For months, I was frustrated with AI outputs. They were competent but generic. Safe but forgettable. I kept thinking the problem was the AI's capability. "Maybe GPT-5 will be better."
Then I realized: the AI wasn't the bottleneck. Context was.
When I asked "write a presentation," I was essentially asking: "Based on your statistical average of what presentations look like, make one." That's not a recipe for excellence. That's a recipe for a B-minus output that sounds like AI.
The pattern that fixed this is simple: before asking for creation, teach the AI what "great" means for your specific situation. Three prompts, each one building context for the next.
The Pattern
- Prompt 1: "What are the characteristics of a great [thing]?"
- Prompt 2: "Who are the masters of [thing]? What makes them great?"
- Prompt 3: "Create [thing] applying those characteristics and channeling those masters."
That's it. Each prompt builds on the previous one. By the time you ask for creation, the AI's context window contains explicit quality criteria, specific techniques from recognized experts, and a clear target to aim for.
Try It Now
Enter what you're trying to create and get all three prompts ready to copy.
Why This Works (The Mechanisms)
This isn't just a trick that happens to work. There are clear reasons why it outperforms single-shot prompting.
1. It Fights Underspecification
Most disappointing outputs come from vague objectives. "Write a great landing page" forces the model to guess what "great" means. Prompt 1 makes it instantiate a rubric: traits, constraints, tradeoffs, and failure modes.
This aligns with mainstream prompt engineering guidance: provide clear goals, context, and explicit expectations so the model can optimize toward something concrete rather than a statistical average.
2. It Leverages In-Context Learning
Prompt 2 pulls named exemplars. Even if the model doesn't "search," naming specific people and their techniques acts like index keys into learned patterns: structures, rhetorical moves, pacing, and domain-specific constraints.
"Masters" aren't only people. They can be artifacts: great teardown posts, PRDs, memos, Nobel lectures, Stripe press releases. Artifacts reduce ambiguity and bias.
3. It's Lightweight Prompt Chaining
The pattern is a 3-link prompt chain: define → retrieve priors → generate. Each step produces an intermediate object you can inspect before committing to the final output. If your criteria look wrong, fix them before creating.
4. It Sets Up Self-Refinement
Once you have a rubric, you can do: draft → critique against rubric → revise. Research shows iterative self-feedback can improve outputs across tasks without additional training. The rubric you built in Prompt 1 becomes the evaluation criteria for Prompt 4.
Real Examples
Sales Email
Prompt 1: "What are the characteristics of a cold email that actually gets responses?"
Characteristics identified: Personalized first line, one clear ask, under 100 words, creates curiosity, offers value before asking for commitment.
Prompt 2: "Who writes the best cold emails? What specific techniques do they use?"
Masters identified: Ramit Sethi (specificity and social proof), Alex Hormozi (directness and clear value prop), Sam Parr (personality and pattern interrupts).
Prompt 3: "Write a cold email for [my situation] applying Ramit's specificity and Alex's directness. It should have [the characteristics we identified]."
Technical Documentation
Prompt 1: "What makes technical docs that developers actually love?"
Characteristics: Working copy-paste examples, explains why not just how, progressive disclosure (quick start → deep dive), scannable structure.
Prompt 2: "Which companies have the best developer docs? What makes them work?"
Masters: Stripe (example-first, working code in every section), Twilio (step-by-step tutorials), Vercel (clean design, fast navigation).
Prompt 3: "Write docs for my API using Stripe's example-first approach. Hit the characteristics: working examples, explain the why, progressive disclosure."
Strategy Presentation
Prompt 1: "What makes a strategy presentation that gets executive buy-in?"
Characteristics: One clear recommendation, three supporting reasons max, quantified impact, addresses obvious objections upfront, ends with specific ask.
Prompt 2: "Who gives the best strategy presentations? What's their approach?"
Masters: Barbara Minto (pyramid principle), McKinsey (situation-complication-resolution), Andy Raskin (narrative arc with enemy).
Prompt 3: "Outline a strategy presentation for [my situation] using Minto's pyramid structure. Hit the characteristics: one recommendation, three reasons, quantified impact."
Related Techniques
The three-prompt pattern sits at the intersection of several established approaches. Understanding where it fits helps you extend it.
| Technique | What It Does | Connection to Three-Prompt |
|---|---|---|
| Chain-of-Thought | Prompts the model to show intermediate reasoning | Add "think step by step" to any prompt for deeper analysis |
| Meta-Prompting | Asks the model to pick a strategy before executing | Prompt 1 is essentially a targeted meta-prompt |
| Few-Shot Prompting | Provides examples to anchor output style | Prompt 2 surfaces examples; you could provide your own |
| Self-Refine | Generate → critique → revise cycles | Natural extension: Prompt 4 critiques against the rubric |
| Persona Prompting | Assigns an expert role to shape output | Masters are more specific than "act like an expert" |
Limitations (Where This Breaks)
No technique works everywhere. Here's where the three-prompt pattern can fail.
Hallucinated or Biased Masters
LLMs can confidently invent experts, misattribute work, or over-index on Western/Anglophone figures. In niche domains, the model may name people who don't exist or whose work it doesn't actually know well.
Mitigation: Add diversity constraints ("include non-Western practitioners"), ask for specific works ("what book or article demonstrates this?"), and verify the names exist before relying on them.
Shallow Theories of Greatness
When asked "what makes great X," models tend to produce plausible but generic checklists that reflect common narratives more than deep craft knowledge. The rubric can sound right but miss what actually drives quality.
Mitigation: Ask for anti-patterns ("what makes X fail?"), observable signals ("how would I tell if X is good?"), and falsification criteria ("what would prove this rubric wrong?").
Overfitting to Imitation
A masters-based approach can push outputs toward past styles rather than innovation. In creative or strategic contexts where differentiation matters, mimicking experts may sabotage originality.
Mitigation: Ask for "distinctive approaches that differ from the masters" or "where the conventional wisdom is wrong."
Context Decay
In long conversations, the model may lose track of earlier criteria by Prompt 3. The more tokens between the rubric and the creation, the weaker the connection.
Mitigation: Explicitly reference the criteria in Prompt 3. Copy-paste the key characteristics into the creation prompt.
Model-Dependent Effectiveness
Research suggests chain-of-thought and multi-step prompting work best with larger models. With smaller models, the overhead may not pay off.
Mitigation: Test with your specific model. If you're using GPT-4 or Claude, the pattern works well. With smaller models, a single detailed prompt may be better.
Upgrades That Make It Better
The basic pattern works. These variations make it materially more effective.
Turn Prompt 1 Into a Weighted Rubric
Instead of just asking for characteristics, ask for:
- 5-8 dimensions with weights that sum to 100
- Common failure modes / anti-patterns
- 2 mini-examples: one weak, one strong (very short)
- How to test whether the rubric is right
"Define what 'great X' means for [audience + use-case].
Return:
- A rubric with 6-8 criteria + weights (sum to 100)
- 5 common failure modes / anti-patterns
- 2 example snippets showing 'weak vs strong' (keep them short)
- 3 ways this rubric could be wrong + how to test it"Extract Principles, Not Just Names
Instead of just getting names in Prompt 2, extract operational principles:
"For each master, summarize their distinctive approach in 3-5 concrete principles.
What specific moves should I copy?
What should I avoid even though they do it?
In what situations does their approach NOT work?"Add Prompt 4: Self-Critique
After the AI creates, ask it to evaluate its own work:
"Score this draft against the rubric we built:
- Rate each criterion 1-5
- Identify the weakest areas
- Suggest specific revisions
- Produce a second draft that addresses the weaknesses"This maps directly to the self-refinement research. Critique-revise loops measurably improve outputs.
Generate Your Self-Critique Prompt
After your AI creates something, use this to generate a critique-and-revise prompt.
Explore Then Converge
First ask for 3-5 distinct takes (different styles, angles, or approaches), then synthesize:
"Generate 3 different versions of [X], each taking a different approach:
1. Version A: [style/angle]
2. Version B: [style/angle]
3. Version C: [style/angle]
Then analyze what works best from each and synthesize into a final version."This borrows from tree-of-thought prompting: explore multiple paths before converging.
When to Skip This
The three-prompt pattern takes 2-3 minutes instead of 30 seconds. It's not always worth it.
Skip for Quick, Low-Stakes Tasks
"Summarize this article" / "Fix this bug" / "What does this error mean?" Just ask directly.
Skip When You Already Know What You Want
If you can articulate the characteristics and exemplars yourself, skip Prompts 1 and 2. Go straight to a detailed creation prompt.
Skip for Exploration
When brainstorming, you might not want to anchor to existing greatness. Sometimes you want the AI to surprise you.
Skip When Time-Sensitive
Three prompts take longer than one. If you need something in 30 seconds, skip it.
The test: Will spending 2 minutes on context save more than 2 minutes of iteration?
Template You Can Copy
PROMPT 1: CHARACTERISTICS
"What makes a great [THING] for [AUDIENCE]?
Give me:
- 6-8 specific, observable traits
- 5 common failure modes
- How I'd evaluate whether it's working"
PROMPT 2: MASTERS
"Who are the best at [THING]? Include diverse approaches.
For each:
- Their signature moves (3-5 specific techniques)
- When their approach works vs. doesn't
- One thing to copy, one thing to avoid"
PROMPT 3: CREATE
"Create [THING] for [SPECIFIC CONTEXT].
Apply these characteristics: [reference key traits from Prompt 1]
Channel these masters: [reference specific techniques from Prompt 2]
Constraints: [your specific requirements]"
OPTIONAL PROMPT 4: CRITIQUE
"Score this against the rubric from Prompt 1.
Rate each criterion 1-5.
Identify the weakest areas and revise."The Takeaway
Generic prompts produce generic outputs because the AI has no context for what "good" means in your situation. The three-prompt pattern fixes this by explicitly building that context before asking for creation.
It's not magic. It's the same thing a good human collaborator would do: ask clarifying questions, understand the standards, learn from exemplars, then execute.
The pattern works because it treats the AI as capable but uninformed. You're not trying to trick it into being smarter. You're giving it the information it needs to do what it already knows how to do.
Related
For the broader framework behind this technique, see Context Engineering.
This is the method I teach hands-on. In a workshop, your team builds these patterns on its own real work and walks out using them, not just reading about them. See AI training for teams in San Diego (on-site or virtual).