PromptCorrectly
PacksBundlesLibraryGamesBlog
Sign inStart free
PromptCorrectlyBLOG
← All articles
TECHNIQUES · 11 min read

Same Question, Different Words: How Small Prompt Changes Swing AI Results (15 Swaps That Work)

promptcorrectly.com · Updated 2026-11-16

A few words — sometimes just the formatting — can change what an AI gives you more than switching to a "better" model. Researchers have measured accuracy swinging by up to 76 percentage points from formatting changes alone, and adding one line asking for step-by-step reasoning has repeatedly lifted results on reasoning tasks. The practical lesson is simple: the words you choose tell the model what kind of expert to be and what a good answer looks like. Below is what the research shows, why it happens, and 15 swaps you can use today.

What the research shows

Formatting alone can swing results enormously

In a paper presented at ICLR 2024, Melanie Sclar and colleagues tested what happens when you change only the format of a prompt — spacing, separators, capitalisation, how examples are laid out — while keeping the meaning identical. On one open model (LLaMA-2-13B) they measured differences of up to 76 accuracy points between equivalent prompts. The effect did not go away with bigger models, more examples or instruction tuning, and they also found wide spreads for GPT-3.5 (Sclar et al., arXiv 2310.11324).

That doesn't mean every prompt is that fragile — the average effect across tasks was much smaller. It means the gap between a careless prompt and a careful one can be far larger than people expect, and that "the AI is bad at this" is often really "this phrasing was bad for the AI."

Asking for the reasoning changes the answer

In the 2022 chain-of-thought paper, Jason Wei and colleagues showed that giving a model a few examples that spell out their reasoning step by step raised Google's PaLM 540B from about 18% to 57% on the GSM8K maths word-problem benchmark (Wei et al., 2022). A follow-up paper by Kojima and colleagues found that simply adding the line "Let's think step by step" before the answer produced large gains on similar tasks with no examples at all (Kojima et al., 2022). Modern models often reason internally, but asking them to show the steps still helps you check the answer. Our chain-of-thought guide covers when it's worth it.

Why a few words matter so much

A language model predicts what text should come next, based on everything it has read. Your words are the only clue it has about which part of everything it has read to draw on.

  • Vague words pull the average. "Make this better" matches millions of mediocre edits. The model gives you the middle of that pile.
  • Precise words pull the expert. "Tighten this to 120 words, keep the three numbers, lead with the customer's problem" matches the way editors actually talk — and editors' work.
  • Domain terms switch on domain knowledge. "Cohort retention by signup month" and "how many people come back" ask the same question, but the first one points straight at analytics practice: the right method, the right chart, the right caveats.
  • Constraints remove guesswork. Every detail you leave out is a decision the model makes for you, usually the most generic one.

This is also why industry vocabulary is so valuable: one correct term can replace three sentences of explanation.

15 swaps that work

Each pair asks for the same thing. The second version tells the model what kind of expert to be and what a good result looks like.

Everyday writing

  1. Instead of "Make this email better" → try "Rewrite this email for a customer who's annoyed about a late delivery. Under 120 words, apologise once, give the new date, no exclamation marks."
  2. Instead of "Summarise this" → try "Summarise this for a busy manager in five bullet points: decision needed, deadline, cost, risks, who's affected."
  3. Instead of "Make it sound professional" → try "Use a plain, confident tone — short sentences, no jargon, no filler phrases like 'I hope this finds you well'."

Marketing and sales

  1. Instead of "Write a catchy headline" → try "Write five benefit-led headlines under eight words for [audience], each naming one concrete outcome."
  2. Instead of "Write a sales email" → try "Write a cold email: one-line personal opener about [their company], the problem we solve in their words, one proof point, and a yes/no question as the call to action."
  3. Instead of "Give me marketing ideas" → try "Give me ten low-budget local marketing ideas for a [business type] in [city], each with cost, time to set up, and how I'd measure it."

Business and numbers

  1. Instead of "Analyse these sales" → try "Show month-over-month revenue change, top five products by margin, and any product whose sales fell three months in a row."
  2. Instead of "How many customers come back?" → try "Calculate cohort retention by signup month for months 1, 3 and 6, and flag cohorts below the average."
  3. Instead of "Should I raise prices?" → try "Model a 10% price rise: break-even customer loss, likely objections, and a short message to regulars announcing it."

