How to Get Better Results From AI
Seven principles that improve AI output by up to 20%. Not by changing the model. By changing how you ask.
A Practical Guide Based on OpenAI’s Prompting Research
Seven principles that improve AI output by up to 20%. Not by changing the model. By changing how you ask.
Introduction
In 2026, OpenAI published a prompting guide for developers. It was based on extensive internal testing and produced a 20% improvement in AI performance. Not by changing the model. By changing the instructions.
The guide was written for people building software. It talks about system prompts, tool-calling, and agentic workflows. But the principles underneath are universal. They apply to anyone who has ever typed a prompt into ChatGPT, Claude, Gemini, or any other AI tool and been disappointed with the result.
The central insight is simple: the model does what you tell it to do, not what you want it to do. The gap between those two things is the gap between a bad prompt and a good one.
These seven principles close that gap.
Principle 1: Define “Done”
What it is: Tell the AI what a complete answer looks like before it starts.
Why it works: AI defaults to the minimum viable response. It gives you one answer and stops. It doesn’t know whether you wanted a paragraph or a dissertation, a list of three options or ten, a surface-level overview or a deep analysis. If you don’t define “done,” the model defines it for you, and its definition is almost always shorter than yours.
The psychology: This is the AI equivalent of delegation. If you tell a new hire “write me a report,” you’ll get something. If you tell them “write me a five-page report covering X, Y, and Z, with data in section 2 and recommendations in section 4, and don’t send it to me until you’ve addressed the counterargument,” you’ll get something useful. The model is the new hire. It’s capable, but it needs the brief.
Examples
Example 1: Content creation
Don’t: “Write me a LinkedIn post about AI in marketing.”
Do: “Write me a LinkedIn post about AI in marketing. 150-200 words. Open with a provocative question. Include one specific data point. End with a call to action that invites comments. Don’t use hashtags. Don’t use emojis. The tone should be confident but not arrogant, like someone who’s done this for 20 years and doesn’t need to prove it.”
Why the second one works: The model now knows the length, the structure, the tone, the content constraints, and what to avoid. The first prompt could produce anything from a single sentence to a 500-word essay, from a corporate tone to a casual one. The second prompt has narrowed the possibility space to something useful.
Example 2: Analysis
Don’t: “Analyse our competitor landscape.”
Do: “Analyse our competitor landscape. Identify the top five competitors. For each, cover their positioning, pricing, target audience, and one thing they do better than us. Then identify the gap that none of them are filling. Format as a table with a summary paragraph at the end.”
Why the second one works: The model knows how many competitors to cover, what dimensions to analyse, what the output format should be, and what the deliverable looks like. The first prompt might give you a paragraph about two competitors. The second gives you a structured, actionable analysis.
Example 3: Decision support
Don’t: “Should we launch in Germany or France first?”
Do: “I’m deciding whether to launch in Germany or France first. Evaluate both markets against five criteria: market size, regulatory complexity, competitive density, existing brand awareness, and logistics cost. Score each criterion 1-5 for both countries. Show me the scoring in a table. Then give me your recommendation with a one-paragraph justification. Don’t recommend until you’ve scored all ten cells.”
Why the second one works: The model has a framework, a scoring system, a format, and a completion condition. It can’t give you a lazy answer because the structure won’t let it. The first prompt gets you an opinion. The second gets you an analysis.
Useful phrases:
- “Keep going until you’ve covered all of the above.”
- “Give me five options, not one.”
- “Don’t stop at the surface answer. Go deeper.”
- “I want the full version, not a summary.”
- “Only stop when you’re confident the answer is complete.”
- “Cover every angle, not just the obvious ones.”
- “I need this to be comprehensive, not concise.”
Principle 2: Don’t Let It Guess
What it is: Tell the AI to flag uncertainty instead of inventing an answer.
Why it works: AI is designed to produce fluent, confident text. It will fill gaps in its knowledge with plausible-sounding fabrications. It doesn’t distinguish between “I know this” and “this sounds right.” You have to tell it to.
The psychology: This is the most important principle for anyone using AI for research, analysis, or factual work. The model is a pattern-completion engine. When it encounters a gap in its training data, it doesn’t stop and say “I don’t know.” It completes the pattern with the most statistically likely answer. Sometimes that answer is correct. Sometimes it’s invented. The model doesn’t know the difference. You have to instruct it to care.
