🎯 Guide

Beginner–Advanced

~2–3h

Prompt Patterns Cheat Sheet

12 Reusable Patterns for Better AI Outputs

Most people write prompts from scratch every time. Experts reuse proven patterns and combine them. This cheat sheet gives you the 12 most useful patterns — with copyable templates, real examples, and guidance on when to use each. For the full theory behind these techniques, see our Prompt Engineering course. To compare which LLMs work best with each pattern, check the LLM Model Family Guide.

Beginner–Advanced

~2–3 hours

Cheat Sheet

TL;DR:

This guide covers 12 prompt patterns organized by difficulty — from Zero-Shot and Few-Shot (beginner) through Chain of Thought and System Prompts (intermediate) to Self-Critique and Meta-Prompting (advanced). Each pattern includes a copyable template, a real-world example, and clear guidance on when to use it and when to skip it. The goal is to give you a reusable toolkit, not a list of tricks.

Guide by

Albert Schaper

Academy review: July 6, 2026

Model names and capabilities change frequently. We use provider families (GPT, Claude, Gemini) in Academy copy — verify current tiers on provider sites or our LLM Model Family Guide.

For live model picks, cross-check with our LLM Comparison guide.

Where this course fits

This course: Prompt pattern reference

Copyable templates and a quick picker — not a full narrative course.

Learn first (we do not repeat this here)

AI FundamentalsRecommendedUnderstand tokens and context before stacking patterns.

Go deeper after this course

Prompt Engineering — Theory, pitfalls, iteration, and advanced techniques.
ChatGPT Mastery — Apply patterns inside ChatGPT projects and custom GPTs.

Course · Updated July 2026

Pattern library refresh

Patterns are stable; placeholders are not. Replace {company}, {audience}, and {tone} every time — and add verification steps for factual tasks.

What we updated for 2026

Add “cite sources” or “say I don’t know” clauses to research patterns.
Code patterns should name file paths and test commands explicitly.
Marketing patterns need brand-voice examples, not adjectives alone.
Pair patterns with Apply It exercises on role guides.

Still true in 2026

Copy-paste without editing produces generic output — customize constraints.

Compare live tools anytime in our tool explorer.

Where this course fits

This course: Prompt patterns

Copy-ready templates with notes — vendor-neutral.

Learn first (we do not repeat this here)

Prompt EngineeringWhy patterns work.

Go deeper after this course

ChatGPT Mastery — OpenAI-specific features.
AI in Practice — Workflow integration.

Quick Reference: All 12 Patterns

Bookmark this table. When you're stuck on a prompt, scan the “When to use” column to find the right pattern.

#	Pattern	Level	When to use
1	Zero-Shot Instruction	Beginner	Simple, well-defined tasks where the model already knows the format (summarize, translate, classify).
2	Few-Shot Examples	Beginner	You need a specific output format, tone, or classification scheme that is hard to describe in words.
3	Chain of Thought (CoT)	Intermediate	Math, logic, multi-step analysis, debugging, or any task where the reasoning path matters.
4	Persona / Role Assignment	Beginner	You want domain-specific language, a particular communication style, or expert-level depth.
5	System Prompt Framing	Intermediate	Building chatbots, assistants, or any multi-turn application where behavior must stay consistent.
6	Output Format Control	Beginner	You need JSON, markdown tables, bullet points, or any structured format for downstream processing.
7	Iterative Refinement	Intermediate	Complex creative tasks, long documents, or when you don't know exactly what you want until you see a draft.
8	Constraint Stacking	Intermediate	You need precise control over length, tone, audience, format, and content simultaneously.
9	Self-Critique / Reflection	Advanced	High-stakes content where quality matters — proposals, code, analysis, or anything going to a client.
10	Task Decomposition	Advanced	Multi-step workflows, research tasks, or anything where a single prompt would produce shallow results.
11	Negative Prompting	Beginner	The model keeps producing unwanted patterns — filler phrases, clichés, disclaimers, or wrong formats.
12	Meta-Prompting	Advanced	You're stuck on how to phrase a complex request, or you want to optimize an existing prompt.

How to Use This Cheat Sheet

Copy & Adapt

Each pattern has a copyable template. Paste it, fill in the blanks, and adjust to your task.

Combine Patterns

The best prompts use 2–4 patterns together. Persona + Output Format + Constraints is a common combo.