Code and building

  1. Instead of "Fix this bug" → try "Write a failing test that reproduces this bug, then the smallest change that makes it pass. Don't touch other files."
  2. Instead of "Make it faster" → try "Profile first: tell me where the time goes, then suggest the one change with the biggest effect and its trade-off."
  3. Instead of "Build a login page" → try "Add email sign-in using [provider]. Done when: sign up, sign in, sign out and password reset all work on mobile. List the files you'll change first."

Careful topics

  1. Instead of "Check this contract" → try "List the clauses on termination, notice period, liability cap and auto-renewal, with clause numbers. Flag anything I should check with a lawyer. Don't give legal advice."
  2. Instead of "Is this right?" → try "List anything in this that you're not sure is accurate, and what I'd need to check to confirm it."
  3. Instead of "Explain this" → try "Explain this to a smart 15-year-old in under 100 words, then give one example from everyday life."

Notice the pattern: audience, length, must-include, must-avoid, and a clear finish line. That's the five-part structure — Role, Context, Task, Constraints, Output — in miniature. See the RCTCO prompt structure for the full version.

A 30-second checklist before you press Enter

  • Who is this for? (Name the reader.)
  • How long? (Give a number.)
  • What must it include? (List the facts.)
  • What must it avoid? (Tone, claims, topics.)
  • What does "done" look like? (Format and finish line.)
  • Did I use the right word for the job? (The term an expert in this field would use.)

Skip the blank page

Getting the wording right every time is work. That's the point of a finished prompt: someone who knows the field has already chosen the terms, the constraints and the output format, and left blanks for your details.

Every PromptCorrectly pack is written that way for one trade or topic, and comes with a start file, 00-GIVE-THIS-TO-YOUR-AI.md, that has your AI ask for your details once and fill them into every prompt in the pack. Want to try the style first? Start with the free prompts or the how-to-prompt guide.

Questions people ask

Does the wording of a prompt really change the answer?

Yes. A 2024 ICLR paper by Sclar and colleagues found that changing only the formatting of equivalent prompts shifted accuracy by up to 76 points on one open model, and the effect persisted with larger models. Clear, specific wording reliably produces better, more consistent answers.

Does 'let's think step by step' still work?

Asking a model to reason step by step was shown to improve results on reasoning tasks in 2022 research. Newer models often reason internally, but asking them to show their steps still makes the answer easier to check.

What words make AI answers better?

Words that name the audience, the length, what to include, what to avoid and what the finished result looks like — plus the exact term an expert in the field would use, such as 'cohort retention' instead of 'how many come back'.

Put this into practice

Grab a ready-made prompt from the Library, or a pack for your trade, fill in your details and take it to the AI you already use.

Browse the prompt libraryShop the packs

Keep reading

AI in 2026: 900 Million People Use It Every Week — So Why Do Most Results Disappoint?

ChatGPT has 900M weekly users and 88% of firms use AI, yet most see little return. What the 2025–26 research says, and how better prompts close the gap.

10 min read

Chain-of-Thought Prompting: How and When to Use It

When chain-of-thought prompting helps, when it hurts, and how to make the model reason step by step then hand you one clean answer.

9 min read

50 AI Prompt Examples You Can Copy, Paste, and Improve

50 ready-to-paste AI prompt examples for writing, work, coding, research, study, and marketing — plus the pattern that makes each one actually work.

11 min read
← All articles
promptcorrectly.comFROM VAGUE INTUITION TO STRUCTURED INSIGHTLIBRARY · PACKS · GAMES
PromptCorrectly

The marketplace for AI prompts, skills and packs. Ready-to-use prompts, installable AI skills and trade packs for ChatGPT, Claude and Gemini. Browse free, or buy a ready pack from €4 — pay once, keep it.

Product
  • Packs
  • Bundles
  • Skills
  • Prompts
  • Free prompts
  • Free skills
  • Games
  • A–Z index
  • Every job we cover
  • Made for your AI
  • Basket
  • Search
Resources
  • Prompts by field
  • All prompts A–Z
  • How to prompt correctly
  • Community uploads
  • AI glossary
Company
  • About
  • Contact
  • FAQ
  • Blog
© 2026 PromptCorrectly · The marketplace for AI prompts, skills and packs.
TermsPrivacyRefundDisclaimerAcceptable useCookies