How to write better AI prompts: 10 patterns that work for any AI in 2026

The ten patterns

Let’s go through each one with real examples.

1. Role priming

The single highest-use move. Tell the AI who it is. Tone, vocabulary, and perspective all follow.

Bad:

Write me 3 Instagram captions for my Melbourne cafe.

Better:

You’re a senior Australian copywriter who has written social for hospitality brands for 10 years. You write captions that feel like a real person, never corporate, never exclamation-marky. Write me 3 Instagram captions for a Melbourne cafe on Sydney Road Brunswick announcing a winter menu (mushroom toast, hot chocolate, dumpling soup). 30-50 words each.

The role + the specifics + the constraint together gives you something usable on the first try.

2. Front-loaded constraints

State your hard constraints in the first 50 words. AI is more likely to follow rules it sees early.

Bad:

Write me a blog post about AI for accountants. Include real examples. Make it useful. By the way, no longer than 800 words and use Australian English.

Better:

Write a blog post in Australian English. 700-800 words. No em-dashes. Australian dates (DD/MM/YYYY). AUD currency. The post is about AI for accountants in Australia, covering Xero categorisation, BAS sanity checks, and monthly close. Use real examples. Don’t sound like a press release.

Constraints land harder when they come first.

3. Show, don’t tell

This is the move that separates 50% better outputs from 200% better outputs. Show 2-3 examples of what good looks like.

Bad:

Write me a tweet about our new product launch in my brand voice.

Better:

Write me a tweet about our new winter menu launch. Here are 3 examples of past tweets from our account that performed well, mirror this voice:

1. “Found these mushrooms at the market this morning. By 5pm they’ll be on toast. Welcome to mushroom season.”
2. “Six new things on the menu today. One of them is a hot chocolate so good we considered keeping it secret.”
3. “If you’ve ever wanted to start your morning with dumpling soup, today is your day. 7am-2pm, Sydney Road, see you.”

Now write 3 tweets in the same voice about our new winter menu (mushroom toast, hot chocolate, dumpling soup).

The examples teach the AI what your voice actually sounds like. Far more useful than describing it.

4. Force the format

Free-form responses are padded. Structured responses are dense.

Bad:

What should I include in a customer onboarding email?

Better:

Give me a 7-row markdown table for a customer onboarding email. Columns: Section, Goal, Sample copy (one sentence), Estimated time-to-write. Sections include: subject line, opening, value reminder, key first action, social proof, support contact, sign-off.

You’ll get exactly the table. No preamble, no “Here are some great suggestions for your onboarding email!”, just the content you need.

This pattern also works with JSON, CSV, YAML, bulleted lists, numbered lists, and anything else with clear structure.

5. The negative list

Tell the AI what NOT to do, especially if you’ve seen it go wrong before. Negative constraints work shockingly well.

Bad:

Write me a draft email reply.

Better:

Write me a draft email reply. Don’t:

• Start with “Thank you for reaching out”
• Use the word “delighted”
• Use any em-dashes
• Use exclamation marks
• Say “I hope this email finds you well”
• Be more than 5 sentences

Be: direct, warm, Australian, and specific. Reply to this email: [paste]

Negative constraints are how you teach the model your editorial standards in one shot.

6. Chain of thought

For complex reasoning, the four magic words: “Think step by step.”

Bad:

Should I take the $30k AUD per month retainer with monthly hours capped at 25, or the $45k AUD per month no-cap retainer?

Better:

I have two consulting offer options. Think step by step through the trade-offs.

Option A: $30k AUD/month, 25 hours capped, anything over billable at $250/hr. Option B: $45k AUD/month, no hour cap.

Variables to consider: my typical month is 30-35 hours of work for this client; my opportunity cost is roughly $200/hr on other work; my cash flow benefits from monthly predictability; client A pays on net-7, client B on net-30.

Show your reasoning, then give a single recommendation.