Build a Library

Save your best prompts. Over time, you'll build a personal library of proven templates for your domain.

Patterns outlast models and products

Persona, constraints, chain-of-thought, and few-shot examples are not ChatGPT tricks — they are communication structures that work across LLMs, search engines, coding assistants, and future systems. The syntax changes; the pattern does not.

Zero-shot — July 2026 depth note

No examples — just instructions. Works when the task is common and format is flexible.

State role + task + format in three lines.
Add constraints last.
Escalate to few-shot if format drifts.

1. Zero-Shot Instruction

Beginner

Give a clear instruction with no examples. The model relies entirely on its training.

When to use

Simple, well-defined tasks where the model already knows the format (summarize, translate, classify).

When to avoid

When the output format is unusual or the task is ambiguous — the model will guess.

Zero-Shot Instruction — Template:

[Task verb] the following [input type].

[Your input here]

Example

Summarize the following customer review in one sentence.

"I've been using this project management tool for 6 months. The Kanban board is excellent, but the reporting dashboard is slow and the mobile app crashes frequently. Customer support responded within 2 hours though."

Few-shot — July 2026 depth note

Examples teach format and tone faster than adjectives.

Use 2–3 consistent examples.
Match example difficulty to real inputs.
Remove examples that contradict each other.

2. Few-Shot Examples

Beginner

Show 2–5 input→output examples before your actual request. The model learns the pattern.

When to use

You need a specific output format, tone, or classification scheme that is hard to describe in words.

When to avoid

When examples are inconsistent or when the task is so simple that zero-shot works fine.

Few-Shot Examples — Template:

[Task description]

Example 1:
Input: [example input]
Output: [example output]

Example 2:
Input: [example input]
Output: [example output]

Now do the same for:
Input: [your actual input]
Output:

Example

Classify the sentiment of product reviews as Positive, Negative, or Mixed.

Example 1:
Input: "Love the interface, hate the price."
Output: Mixed

Example 2:
Input: "Best purchase I made this year!"
Output: Positive

Now do the same for:
Input: "The AI features are impressive but it keeps crashing on my phone."
Output:

Chain-of-thought — July 2026 depth note

Ask for reasoning steps on multi-hop logic — then verify the conclusion separately.

Request numbered steps.
Add “show uncertainty” for ambiguous cases.
Skip for simple lookups.

3. Chain of Thought (CoT)

Intermediate

Ask the model to reason step by step before giving its final answer.

When to use

Math, logic, multi-step analysis, debugging, or any task where the reasoning path matters.

When to avoid

Simple factual lookups or creative writing where step-by-step reasoning adds noise.

Chain of Thought (CoT) — Template:

[Your question or task]

Think through this step by step:
1. First, identify [relevant factor].
2. Then, analyze [relationship].
3. Finally, conclude [answer].

Show your reasoning before giving the final answer.

Example

A SaaS company has 1,200 customers paying $49/month. They want to raise prices to $59/month but expect 15% churn. Should they do it?

Think through this step by step:
1. Calculate current MRR.
2. Calculate new MRR after price increase and churn.
3. Compare and recommend.

Show your reasoning before giving the final answer.

Persona — July 2026 depth note

Role prompts shape tone; they do not add facts.

Define audience and expertise level.
Pair with citation requirements for factual tasks.
Avoid cartoon personas for client work.

4. Persona / Role Assignment

Beginner

Tell the model who it is. A persona shapes tone, vocabulary, depth, and perspective.

When to use

You want domain-specific language, a particular communication style, or expert-level depth.

When to avoid

When you need factual accuracy more than style — a persona can make the model more confidently wrong.

Persona / Role Assignment — Template:

You are a [role] with [years] of experience in [domain].
Your communication style is [style description].
Your audience is [audience description].

[Your actual task or question]

Example

You are a senior product manager with 10 years of experience in B2B SaaS.
Your communication style is direct, data-driven, and concise.
Your audience is engineering leadership.

Write a one-page brief explaining why we should prioritize the API rate-limiting feature over the dashboard redesign this quarter.

System prompt — July 2026 depth note

System layer = persistent rules; user layer = task instance.

Keep system rules under ~20 bullets.
Version system prompts per product.
Test changes on a fixed eval set.

5. System Prompt Framing

Intermediate

Use the system message to set persistent rules, constraints, and behavior for the entire conversation.

