Prompt Patterns Cheat Sheet
Most people use AI like a search engine — type a question, get an answer. These 5 patterns change the dynamic: you're directing an expert, not querying a database. The difference in output quality is immediate and significant.
Pattern 1: Role Assignment
Tagline: Tell Claude who it is before you tell it what to do
When to use: Always. Use this at the start of any prompt that needs expertise.
Template:
You are an expert [role] who specializes in [domain]. Your task is to [goal]. [Additional constraints here.]
Example:
"You are an expert Next.js developer who specializes in building SaaS apps for non-technical founders. Your task is to write a Supabase query that returns all users who signed up in the last 30 days and have not yet completed onboarding. Return only the SQL, no explanation."
Why it works: Claude plays the role you assign. 'Expert Next.js developer' gets you expert output. No role assignment gets you average output. This single change is responsible for 80% of quality improvement for most users.
Pattern 2: Plan Before Build
Tagline: Make Claude think before it codes
When to use: Any task that involves writing more than ~20 lines of code.
Template:
Before writing any code, write out your plan: what files you will create or edit, what each function does, and any edge cases you anticipate. Then write the code.
Example:
"Before writing any code, write out your plan for adding email verification to our Next.js app using Supabase Auth. List: which files you'll create or modify, what each route does, and any edge cases. Then implement it."
Why it works: When you ask Claude to plan first, it catches design problems before they become code problems. Planning takes 10 seconds. Debugging bad architecture takes hours.
Pattern 3: Step-by-Step Decomposition
Tagline: Break big tasks into numbered steps
When to use: Complex multi-file changes, setup tasks, debugging sessions.
Template:
Break this into numbered steps. After each step, pause and confirm before continuing.
Example:
"Break adding Stripe subscription payments into numbered steps. Include: schema changes needed, API routes to create, frontend components, and webhook handling. After listing all steps, implement step 1 only."
Why it works: Big requests produce big walls of code. Decomposed requests produce testable, debuggable chunks. You can stop after any step, verify it works, and continue — which is how real engineers work.
Pattern 4: The Improve Pattern
Tagline: Show Claude what you have, ask what's wrong
When to use: Reviewing code, writing, specs, or any artifact you want to make better.
Template:
Here is my current [thing]. What's wrong with it and how would you improve it? Be specific about what to change and why.
Example:
"Here is my current PRD for the notification system. What's wrong with it and how would you improve it? Be specific — I want to know what's missing, what's vague, and what would confuse an engineer reading this."
Why it works: Asking 'is this good?' gets you a polite 'yes, but...'. Asking 'what's wrong?' gets you a real critique. The Improve Pattern unlocks Claude's critical thinking. Use it on everything before you ship.
Pattern 5: Anti-Obvious Brainstorm
Tagline: Explicitly ask Claude to skip the generic ideas
When to use: Ideation, naming, feature brainstorming, marketing copy.
Template:
Give me [N] ideas for [X]. Avoid the obvious ones everyone else would suggest. I want ideas that are specific, surprising, or counterintuitive.
Example:
"Give me 8 name ideas for a SaaS tool that helps product managers track feature adoption. Avoid generic names like 'ProdMetrics' or 'FeatureFlow.' I want names that feel specific, memorable, and a little unexpected."
Why it works: Left to its own devices, Claude will give you the median response — the average of everything it's seen. Explicitly banning the obvious forces it to explore the edges, which is where the interesting ideas live.
Power Phrases
Short phrases you can add to any prompt to dramatically improve the output.
| Phrase | Effect |
|---|---|
Think step by step. | Forces structured reasoning instead of guessing |
Be specific. | Eliminates vague advice like "handle edge cases" |
Show your work. | Surfaces assumptions you can catch and correct |
What are you uncertain about? | Surfaces the gaps Claude is filling in with assumptions |
Assume I have no engineering background. | Gets you explanations you can actually follow |
Don't write code yet. | Forces planning before implementation |
What would make this worse? | Pre-mortem thinking — catches failure modes early |
Give me the short version. | Stops Claude from writing an essay when you need a sentence |
What am I missing? | Opens the door to things you didn't know to ask about |
Here's what I've already tried: | Prevents Claude from suggesting things you've ruled out |
5-Question Evaluation Framework
Before you copy-paste AI output into your code or share it with a stakeholder, run it through these 5 questions. They catch 90% of the problems.
- Did it answer the question I actually asked, or a slightly different one?
- Is there anything in this response I can't verify or don't understand?
- What assumption is this response making that might be wrong?
- If this advice is wrong, what's the worst that happens?
- What would a skeptical senior engineer say about this?
The meta-pattern behind all of these
Every pattern on this page does the same thing: it gives Claude more context about who you are, what you want, and what you don't want. AI output quality is almost entirely a function of how well you specify the problem. Better prompts don't require technical knowledge — they require clarity about what you actually want.