Most AI prompts produce mediocre output. The reason is rarely the model. It is the prompt. A poorly structured prompt gets a generic answer no matter how capable the model is behind it.
A well structured prompt is built on patterns. Not magic words, not secret jailbreaks, just a small set of structural choices that consistently produce better output. After analyzing thousands of prompts in the PromptLeadz library and the wider AI ecosystem, twelve patterns show up in almost every prompt that actually works. The rest is variation.
This post breaks down all twelve. Each one has a name, a definition, a bad and good example, and a clear rule for when to use it. Read once and your next prompt gets sharper. Apply all twelve and you stop guessing.
What Is a Prompt Pattern?
A prompt pattern is a repeatable structural choice you make when writing a prompt. It is not the topic, not the question, not the wording. It is the structure underneath the wording.
Prompt patterns are reusable across models. The same twelve patterns work in ChatGPT, Claude, Gemini, Copilot, Mistral, and any model that follows instructions. The patterns also work across tasks. You can write a marketing email, a code review, an SOP, or a legal summary using the same twelve patterns. The content changes. The patterns do not.
A prompt template is a fill in the blanks scaffold. A prompt pattern is the design principle behind the scaffold. Templates are downstream. Patterns are upstream. Learn the patterns and you can write any template.
The 12 Pattern Framework
The twelve patterns split into four groups based on what they control.
The first group controls who the model is. Role Anchoring, Persona Calibration, Audience Targeting.
The second group controls what the model produces. Format Specification, Constraint Stacking, Success Criteria.
The third group controls what the model reasons over. Context Injection, Few Shot Examples, Stepped Reasoning.
The fourth group controls what the model avoids and how it iterates. Negative Specification, Edge Case Coverage, Iteration Hook.
Use all twelve when the stakes are high. Use a subset when the task is light. The list below is in the order you usually write a prompt, not the order of importance.
Pattern 1: Role Anchoring
Role Anchoring tells the model who it is before it answers. The role frames the entire response.
Why it works. Large language models adjust tone, vocabulary, depth, and structure based on the role you assign. A model told to act as a senior tax accountant produces different output than the same model told to act as a college tutor, even on the same question.
Without the pattern: "How do I price a SaaS product?"
With the pattern: "Act as a senior pricing strategist with experience in vertical SaaS. How do I price a SaaS product for the dental practice market with 200 to 2000 monthly users per practice?"
The second prompt gets a strategist's answer. The first gets a Wikipedia summary.
When to use it. Always. Role Anchoring is the highest leverage pattern and costs you one sentence.
Pattern 2: Persona Calibration
Persona Calibration goes deeper than Role Anchoring. It tells the model how to communicate, not just who to be.
Why it works. Two senior pricing strategists can write completely different memos. One is dry and academic. One is sharp and operator focused. Persona Calibration controls that dimension.
Without the pattern: "Act as a senior pricing strategist."
With the pattern: "Act as a senior pricing strategist. Write in a direct, operator focused style. No hedging, no academic disclaimers, no five paragraph essays where one paragraph would work. Use short sentences when the point is sharp and longer ones when nuance is required."
The model produces tighter, more usable prose.
When to use it. Whenever the output will be read by another human and tone matters. Almost always for written deliverables.
Pattern 3: Audience Targeting
Audience Targeting tells the model who the output is for. The audience shapes vocabulary, depth, examples, and what gets explained versus assumed.
Why it works. The same answer rewritten for a CFO, a junior engineer, or a customer support agent produces three completely different documents. Without an audience the model defaults to a generic intermediate level that satisfies nobody.
Without the pattern: "Explain how vector databases work."
With the pattern: "Explain how vector databases work to a startup CFO who needs to evaluate whether to invest in one. Avoid implementation detail. Focus on cost, vendor lock in, and what they enable that a traditional database cannot."
When to use it. Whenever the output has a specific reader. Skip only for personal scratch work where the audience is you and you already know the context.
Pattern 4: Format Specification
Format Specification tells the model exactly how to structure the output. Tables, numbered lists, headers, paragraph counts, word counts.
Why it works. Models are good at following structural instructions and bad at guessing the structure you wanted. If you do not say "output as a three column table," you will get a flowing paragraph that you then have to convert.
Without the pattern: "Compare AWS, Azure, and GCP for a startup."
With the pattern: "Compare AWS, Azure, and GCP for a startup. Output as a markdown table with columns: Criterion, AWS, Azure, GCP, Recommended For. Include these criteria: pricing predictability, free tier generosity, hiring pool, lock in risk, ecosystem maturity. End the table with a one paragraph recommendation."
