step. Then give the final answer on a new line
prefixed with "ANSWER:".

• Use when: maths, logic, multi-hop reasoning, anything where wrong answers cost you. Pairs well with reasoning models (o-series, Claude with extended thinking).
• Avoid when: high-volume extraction. The thinking eats tokens and slows responses.

4. Self-critique loop

Ask the model to draft an answer, then critique its own draft against criteria, then revise. Quality jumps, sometimes dramatically.

Step 1: draft a reply.
Step 2: critique your draft against these criteria: (a) is the price
clear? (b) is the next step obvious? (c) does it sound like a
person?
Step 3: rewrite incorporating the critique. Show only the rewrite.

• Use when: customer-facing writing, where one extra pass meaningfully improves quality.
• Avoid when: latency matters more than polish.

5. Decomposition

Break a big task into named sub-tasks and ask the model to do them in order. Big tasks tend to fail; small named ones tend to succeed.

Goal: turn the meeting transcript into a weekly client update.

Subtasks (do in order):
1. Pull out decisions made.
2. Pull out action items with owners.
3. Pull out open questions.
4. Compose a 200-word summary using the above.

Output each subtask under its own heading.

• Use when: the task naturally has distinct parts.
• Avoid when:the subtasks aren’t actually independent. You can end up contradicting earlier work.

6. Output schema (constrain to JSON)

Define the exact shape of the reply. Modern models support this natively (Anthropic’s tool-use, OpenAI’s structured outputs, Gemini’s response schemas). When the schema is enforced, the model can’t invent fields.

Extract the customer details from the email below.
Return ONLY a JSON object matching this schema:

{
  "name": string,
  "phone": string | null,
  "postcode": string | null,
  "job_type": "boiler" | "leak" | "blockage" | "other",
  "urgency": "today" | "this_week" | "flexible"
}

If a field isn't present, use null. Do not invent values.

• Use when: the output feeds another system. The single biggest reliability win for production.
• Avoid when: the task is creative writing. Schemas crush voice.

7. Retrieval-augmented (cite your sources)

Don’t let the model rely on its training. Pass the relevant facts in and demand citations. The single biggest anti-hallucination lever.

Answer the question using ONLY the provided context.
If the context doesn't contain the answer, reply: "Not in the
provided documents."

For every claim, cite the source in square brackets like [doc-3].

CONTEXT:
[doc-1] {{first chunk}}
[doc-2] {{second chunk}}
[doc-3] {{third chunk}}

QUESTION: {{user question}}

• Use when: the answer needs to come from your data (policies, contracts, product docs, knowledge base).
• Avoid when: general world knowledge is fine. RAG adds latency and cost.

8. Tool-use prompt

For agents. Give the model a list of tools with clear descriptions, and tell it when to call them vs answer directly. Description quality is the whole game - the model picks tools by reading what you wrote.

You have access to these tools:

- lookup_customer(email): returns customer record or null.
- check_calendar(date): returns available slots for the date.
- send_quote_draft(to, body): saves an email draft. Does NOT send.

Rules:
- Always lookup_customer first.
- Never send_quote_draft until you've confirmed availability.
- If you can answer without a tool, do so.
- Reply to the human in plain English; tool calls are not visible.

• Use when: building any agent.
• Avoid when:a single fixed workflow would do. Agents are overkill for “always do step A then step B”.

The combination cheat sheet

Three rules to write better prompts immediately: tell the model who it is, show it one example, and constrain the output. That alone gets you 80% of the way to a useful result on most tasks.

Where to go next

• Context windows, hallucinations and agents - the three concepts these patterns exist to manage.
• What is an AI agent? - where pattern 8 above actually lives.
• The AI readiness self-assessment - check your foundations before getting fancy with prompts.