The short answer: say what result you want and who it's for, give the facts the AI can't guess, say the format and length you want back, and add one or two hard limits. That's it. Most of the famous tricks ("act as a world-class expert," "think step by step," "I'll tip you $200," ALL CAPS) don't reliably help today's models, and some backfire.
We read what OpenAI, Anthropic and Google say about their current models, went through the research on the tips people argue about, and measured the tokens. Here's what holds up in 2026.
The 30-second version
- Lead with the result. "Write a reply to this customer that…", not "I need help with something."
- Say who it's for. A board, a customer, a 10-year-old, yourself.
- Give the facts only you know. Names, numbers, what you can offer, what already happened.
- Say the format and length. "Five bullets." "Under 150 words." "A table with three columns."
- Add one or two hard limits. "Don't change the dates." "Draft only." "If you're not sure, say so instead of guessing."
- Read the answer critically and fix one thing at a time.
What changed in 2025 and 2026 (and why old prompt tips backfire)
Many prompt guides that rank today still teach advice written for GPT-4-era models. The current models from all three companies behave differently, and the companies' own advice has changed with them.
Today's models think before they answer
Reasoning is now built in. OpenAI's guidance for its reasoning models says to "avoid chain-of-thought prompts" (OpenAI). Anthropic says that for Claude Opus 5.5 in chat apps, removing a "think carefully before answering" line "made replies start sooner, with no clear decline in the quality of the reply" (Anthropic). On its newest models, including Opus 5.5 and Sonnet 5.5, a prompt asking the model to write out its reasoning in tags "may be declined" (Anthropic). Google says it's "generally not necessary" to have Gemini lay out its reasoning steps in the answer (Google).
In the apps, the better lever is the thinking or reasoning mode. Turn it on for hard problems; leave it off for quick ones.
They follow instructions more literally, so shouting overshoots
OpenAI's guidance for GPT-5.6 says: "Avoid unnecessary absolute rules. Use ALWAYS, NEVER, must, and only for true invariants such as safety rules, required fields, or actions that should never happen" (OpenAI). For GPT-6 Astra, OpenAI warns that strong wording written for older models means Astra "could take it too seriously and may stop work" (OpenAI). Anthropic suggests replacing tool instructions like "CRITICAL: You MUST use this tool when..." with plain wording like "Use this tool when...", because its newer models can overtrigger (Anthropic). Write the rule once, calmly, and say why it matters.
Leaner prompts scored better in OpenAI's own tests
OpenAI recommends removing repeated instructions and examples, and reports that in its internal coding-agent tests on GPT-5.6, leaner system prompts "improved evaluation scores by roughly 10–15% while reducing total tokens by 41–66%," results it calls directional (OpenAI). More words aren't more clarity.
| Old tip | What the guidance and research say now | Do this instead |
|---|---|---|
| "Think step by step" | Reasoning is built in; scripting it adds little | Use the app's thinking mode for hard tasks |
| "Act as a world-class expert" | Studies find roles shape tone, not correctness; Anthropic and Google still suggest one for tone | Say who the answer is for and the tone you want |
| ALL CAPS and "VERY IMPORTANT" | Save absolutes for real must-nots | State the rule once and explain why |
| "Be as detailed as possible" | Defaults already differ by model | Give a specific length or structure |
| Long, repeated instructions | Leaner prompts scored better | Include every needed fact, cut the repetition |
How to write a good prompt: the 4 building blocks
All three companies' current guides agree on these.
1. Lead with the result and the reader
OpenAI's guide for ChatGPT business users suggests asking yourself one question first: "What is the single output I want?" Google's advice is to "state your goal clearly and concisely." Put that first, along with who it's for: "Write a one-page summary of this report for our board, who care about cost and risk."
2. Give the context the AI can't guess
Anthropic's comparison is a good test: "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows." If a smart new colleague would need a fact to do the job well, the AI needs it too: the customer's name, what you can offer, the deadline, what you've already tried. Leave out what doesn't change the answer.