When to use

Building chatbots, assistants, or any multi-turn application where behavior must stay consistent.

When to avoid

One-off tasks where a simple instruction is enough — system prompts add complexity.

System Prompt Framing — Template:

[System message]
You are [role]. You follow these rules:
1. [Rule 1 — e.g., always respond in JSON]
2. [Rule 2 — e.g., never make up information]
3. [Rule 3 — e.g., ask clarifying questions before answering]
4. [Tone/style constraint]

[User message]
[The actual user query]

Example

[System message]
You are a customer support agent for a SaaS analytics platform. You follow these rules:
1. Always be empathetic and professional.
2. Never promise features that don't exist — say "I'll pass this to our product team."
3. If the user describes a bug, ask for their browser, OS, and steps to reproduce.
4. Keep responses under 150 words.

[User message]
The dashboard has been loading for 5 minutes and I have a client presentation in 30 minutes. Help!

Output format — July 2026 depth note

JSON, tables, and schemas reduce parsing pain — especially with tool calling.

Provide a one-line schema or table header row.
Validate machine-readable output in code.
Allow a human-readable fallback field.

6. Output Format Control

Beginner

Explicitly specify the structure, format, and constraints of the output you want.

When to use

You need JSON, markdown tables, bullet points, or any structured format for downstream processing.

When to avoid

When you want the model to be creative about presentation — over-constraining kills creativity.

Output Format Control — Template:

[Your task description]

Output format:
- Format: [JSON / Markdown table / Bullet list / etc.]
- Length: [word count or number of items]
- Include: [required fields]
- Exclude: [things to leave out]
- Tone: [formal / casual / technical]

Example

Analyze these 3 competitor products and compare them.

Output format:
- Format: Markdown table
- Columns: Product Name | Pricing | Key Strength | Key Weakness | Best For
- Length: One row per product, max 15 words per cell
- Tone: Neutral and factual

Iterative refinement — July 2026 depth note

Explicit “v2” instructions beat vague “try again.”

Quote what failed in the prior output.
Change one dimension per iteration.
Stop when checklist passes.

7. Iterative Refinement

Intermediate

Start broad, then refine with follow-up prompts. Each turn narrows the output closer to what you want.

When to use

Complex creative tasks, long documents, or when you don't know exactly what you want until you see a draft.

When to avoid

When you can specify everything upfront — iterating wastes tokens and time.

Iterative Refinement — Template:

Turn 1: [Broad initial request]
Turn 2: "Good, but [specific feedback]. Revise to [direction]."
Turn 3: "Almost. Change [specific element] and [specific element]."
Turn 4: "Final version. Polish for [audience/context]."

Example

Turn 1: "Write a landing page headline for an AI writing tool aimed at marketers."
Turn 2: "Good, but too generic. Make it more specific about saving time on blog posts."
Turn 3: "Almost. Remove the exclamation mark, make it sound confident not excited."
Turn 4: "Final version. Polish for a SaaS landing page — professional, benefit-driven."

Constraint stacking — July 2026 depth note

Multiple constraints work when they don’t conflict — order from hardest to softest.

List must-haves vs. nice-to-haves.
Drop lowest-priority constraint if model stalls.
Use negative prompting sparingly.

8. Constraint Stacking

Intermediate

Layer multiple constraints to narrow the output space. More constraints = more predictable results.

When to use

You need precise control over length, tone, audience, format, and content simultaneously.

When to avoid

When constraints conflict with each other — the model will silently drop some.

Constraint Stacking — Template:

[Task description]

Constraints:
- Audience: [who]
- Tone: [how]
- Length: [how much]
- Must include: [required elements]
- Must avoid: [forbidden elements]
- Format: [structure]

Example

Write a product announcement for our new AI-powered search feature.

Constraints:
- Audience: Existing enterprise customers (technical decision-makers)
- Tone: Professional, confident, not hypey
- Length: 150–200 words
- Must include: One concrete use case, a link placeholder [CTA_LINK]
- Must avoid: Words "revolutionary", "game-changing", "cutting-edge"
- Format: Short paragraph + 3 bullet points + CTA

Self-critique — July 2026 depth note

Second pass for quality — doubles cost; use on high-stakes drafts only.

Ask for rubric-based critique.
Require fixed issues in a third pass.
Skip for low-risk internal notes.