When to use it. Whenever you will paste, share, or process the output. Skip only for true brainstorming.
Pattern 5: Constraint Stacking
Constraint Stacking layers explicit limits on the output. Length, tone, scope, what to include, what to exclude.
Why it works. Models default to longer, hedged, more inclusive output because that is what training rewards. Constraints pull the output back to what you actually need.
Without the pattern: "Write a product launch email."
With the pattern: "Write a product launch email. Maximum 120 words. Three short paragraphs. No subject line. No greeting. Plain language no marketing adjectives. Include exactly one call to action at the end."
The output becomes immediately usable.
When to use it. Whenever the output has a target length or strict format. Especially valuable for short form output where every word counts.
Pattern 6: Success Criteria
Success Criteria tells the model how to evaluate its own output. The criteria become a self check before the answer is finished.
Why it works. When the model knows what good looks like, it produces output closer to good. Self evaluation is a real mechanism that consistently improves quality, especially for longer outputs.
Without the pattern: "Write a job description for a Head of Operations."
With the pattern: "Write a job description for a Head of Operations. The output is good if: it explains the company stage and why this role matters now, it lists 5 to 7 specific responsibilities (not generic), it includes 3 to 5 traits we are screening for, and it ends with a clear next step for the applicant. Before finalizing, check the output against these criteria and revise."
When to use it. For high stakes outputs where quality matters more than speed. Adds 10 to 15 seconds of model thinking and meaningfully improves the result.
Pattern 7: Context Injection
Context Injection feeds the model the source material it needs to reason over. Without it the model guesses. With it the model grounds its output in your reality.
Why it works. Most prompt failures come from missing context, not bad instructions. If you want the model to summarize a meeting, paste the transcript. If you want it to analyze a contract, paste the contract. If you want a product positioning, paste the existing product page.
Without the pattern: "Help me improve my landing page."
With the pattern: "Help me improve my landing page copy. Here is the current copy: [paste]. Here is the audience: [describe]. Here is the conversion goal: [state]. Suggest three specific rewrites for the hero section, with rationale for each."
When to use it. Whenever the task depends on specific source material. Most operations, finance, marketing, and analysis tasks need this.
Pattern 8: Few Shot Examples
Few Shot Examples show the model two or three completed examples before asking for a new one. The examples teach the pattern faster than instructions can describe it.
Why it works. Some patterns are easier to demonstrate than describe. Tone, voice, formatting quirks, and structural choices that vary by team are often invisible to instructions but obvious from examples.
Without the pattern: "Write a LinkedIn post in my voice."
With the pattern: "Write a LinkedIn post in my voice. Here are three previous posts that performed well, for reference:
Example 1: [paste]
Example 2: [paste]
Example 3: [paste]
Now write a new post about [topic] in the same voice and structure."
When to use it. Whenever the output has stylistic conventions you cannot fully describe. Especially valuable for voice matching, recurring formats, and brand specific output.
Pattern 9: Stepped Reasoning
Stepped Reasoning asks the model to walk through its thinking before delivering the final answer. The thinking improves the answer.
Why it works. Models that show their work catch their own mistakes. Asking for a structured reasoning trace before the final output reliably produces better answers on analytical tasks, planning tasks, and anything that requires comparison or tradeoff.
Without the pattern: "Which of these three vendors should we pick?"
With the pattern: "Which of these three vendors should we pick? First, list the criteria that matter for this decision and weight them. Second, score each vendor against each criterion with one sentence of reasoning. Third, calculate the weighted total. Fourth, write a one paragraph recommendation. Show all four steps."
When to use it. For decisions, comparisons, root cause analysis, planning, and anything that requires multi step reasoning. Skip for short factual lookups where the steps would be padding.
Pattern 10: Negative Specification
Negative Specification tells the model what not to do. Often more useful than positive instructions.
Why it works. Models default to common failure modes. Marketing copy defaults to adjectives. Job descriptions default to corporate cliches. Technical writing defaults to hedging. Telling the model what to avoid directly attacks those defaults.
Without the pattern: "Write a product page for our software."
With the pattern: "Write a product page for our software. Do not use the words seamless, innovative, robust, cutting edge, leverage, empower, or unlock. Do not write more than 200 words. Do not include any feature the user did not provide."
When to use it. Whenever you can name specific failure modes you have seen the model produce. Especially valuable in marketing, writing, and tone sensitive output.
Pattern 11: Edge Case Coverage
Edge Case Coverage tells the model how to handle situations the main instructions did not anticipate.