Examples
Example 1: Research
Don’t: “What are the latest NPS scores for the UK student accommodation sector?”
Do: “What are the latest NPS scores for the UK student accommodation sector? I need actual data, not estimates. If you don’t have specific numbers, tell me you don’t have them and suggest where I might find them. Do not approximate or extrapolate.”
Why the second one works: The model can’t hide behind vague language. It either has the data or it doesn’t. The instruction “suggest where I might find them” also turns a limitation into a useful output.
Example 2: Fact-checking
Don’t: “Is it true that 70% of students choose accommodation based on social media?”
Do: “Is it true that 70% of students choose accommodation based on social media? I’ve seen this figure cited but I can’t find the original source. If you know the source, tell me. If you don’t, say so clearly and tell me whether the claim is plausible based on what you do know.”
Why the second one works: The model is being asked to verify, not assert. The instruction “tell me whether the claim is plausible” gives it a useful output path even when it can’t confirm the specific number.
Example 3: Strategic advice
Don’t: “What’s the best pricing strategy for student accommodation?”
Do: “What are the three most common pricing strategies for purpose-built student accommodation? For each, explain when it works best and when it fails. If you’re drawing on general business pricing theory rather than PBSA-specific knowledge, say so.”
Why the second one works: The model is forced to reveal its sources. “If you’re drawing on general business pricing theory rather than PBSA-specific knowledge, say so” prevents it from presenting generic advice as sector-specific expertise.
Example 4: Legal and compliance
Don’t: “Can I use student data for targeted marketing under GDPR?”
Do: “I need to understand GDPR constraints on using student data for targeted marketing. Don’t give me a definitive legal answer, because you’re not a lawyer. Instead, outline the key considerations, flag the areas of risk, and tell me what questions I should be asking our legal team. Be explicit about where you’re uncertain.”
Why the second one works: The model is instructed not to overstate its authority. The output becomes a preparation tool for a real legal conversation rather than a risky substitute for one.
Useful phrases:
- “If you’re not sure, tell me you’re not sure.”
- “Don’t make up statistics. If you don’t have the data, say so.”
- “Flag any claims you can’t verify.”
- “Distinguish between what you know and what you’re inferring.”
- “If this requires speculation, label it as speculation.”
- “Tell me your confidence level for each claim.”
- “If you’re pattern-matching rather than recalling, say so.”
Principle 3: Plan Before Acting
What it is: Ask the AI to outline its approach before it starts the work.
Why it works: This forces the model out of pattern-matching mode and into reasoning mode. The difference between a mediocre output and a great one is usually whether the model was given space to plan first. It also lets you redirect before you’ve wasted a full output cycle.
The psychology: Without a planning instruction, the model starts generating text immediately, token by token, without a roadmap. The result often starts strong and then meanders, because the model is making decisions about structure in real time rather than in advance. Planning forces the model to make those decisions upfront, which produces more coherent, more logically structured output.
Examples
Example 1: Writing
Don’t: “Write me an article about AI anxiety.”
Do: “I want to write an article about AI anxiety. Before you write anything, give me: (1) the core argument in one sentence, (2) the structure with section headers and one-line descriptions, (3) the opening hook, and (4) the closing line. I’ll review the plan before you write.”
Why the second one works: You get to see the bones before the flesh. If the argument is wrong, you fix it before 1,200 words have been written. If the structure is weak, you reorganise before the prose exists. This saves time and produces better output.
Example 2: Problem-solving
Don’t: “How do we reduce student churn in the first term?”
Do: “I want to reduce student churn in the first term. Before you give me solutions, do this: (1) identify the five most common reasons students leave in the first term, (2) for each reason, tell me whether it’s something we can influence or something we can’t, (3) rank the influenceable reasons by likely impact, (4) then give me three specific interventions for the top two reasons. Show me the reasoning at each step.”
Why the second one works: The model can’t skip to a generic solution. It has to diagnose before it prescribes. The step-by-step structure also makes it visible where the model’s reasoning is strong and where it’s weak.
Example 3: Strategy
Don’t: “Help me build a marketing strategy for our new property.”