9. Self-Critique / Reflection

Advanced

Ask the model to evaluate its own output, find weaknesses, and improve.

When to use

High-stakes content where quality matters — proposals, code, analysis, or anything going to a client.

When to avoid

Quick tasks where "good enough" is fine — self-critique doubles the token cost.

Self-Critique / Reflection — Template:

[Generate initial output]

Now review your response:
1. What are the 3 weakest points?
2. What assumptions did you make that might be wrong?
3. What would a skeptical expert criticize?

Rewrite an improved version addressing these issues.

Example

Write a technical proposal for migrating our monolith to microservices.

Now review your proposal:
1. What are the 3 weakest points in your argument?
2. What assumptions did you make about our team's capabilities?
3. What would a skeptical CTO criticize?

Rewrite an improved version addressing these issues.

Decomposition — July 2026 depth note

Break complex deliverables into sub-prompts with clear handoffs.

Define subtask outputs as inputs to the next step.
Human-review between stages for client work.
Map to agent plans in advanced setups.

10. Task Decomposition

Advanced

Break a complex task into subtasks, solve each one, then combine. Reduces errors on hard problems.

When to use

Multi-step workflows, research tasks, or anything where a single prompt would produce shallow results.

When to avoid

Simple tasks that don't benefit from decomposition — it adds overhead.

Task Decomposition — Template:

I need to [complex goal]. Let's break this into steps:

Step 1: [First subtask]
Step 2: [Second subtask, building on Step 1]
Step 3: [Third subtask, building on Step 2]
Step 4: [Synthesis / final output]

Start with Step 1.

Example

I need to write a competitive analysis of the top 5 project management tools. Let's break this into steps:

Step 1: List the top 5 tools by market share and briefly describe each.
Step 2: For each tool, identify 3 strengths and 3 weaknesses.
Step 3: Create a comparison matrix across pricing, features, integrations, and user ratings.
Step 4: Write a 200-word executive summary with a recommendation.

Start with Step 1.

Negative prompting — July 2026 depth note

“Don’t” lists help image and style tasks; in text, prefer positive constraints.

Cap at 3–5 negatives for LLM text.
State what to do instead of each don’t.
Heavy negatives for image/video tools.

11. Negative Prompting

Beginner

Tell the model what NOT to do. Negative constraints are often more effective than positive ones.

When to use

The model keeps producing unwanted patterns — filler phrases, clichés, disclaimers, or wrong formats.

When to avoid

Overloading with negatives — too many "don't" instructions can confuse the model.

Negative Prompting — Template:

[Your task]

Important: Do NOT:
- [Unwanted behavior 1]
- [Unwanted behavior 2]
- [Unwanted behavior 3]

Example

Write a LinkedIn post about our Series A funding round.

Important: Do NOT:
- Start with "I'm thrilled to announce" or "Excited to share"
- Use the word "journey"
- Include more than 3 hashtags
- Exceed 200 words
- Use emojis in the first sentence

Meta-prompting — July 2026 depth note

Ask the model to draft or improve the prompt itself — then you edit and run.

Provide goal, audience, and failure examples.
Never run meta-output without human edit.
Great for discovering missing constraints.

12. Meta-Prompting

Advanced

Ask the model to write a better prompt for your task. Use AI to improve your AI instructions.

When to use

You're stuck on how to phrase a complex request, or you want to optimize an existing prompt.

When to avoid

When you already know exactly what you want — meta-prompting adds a round-trip.

Meta-Prompting — Template:

I want to [describe your goal in plain language].

Write me an optimized prompt that I can use with an LLM to achieve this goal. The prompt should:
- Be specific and unambiguous
- Include relevant constraints
- Specify the output format
- Anticipate edge cases

Example

I want to turn customer support tickets into structured bug reports that my engineering team can act on.

Write me an optimized prompt that I can use with an LLM to achieve this goal. The prompt should:
- Be specific and unambiguous
- Include relevant constraints (what fields to extract)
- Specify the output format (JSON)
- Anticipate edge cases (vague tickets, feature requests disguised as bugs)

Pattern Combinations That Work

Real-world prompts rarely use a single pattern. Here are five proven combinations for common tasks.

Business Writing

Persona + Output Format + Constraint Stacking

Set a role (CMO), specify format (email), add constraints (200 words, no jargon).