Asking the model to reason out loud catches mistakes the model would otherwise commit silently. Use this for anything involving numbers, comparisons, or trade-offs.

7. Scratchpad

For multi-step analysis where you want the AI to plan before answering:

Use a scratchpad first to think through this problem. Then give me your final answer in a separate block.

Problem: [paste]

The model writes its thinking openly, then commits to an answer. You can read both. Catches all sorts of mid-reasoning errors.

8. The audience switch

Sometimes the prompt isn’t wrong; the audience is. Tell the AI who’s reading.

Bad:

Explain MCP (Model Context Protocol).

Better (for a developer audience):

Explain MCP to a senior backend engineer who knows REST and gRPC but hasn’t seen MCP yet. Focus on the wire format, the discovery flow, and the security model. 200 words.

Better (for a small business owner):

Explain MCP to a Melbourne cafe owner who knows what Zapier is but hasn’t used AI tools beyond ChatGPT. Avoid jargon. Focus on what it lets them do. 200 words.

Same topic, three different audiences, three completely different (and all correct) outputs.

9. Self-critique loop

For final polish, run your output back through the model with a critique prompt.

Step 1: Generate the first draft.

Step 2:

Here’s a draft. Critique it as a senior Australian editor. What are the three weakest sentences? What would you change?

[paste draft]

Step 3:

Now apply those critiques and rewrite the draft.

Three prompts. Substantially better output than asking for the polished version up front. Works because the model is better at finding flaws than avoiding them.

10. The “ask if unsure” guardrail

For anything where hallucination is a real risk, give the AI permission to admit it doesn’t know.

Answer the following question. If you don’t have enough information to answer confidently, ask me a specific follow-up question rather than guessing.

Question: [paste]

Single line at the start. Dramatically reduces the model fabricating facts.

The bonus pattern: “use a CLAUDE.md”

If you’re using Claude Code or building a Custom GPT, the highest-use thing you can do is write a configuration file (CLAUDE.md for Claude Code; the “Instructions” field for Custom GPTs) that puts your role + your constraints + your examples + your negative list in there once, permanently. Every future conversation starts pre-loaded.

See our Claude Code install guide for the CLAUDE.md template.

A real example: before and after

We use prompts like the one below to draft customer-facing email replies. Same model (Claude Pro), same task, before-and-after.

Before (slop):

Write me a customer email reply about a delayed order.

Output: “Thank you for reaching out about your order. We sincerely apologise for the delay and appreciate your patience…” (etc, etc, 200 words of corporate boilerplate)

After (usable):

You’re a Melbourne small-business owner replying to a customer email. Tone: warm, honest, Australian. Constraints: no “thank you for reaching out”, no “apologise for any inconvenience”, no exclamation marks, no em-dashes, under 80 words. Be specific about the delay reason if I tell you. Include one concrete next-step.

Their email: “Hi, just wondering where my order is? I ordered the moisturiser last Wednesday and tracking still shows ‘preparing for dispatch’. Was hoping to take it on a trip this weekend.”

Context: We had a printer outage on Thursday that delayed all packing by 48 hours. Their order shipped Monday and is showing in transit now. Replace the delay with a $10 store credit for the inconvenience.

Output: “Hi, sorry for the wait. Our printer went down on Thursday and held everything up by about 48 hours. Your order shipped Monday and is in transit now. I’ve added $10 store credit to your account for the slow start. Tracking should update by tomorrow morning. If it doesn’t, hit reply and I’ll chase. Have a good trip.”

Same model. Different prompt. Reads like a real person wrote it. That’s the entire point.

What’s next

• How to install Claude Code if you want to script repeated prompt workflows.
• Claude vs ChatGPT for Australian small business for picking the right model.
• How to use AI for SEO for a longer applied workflow.

If you want help building production prompts (Custom GPTs, Claude Projects, automated pipelines), our Quick Start build at $497 AUD includes setting up your first one as part of the engagement.