Do: “I’m building a marketing strategy for a new PBSA property in Manchester. Before you write the strategy, ask me five questions that would help you do a better job. Don’t write the strategy until I’ve answered them.”
Why the second one works: This reverses the flow. Instead of the model guessing what you need, it asks. The questions themselves are often useful, because they reveal assumptions you hadn’t considered. The final strategy is better because it’s built on your actual context, not the model’s assumptions.
Example 4: Presentations
Don: “Help me prepare for a board presentation on our AI strategy.”
Do: “I’m presenting our AI strategy to the board on Thursday. Before you help me with slides or talking points, outline what you think the board’s five biggest concerns will be. Then I’ll tell you which ones are right and which ones you’ve missed. Then we’ll build the presentation to address them.”
Why the second one works: The model is doing preparation work, not final work. The concerns list becomes a diagnostic tool. You correct it, and the model uses the corrected version to build something that actually addresses the real audience.
Useful phrases:
- “Outline your approach before you start.”
- “Walk me through your reasoning first.”
- “What’s your plan before you begin?”
- “Think step by step.”
- “Give me the structure, then I’ll decide if you should write it.”
- “Explain your logic before you give me the conclusion.”
- “Ask me questions before you start.”
- “What would you need to know to do this well?”
Principle 4: Be Specific
What it is: Remove ambiguity from your prompts. State exactly what you want.
Why it works: Modern AI models follow instructions literally. They don’t infer intent the way earlier models did. If your prompt is ambiguous, the model picks one interpretation and commits. If that interpretation is wrong, the output is wrong, and the model had no way of knowing.
The psychology: The model is a literalist. It does not read between the lines. It does not understand that “make it better” means “fix the three specific problems I can see but haven’t described.” It takes the word “better” and interprets it in the most common way its training data suggests, which is usually “use more sophisticated vocabulary” or “add more detail.” That’s almost never what you meant.
Examples
Example 1: Editing
Don’t: “Make it better.”
Do: “Make three changes: (1) cut the first paragraph entirely, it’s throat-clearing, (2) the section on pricing needs data, add a comparison table, (3) the conclusion is weak, end with a specific recommendation instead of a general observation.”
Why the second one works: The model knows exactly what “better” means. Each change is specific, actionable, and verifiable. You can check whether it was done.
Example 2: Tone
Don’t: “Make it more professional.”
Do: “Rewrite this for a CEO audience. Remove contractions. Shorten the sentences. Cut anything that sounds like marketing copy. If a sentence could appear on a billboard, delete it. The tone should be: I’ve done the analysis, here’s what it says, here’s what we should do.”
Why the second one works: “Professional” means different things in different contexts. A law firm’s version of professional is different from a tech startup’s. The second prompt defines professional by what it’s not (marketing copy, billboard language) and what it is (analytical, direct, evidence-based).
Example 3: Length
Don: “Keep it short.”
Do: “Target 500 words. If you’re over 600, cut. Every sentence must earn its place. If a sentence repeats something already said, delete it.”
Why the second one works: “Short” is relative. 500 words is a number. The model can count. “Every sentence must earn its place” is a principle the model can apply. “If a sentence repeats something already said, delete it” is a rule that enforces concision.
Example 4: Format
Don’t: “Summarise this for me.”
Do: “Summarise this in exactly three bullet points. Each bullet should be one sentence. The first bullet should cover what happened. The second should cover why it matters. The third should cover what we should do next.”
Why the second one works: The model knows the structure, the length, and the purpose of each section. “Summarise” alone could produce anything from a paragraph to a table to a one-liner.
Example 5: Audience
Don’t: “Write this for a general audience.”
Do: “Write this for a 25-year-old who has never heard of STAR, has no background in psychology, and is reading this on their phone during a commute. Short sentences. No jargon. If you use a technical term, define it immediately. Assume they’ll stop reading if they’re bored in the first 30 seconds.”
Why the second one works: “General audience” is an abstraction. A 25-year-old on a phone is a person. The model can write for a person. It can’t write for an abstraction.
Useful phrases:
- “Cut every sentence that doesn’t advance the argument.”
- “Target [X] words. If you’re over [Y], cut.”