Why it works. Real input is messy. Data has gaps, sentences are ambiguous, instructions are incomplete. If the model is not told how to handle edge cases, it makes a silent guess. Edge Case Coverage makes the handling explicit.
Without the pattern: "Categorize each customer support ticket by issue type."
With the pattern: "Categorize each customer support ticket by issue type. Use these categories: Billing, Bug, Feature Request, How To, Other. If a ticket fits more than one category, pick the primary category and add a secondary in parentheses. If a ticket is unclear, mark it Unclear and explain in one sentence what additional information would resolve it. Do not invent new categories."
When to use it. For repetitive tasks, classification, extraction, and any prompt that will be applied to many inputs. Saves you from cleaning up garbage on the back end.
Pattern 12: Iteration Hook
Iteration Hook ends the prompt with an explicit invitation to refine. The first output becomes a draft, not the final answer.
Why it works. Most prompts produce better output on the second or third pass. The Iteration Hook reminds you to iterate and gives the model the structure to handle revisions cleanly.
Without the pattern: "Write a 200 word company bio."
With the pattern: "Write a 200 word company bio. After you produce the draft, list three specific questions you would ask to make the second draft sharper, and wait for my answers before revising."
The model produces a draft and a list of clarifying questions. You answer the questions. The second draft is materially better.
When to use it. For any prompt you would normally need to run two or three times. The Iteration Hook builds the loop into the first run.
How to Combine the Patterns
You do not need all twelve patterns in every prompt. Most working prompts use four to seven.
The minimum viable prompt for serious work uses Role Anchoring, Format Specification, Constraint Stacking, and Context Injection. Those four cover who the model is, how it responds, what limits to respect, and what to reason over. That is enough for 70 percent of real work.
When the stakes are higher, add Audience Targeting, Success Criteria, and Negative Specification. Now you have specified who the output is for, what good looks like, and what to avoid. That covers another 20 percent of cases.
When the task is analytical or involves judgement, add Stepped Reasoning and Few Shot Examples. When the task will run many times, add Edge Case Coverage. When the first pass will not be the final pass, add Iteration Hook. Persona Calibration goes on whenever tone matters.
The order is consistent. Role first, then audience, then format and constraints, then context and examples, then reasoning instructions, then negative and edge case rules, then the iteration hook at the end. Following the order makes prompts easier to read, easier to debug, and easier to reuse.
Once you can write good prompts from patterns, the next move is composing them into prompt stacks that chain multiple prompts together to produce complete artifacts instead of pieces.
Frequently Asked Questions
What is the difference between a prompt template and a prompt pattern?
A prompt template is a fill in the blanks scaffold with specific wording. A prompt pattern is the underlying structural choice the template uses. Patterns are reusable across templates. Templates are reusable across users. Patterns are the upstream concept that lets you build any template from scratch.
Do these patterns work for ChatGPT, Claude, and Gemini equally?
Yes. All twelve patterns work in any modern instruction following model. The patterns target how the model parses your request, not any model specific quirk. Output quality may vary by model and task, but the structural improvement from applying the patterns is consistent across the major models.
Can I combine patterns in one prompt?
Yes, and you should. Most serious prompts combine four to seven patterns. Combining them is additive, not conflicting. Each pattern controls a different dimension of the output.
Do these patterns work for image prompts?
Partially. Role Anchoring, Constraint Stacking, Audience Targeting, Negative Specification, and Few Shot Examples translate directly. Format Specification and Stepped Reasoning have weaker analogs in image generation. The remaining patterns are mostly text specific.
Is this prompt engineering or prompt design?
Both terms describe the same skill. Prompt engineering is the more common name in technical circles. Prompt design is preferred in marketing and writing contexts. The twelve patterns apply under either name.
How long should an effective prompt be?
There is no fixed length. Short tasks need short prompts (two or three patterns, 30 to 60 words). Complex tasks need longer prompts (seven to twelve patterns, 200 to 500 words). The number of words is a side effect of how many patterns the task needs.
Do I need to use the exact wording in the examples?
No. The examples show the pattern. The wording can be anything that expresses the same structural choice. Once you understand the pattern, you can write it in your own voice.
Get the Full Prompt Pattern Library
The 12 patterns above are the framework. Building prompts from patterns rather than from scratch is the number one habit that marks a pro AI user. The PromptLeadz prompt library applies the framework across 50 plus role packs and 1000 plus prompts, every one structured around the 12 patterns and formatted three ways for Claude, ChatGPT, and Gemini. Browse the role packs to find prompts already calibrated to your function, or start with the Freebie Vault for free starter prompts in every role.
コメントを残す: