There's a cottage industry of people selling "the 10,000 best ChatGPT prompts" in downloadable PDFs for $19. Skip all of it.
Prompts aren't a catalogue you browse. They're a skill you build. And like most skills, 80% of the results come from 20% of the techniques. In this case, three techniques.
Once you know these three patterns, you'll stop staring at a blank text box wondering how to ask. You'll know which pattern fits the task, write the prompt in under two minutes, and get a draft that's actually useful. Every time.
These patterns work across ChatGPT, Claude, and Gemini. They're not platform-specific. They're how to talk to any AI model when you mean business.
---
Pattern 1: Role, Task, Constraints, Format (The Operator Default)
This is your everyday workhorse. Use it for anything you don't have an example of already, and anything where the output format matters.
The structure: tell the AI who it's being, what you need, what the limits are, and what the output should look like.
The structure:
Role: You are [a specific type of expert relevant to the task].
Task: [What you want done, one sentence, specific].
Constraints: [What it must not do, include, or assume].
Format: [Length, structure, tone, be exact].Bad version:
Write a job ad for a carpenter.The output will be a generic job ad that could apply to any carpentry job anywhere. Nothing about your business, your culture, the specific skills you're after, or why someone would actually want to work there.
Good version (real prompt):
Role: You are an experienced HR professional who writes job ads that attract
tradespeople rather than office workers.
Task: Write a job ad for a residential carpenter position at a small Melbourne
renovation company. We do kitchens, bathrooms, and structural work. The role
is full-time. $45-$55/hour depending on experience.
Constraints: Don't use corporate language. Don't list more than 5 requirements.
Don't mention "fast-paced environment" or "team player." Don't promise anything
we haven't confirmed, no mention of specific benefits unless I've listed them.
Format: Three sections, the job in plain English, what we're looking for
(dot points), and how to apply. Total length: under 300 words. Tone: direct
and honest, like a business owner speaking to a tradie.The difference in output is significant. The constrained, specific version produces something you can actually post. The vague version produces something you'd need to rewrite entirely.
Use this pattern as your default. When in doubt, build the four-part structure. The extra 60 seconds in the prompt saves 20 minutes of editing.
---
Pattern 2: Examples-First (The Quality Shortcut)
This is the fastest way to get AI to match a specific style, tone, or format you already like.
Instead of describing what you want, you show it. You give the AI two or three real examples of the kind of output you're after, then ask it to produce a fourth in the same vein.
This works exceptionally well for: social media captions, email subject lines, SMS follow-ups, review responses, product descriptions.
Real prompt (review response template):
Here are three examples of how I respond to Google reviews for my electrical
contracting business:
Example 1:
"Thanks so much, David. Really glad the switchboard upgrade went smoothly,
appreciate you taking the time to write this. We're here if you need anything
down the track."
Example 2:
"Appreciate the kind words, Sarah. That bathroom reno was a good one.
Call us anytime."
Example 3:
"Thanks John, the team loved working on that project. Glad the timeline
held up. Happy to help again whenever you need."
Now write five more review responses in exactly this voice. Keep them under
40 words each. Each should feel like it was written by the business owner,
not a marketing department. Use the reviewer's first name.This is much faster than explaining the voice you want. Showing is almost always better than describing, especially for tone.
Where this pattern breaks down: if your examples are inconsistent (three different tones, different lengths, different formality levels), the model will average them rather than match any one. Before you paste your examples in, make sure they all have the same voice. Pick your best three.
---
Pattern 3: Critique-Then-Rewrite (The Self-Improvement Loop)
This is the highest-impact pattern for getting something from "pretty good" to "actually good."
Instead of accepting the first draft, you ask the AI to critique its own output against your specific criteria, then rewrite it based on that critique.
Why this works: AI models generally produce a reasonable first pass within their default parameters. But their default parameters often favour completeness, hedging, and covering all bases. Your criteria are different. You want specific, brief, honest, and on-brand. Asking the model to critique against YOUR standard often produces a better result than asking it to just "make it better."
Real prompt sequence:
First, get a draft:
Write a short email I can send to a client who hasn't paid their invoice
after 30 days. We're a small landscaping business. The invoice is $2,400.
This is the first follow-up. Tone: professional but direct.Then, after it produces a draft, send a follow-up in the same chat:
Now critique that email against these criteria:
1. Is it under 100 words?
2. Does it state the amount and due date clearly?
3. Does it avoid being passive-aggressive or overly apologetic?
4. Does it include a clear next step?
5. Does it sound like a person, not a collections department?
For each criterion, give a yes/no and one sentence of reasoning. Then rewrite
the email to address any criteria it failed.The self-critique step forces precision. You'll often find that the model agrees its first draft was too long, too hedging, or too formal. The rewrite after critique is almost always significantly better.
This pattern is also useful for copy you've written yourself. Paste in your own draft, give the model your criteria, ask it to critique and rewrite. It's a fast editorial pass when you don't have another person to read your work.
---
The Bad Version vs The Good Version
Here's what each pattern looks like when used poorly versus well, applied to a real operator task: writing an onboarding email for a new client.
Pattern 1 used badly:
Write a welcome email for a new client.Produces: A generic welcome email that reads like an out-of-office reply.
Pattern 1 used well:
Role: You are the owner of a small bookkeeping firm writing to a new client
who just signed up for monthly bookkeeping at $350/month.
Task: Write a welcome email that confirms what happens next and makes them
feel they made the right decision.
Constraints: No corporate speak. Don't promise things I haven't confirmed.
Don't mention software they haven't signed up for.
Format: Under 150 words. Plain text. Friendly, confident tone. End with a
specific next step (e.g. "I'll send through the bank access form by Thursday").Pattern 2 used badly:
Here's one example of my writing style. Match it.
[One example]One example gives the model almost nothing to work with. It'll produce something that takes one trait from the example and runs with it. Three examples is the minimum. Five is better.
Pattern 3 used badly:
Can you improve this?"Improve" means nothing to a language model. It'll make it longer (usually), more formal (often), and add a few synonyms. Tell the model exactly what you want improved, against specific criteria, and it'll actually improve it.
---