- “The tone should be [specific description].”
- “Write for [specific person/role/context].”
- “Remove all [specific thing: jargon, bullet points, questions, etc.].”
- “Every [paragraph/sentence/section] must [specific requirement].”
Principle 5: Show, Don’t Tell
What it is: Paste an example of the output you want.
Why it works: The model can match a pattern far more reliably than it can interpret an abstract description. “Write like this” is the most precise instruction you can give.
The psychology: Describing tone in words is like describing a colour to someone who’s never seen it. “Write in an engaging style” means nothing concrete. But paste a paragraph that you find engaging, and the model can extract the sentence rhythm, the vocabulary level, the pacing, the use of concrete vs. abstract language, and a dozen other features that you couldn’t have articulated. Pattern matching is what the model does best. Let it match.
Examples
Example 1: Matching a style
Don’t: “Write like a journalist.”
Do: “Write in this style:
‘The meeting lasted three hours. By the end, nobody had agreed on anything except that the coffee was terrible. The CFO had left twice, once for a call he claimed was urgent and once, apparently, just to breathe. The strategy document that was supposed to be finalised by Friday sat untouched on the table, its margins filling with increasingly desperate annotations.’
Match the sentence length, the dry tone, the specific detail, and the way it shows rather than tells.”
Why the second one works: The model now has a concrete reference. It can see the short sentences, the observational detail, the absence of adjectives, the deadpan humour. “Write like a journalist” could mean a hundred different things. The example means one thing.
Example 2: Replicating a format
Don: “Write this like a Substack post.”
Do: “Write this in the format of this Substack post:
[Title] [Subtitle in italics]
Opening paragraph that hooks with a specific observation, not a generalisation.
Section header
Body text that mixes short declarative sentences with longer analytical ones. Every section ends with a line that makes the reader want to keep going.
Section header
Same pattern.
[Sign-off paragraph in italics]”
Why the second one works: The format is explicit. The model can see the structure, the rhythm, and the conventions. It will replicate them.
Example 3: Getting consistent output
Don’t: “Write me five email subject lines.”
Do: “Write me five email subject lines in this style:
‘Students are choosing based on vibes. Here’s the data.’ ‘Your competitor’s NPS just dropped. Here’s why.’ ‘The one metric your board isn’t looking at.’
Short. Specific. Slightly provocative. End with a period, not an exclamation mark.”
Why the second one works: The model has three examples and two rules. The output will be consistent with the examples because that’s what pattern matching produces.
Example 4: Communicating what you can’t describe
Don’t: “Write with more personality.”
Do: “Write like this:
‘I spent my twenties teaching people not to die at forty thousand feet. Turns out the skills transfer. Consumer psychology is just a cockpit where the instruments are feelings and the stall warning is a bounce rate.’
That’s the level of personality I want. Direct. Slightly unexpected. A little self-deprecating. Never trying too hard.”
Why the second one works: “Personality” is the vaguest instruction in existence. The example makes it concrete. The three adjectives at the end (“direct, slightly unexpected, a little self-deprecating”) give the model additional calibration.
Useful phrases:
- “Write it like this: [example]”
- “Match this format: [example]”
- “Use this as a reference for tone: [example]”
- “Here’s an example of what I mean: [example]”
- “This is the standard I’m looking for: [example]”
- “Match the sentence length, vocabulary, and rhythm of this: [example]”
Principle 6: Ask “Why” Before “What”
What it is: Request the reasoning before the conclusion.
Why it works: Two reasons. First, the model produces better conclusions when it has to build the argument first. Second, you can evaluate the reasoning before committing to the answer. If you only see the conclusion, you have no way of knowing whether it’s built on solid ground.
The psychology: Without a reasoning instruction, the model goes straight to the answer. This feels efficient but is actually risky, because you can’t see the logic that produced it. If the logic is flawed, the answer is flawed, and you won’t know until you act on it. Asking for reasoning first is like asking a accountant to show their calculations before giving you the final number. It’s slower, but it’s trustworthy.
Examples
Example 1: Decision-making
Don: “Which pricing model should we use?”