Code Generation

System Prompt + Few-Shot + Output Format

Set coding rules in system prompt, show 2 examples, specify language and style.

Research & Analysis

Task Decomposition + Chain of Thought + Self-Critique

Break into subtasks, reason step by step, then review for gaps.

Content at Scale

Persona + Few-Shot + Negative Prompting

Set brand voice, show tone examples, ban clichés and filler.

Prompt Optimization

Meta-Prompting + Iterative Refinement

Ask AI to write the prompt, test it, refine based on output quality.

Frequently Asked Questions

Which pattern should I learn first?▾

Start with Zero-Shot Instruction and Few-Shot Examples. They cover 80% of daily use cases. Add Chain of Thought when you need reasoning, and Persona when you need a specific voice.

Can I combine multiple patterns?▾

Absolutely — and you should. Most production prompts combine 2–4 patterns. For example: Persona + Output Format + Constraint Stacking is a very common combination for business writing.

Do these patterns work with all LLMs?▾

Yes. These patterns are model-agnostic. They work with ChatGPT, Claude, Gemini, Llama, Mistral, and others. Some models respond better to certain patterns — experiment with your specific model.

How many examples do I need for Few-Shot?▾

2–5 examples is the sweet spot. More examples improve consistency but cost more tokens. If 2 examples give you the right format, don't add more.

Is Chain of Thought always better?▾

No. CoT helps with reasoning tasks but adds length and cost. For simple tasks (translate, summarize, classify), zero-shot is faster and cheaper.

What's the difference between a system prompt and a persona?▾

A persona is a role assignment ("You are a senior developer"). A system prompt is a broader set of rules that can include a persona plus constraints, output rules, and behavioral guidelines. System prompts persist across turns; personas can be set per-message.

Apply It: Build a 3-Pattern Combo

1Pick a real task from your work this week (e.g., writing an email, analyzing data, generating code).
2Choose 3 patterns from this cheat sheet that fit the task.
3Write a single prompt that combines all 3 patterns. Try with ChatGPT
4Run it. Then try removing one pattern — does the output get worse? That tells you which patterns matter most for this task.

Reflect: Over the next week, try to identify which 3–4 patterns you use most often. Those are your 'core toolkit' — master them deeply before exploring the rest.

Completed

You now have a reusable toolkit of 12 prompt patterns.

Ready to Apply What You Learned?

Prompt Engineering

Go deeper — learn the theory behind why these patterns work.

LLM Comparison

Different models respond differently to patterns — learn which to pick.

ChatGPT Deep Dive

Apply these patterns specifically to ChatGPT workflows.

Recommended Next

AI in Practice

Put these patterns to work in real daily workflows.

Start Learning

Related Courses & Guides

AI Tools for Writers

🎯 Guide

Intermediate

AI Tools for Students

🎯 Guide

Beginner–Intermediate

AI Tools for Marketers

🎯 Guide

Intermediate

Test Your Knowledge

Test your understanding of prompt patterns and when to use them.

Loading quiz...

Key Insights: What You've Learned

Prompt patterns are reusable structures — not magic phrases. The 12 patterns in this guide cover the spectrum from simple instructions (Zero-Shot) to advanced techniques (Meta-Prompting, Self-Critique).

The best prompts combine 2–4 patterns. Persona + Output Format + Constraints is the most common production combination. Chain of Thought + Self-Critique is the most powerful for quality.

Build a personal prompt library. Save your best prompts, tag them by task type, and refine them over time. A good library of 20 prompts will serve you better than memorizing 200 techniques.

Risks & Responsible Use

Know these before you go further.

Pattern Cargo-Culting

Copying a CoT or few-shot template without adapting context often yields verbose or wrong outputs.

What this means for you

Change one variable at a time; keep a small eval set for your domain before calling a pattern “production-ready.”

Secret Leakage in Examples

Few-shot examples frequently contain real customer snippets pasted from tickets or docs.

What this means for you

Sanitize examples; use synthetic data in shared prompt libraries.

Over-Long Prompts

Stuffing every pattern into one mega-prompt burns context and confuses smaller models.

What this means for you

Chain prompts or use retrieval for reference material instead of inline walls of text.

False Precision from JSON Mode

Structured output can be syntactically valid but semantically wrong — especially for numbers and dates.

What this means for you

Validate with schemas plus business rules; never trust JSON alone for financial or medical fields.