Do: “We’re choosing between three pricing models: fixed, tiered, and usage-based. For each model, analyse: (1) how it affects revenue predictability, (2) how it affects customer perception, (3) how it compares to competitors, (4) what the operational complexity is. Show me the analysis in a table. Then give me your recommendation, and tell me what the strongest argument against your recommendation would be.”
Why the second one works: The model can’t just pick one and justify it after the fact. It has to evaluate all three against consistent criteria. The “strongest argument against” instruction prevents the model from presenting its recommendation as the only reasonable option.
Example 2: Analysis
Don’t: “Is our marketing strategy working?”
Do: “Here’s our marketing data from Q1. Before you tell me whether it’s working, do this: (1) identify the three metrics that matter most for our business model, (2) calculate the trend for each, (3) compare them to industry benchmarks if you have them (if you don’t, say so), (4) then tell me what’s working, what’s not, and what you’d change. Show your work at each step.”
Why the second one works: The model has to demonstrate the reasoning, not just assert the conclusion. If it miscalculates a trend or uses a wrong benchmark, you can catch it before you act on the recommendation.
Example 3: Creative work
Don: “Write me an opening paragraph for a keynote speech.”
Do: “Write me three opening paragraphs for a keynote speech about AI anxiety. For each one, explain: (1) what psychological principle it’s using, (2) why it would work for this audience, (3) what the risk is. Then tell me which one you’d recommend and why.”
Why the second one works: The model isn’t just generating, it’s evaluating. The explanation of psychological principle forces it to draw on actual knowledge rather than vibes. The risk assessment prevents it from presenting all three as equally good.
Example 4: Recommendations
Don: “What should we do about declining student satisfaction?”
Do: “Our student satisfaction scores dropped 8 points in Q3. Before you recommend solutions, walk me through: (1) what the most likely causes are, (2) which causes we can influence and which we can’t, (3) for the influenceable causes, what the evidence says about effective interventions, (4) then give me your top three recommendations, ranked by likely impact and ease of implementation. Show the reasoning for each.”
Why the second one works: The model has to diagnose before it prescribes. The ranking criteria (impact and ease) force it to be practical, not theoretical. The “show the reasoning” instruction makes the logic auditable.
Useful phrases:
- “Show me your reasoning before you give me the answer.”
- “What assumptions are you making?”
- “Walk me through the logic.”
- “What are the strongest arguments against your conclusion?”
- “Where might this reasoning be wrong?”
- “Steel-man the opposing view before you make your case.”
- “What would have to be true for your recommendation to be wrong?”
- “Don’t just tell me what, tell me why.”
Principle 7: Structure Your Prompt
What it is: Use headers, bullet points, and labels in your prompts.
Why it works: Structured input produces structured output. The model parses your prompt by format. Headers signal topic changes. Bullet points signal lists. Labels signal categories. If you write a wall of text, you get a wall of text back.
The psychology: Structure is a signalling mechanism. When you label a section “Audience,” the model knows that information constrains the output. When you label a section “Tone,” the model knows to apply that tone throughout. Without labels, the model has to guess which parts of your prompt are constraints and which are context. It often guesses wrong.
Examples
Example 1: Brief-style prompt
Don’t: “I need to write a presentation about our Q3 results and it should cover revenue by segment, customer acquisition costs, retention rates, and what we’re doing differently next quarter, also make it for the board so keep it high-level but include the data they’ll ask about, oh and make sure there’s an executive summary.”
Do:
Task: Board presentation on Q3 results. Audience: Board of directors. High-level, data-ready. They’ll drill into numbers, so include specifics behind each headline. Sections:
- Executive summary (1 slide)
- Revenue by segment (with YoY comparison)
- Customer acquisition costs (trend and benchmark)
- Retention rates (by cohort)
- Q4 strategic priorities (3 max)
- Q&A prep (likely board questions and suggested answers) Tone: Authoritative, concise. No jargon. No marketing language. Length: 10-12 slides.
Why the second one works: Every piece of information is labelled and positioned. The model knows what’s a constraint, what’s content, and what’s format. The output will mirror the structure.
Example 2: Multi-part request
Don’t: “Help me with our rebrand. We need new messaging, a tone of voice guide, and some examples of how it would look in practice.”
Do:
Context: We’re rebranding from a budget student accommodation brand to a premium one. Our audience is international students who value quality and experience over price.
Task 1: Messaging framework
- Core positioning statement (one sentence)
- Three key messages (one sentence each)
- Proof points for each message
Task 2: Tone of voice guide
- Describe the tone in three words
- Give me a “we are / we are not” table (5 rows)
- Show me one paragraph written in this tone
Task 3: Practical examples
- Write one Instagram caption in this tone
- Write one email subject line and opening paragraph
- Write one property listing description
Why the second one works: The three tasks are clearly separated. Each has its own deliverables. The model won’t mix them together or skip one. The context at the top informs all three tasks.
Example 3: Feedback request
Don’t: “Give me feedback on this email.”
Do:
What I’m sending: Cold email to university accommodation managers introducing our new property.
What I want feedback on:
- Is the subject line strong enough to get opened?
- Does the opening line earn the reader’s attention in the first 5 seconds?
- Is the value proposition clear by the second paragraph?
- Is the call to action specific enough?
- What would you cut?
Tone of feedback: Direct. Don’t be polite. Tell me what’s wrong.
Why the second one works: The model knows exactly what to evaluate. Without the list, it might comment on tone when you cared about structure, or fix grammar when you cared about the value proposition. The “don’t be polite” instruction prevents the model from softening its critique.
Example 4: Research brief
Don’t: “Research the UK PBSA market for me.”
Do:
Research brief: UK PBSA market
Scope: Purpose-built student accommodation in the UK, focusing on the top 10 university cities.
Questions to answer:
- What’s the total market size (beds and revenue)?
- Who are the top 5 operators and what’s their market share?
- What’s the average occupancy rate and how has it changed in 3 years?
- What are the top 3 student complaints about PBSA?
- What trends are shaping the market in 2026?
Format: One page per question. Include data sources where available.
Constraints: If you don’t have data for a specific question, say so. Don’t estimate.
Why the second one works: The model has a clear brief with defined scope, specific questions, a format, and a constraint about uncertainty. The output will be structured, focused, and honest about gaps.
Useful phrases:
- Use labels: Task:, Audience:, Tone:, Format:, Constraints:
- Use numbered lists for sequential requirements
- Use headers to separate distinct parts of a request
- Use bullet points for parallel requirements
- State the output format explicitly
Combining Principles
The principles work best in combination. Here’s an example that uses all seven:
Prompt:
“I need a competitive analysis of the top five PBSA operators in Leeds.
Before you write anything, give me the structure: what sections you’ll cover and what data you’ll include in each. I’ll review before you proceed. [Principle 3: Plan]
Analyse each operator on: pricing, occupancy, Google reviews score, social media presence, and one unique differentiator. [Principle 4: Be Specific]
Format: A table with one row per operator, followed by a summary paragraph that identifies the gap none of them are filling. [Principle 7: Structure]
Only use data you’re confident about. If you don’t have a specific number, say so rather than estimating. [Principle 2: Don’t Guess]
Don’t stop at five operators if there are others worth mentioning. If a sixth or seventh is relevant, include them. [Principle 1: Define Done]
After the analysis, tell me which operator you’d be most worried about if we were entering Leeds, and show me your reasoning. [Principle 6: Why Before What]
Write in this tone: [paste example] [Principle 5: Show]”
Quick Reference
- AI gives shallow answers → Define “Done” → Tell it what complete looks like
- AI makes things up → Don’t Let It Guess → “If you’re not sure, say so”
- AI jumps straight to output → Plan Before Acting → “Outline your approach first”
- AI misinterprets your intent → Be Specific → Add constraints: length, tone, audience, format
- AI’s tone is wrong → Show, Don’t Tell → Paste an example of what you want
- AI’s answer feels unsupported → Ask “Why” Before “What” → “Show me the reasoning first”
- AI’s output is messy → Structure Your Prompt → Use headers, bullets, and labels
The One Rule
The model does what you tell it to do, not what you want it to do.
If the output is wrong, the instruction was wrong. Fix the prompt, not the model.
Adapted from OpenAI’s GPT-4.1 Prompting Guide (2026). The original guide is written for developers building agentic systems. This version is for anyone using AI.
The STAR Framework
If you enjoyed this essay, you'll find the full argument — and the framework behind it — in the book.