3. Say the format and length
This matters more than any trick. In a Wharton study of multiple-choice questions, removing the instruction telling the model how to format its answer cut accuracy by 8.6 points on GPT-4o and 12.1 points on GPT-4o mini, far more than politeness moved overall accuracy (Meincke et al., 2025).
Defaults also differ. OpenAI says GPT-6 Astra "tends toward detailed, formatted responses" with lists, tables and Markdown (OpenAI); Google says Gemini 3 models give "direct and efficient answers" unless you ask for more. Without a length, you get whatever that model's habit is. Short targets ("under 150 words," "5 bullets") work well; very long ones ("write 5,000 words") don't, according to a NeurIPS 2025 benchmark of 26 models (LIFEBench).
4. Add one or two hard limits
Limits stop the most expensive mistakes: "Keep the approved dates and budget unchanged." "Draft only; don't send." "If the answer depends on something I haven't told you, say so instead of guessing." OpenAI's guide for business users suggests a similar self-check: "Before finishing, list any assumptions you made."
When to add an example
The companies disagree. Google says "always include few-shot examples," Anthropic suggests 3 to 5, and OpenAI says to try without first. A Microsoft study of OpenAI's o1 reasoning model found that examples hurt its performance on medical questions (Nori et al., 2024). A practical rule: add an example when format or voice matters (paste one email in your style), and skip it for analysis and problem-solving.
Long documents: paste first, ask last
When you paste a long document, put it first and your question at the end. Anthropic says questions at the end "can improve response quality by up to 30 percent in tests," especially with several long documents (Anthropic). Google gives the same advice. Mark where the pasted text starts and ends, for example with """ quotes, and say what it is ("This is the customer's email").
Before and after: two real rewrites
Example A: a customer reply, with the tricks taken out
Before (86 words, 105 tokens):
Hi ChatGPT! I hope you're doing well. Act as a world-class customer service manager with 20 years of experience. Please, please help me. A customer named Maria emailed us really upset that her order #48213 arrived late and one of the mugs was chipped. I need you to write a reply to her. It's VERY IMPORTANT that you are EXTREMELY professional and NEVER blame the customer. Take a deep breath and think step by step. I'll tip you $200 if it's perfect. Thank you so much!!
After (73 words, 94 tokens):
Write a reply to this customer email. Goal: Maria feels taken care of and orders from us again.
Facts: order #48213 arrived late and one mug was chipped. We can ship a free replacement mug today or refund that mug; let her choose.
Don't blame her or the shipper, and don't promise a delivery date.
Under 120 words, warm and plain. Sign off as [your name].
Her email: """ [paste Maria's email] """
The "after" is shorter, yet it adds the one fact that changes the reply: what you can actually offer her. It also states the goal, a limit, the tone and the length. Here's what the tricks in the "before" cost, counted with OpenAI's tokenizer:
| Phrase | Tokens |
|---|---|
| "Act as a world-class customer service manager with 20 years of experience." | 15 |
| "Hi ChatGPT! I hope you're doing well." | 10 |
| "Take a deep breath and think step by step." | 10 |
| "I'll tip you $200 if it's perfect." | 9 |
| "Please, please help me." | 6 |
| "Thank you so much!!" | 5 |
| Total | 55 of 105 (52%) |
Example B: a longer prompt that gets a cheaper answer
Before (21 words, 23 tokens):
Can you tell me everything I need to know about starting a small bakery business? Please be as detailed as possible.
After (70 words, 88 tokens):
I want to start a home bakery in [your state], selling cakes at weekend farmers markets, with a $3,000 budget.
List my first 10 steps in order (licenses, kitchen rules, pricing, where to sell). For each step give one line on what to do and one on the likely cost.
Keep it under 300 words. Where the answer depends on my state's cottage food law, say so instead of guessing.
This prompt is 65 tokens longer, and it should usually be the cheaper one. Every current flagship model from OpenAI, Anthropic and Google charges about five times more per token for the answer than for your prompt. So 65 extra prompt tokens pay for themselves if the answer comes back just 13 tokens shorter, and "as detailed as possible" invites a much longer answer than a 300-word cap. The better prompt is also more useful: it's specific to your situation, and it tells the AI to flag what it doesn't know.
Prompt myths, tested
Most studies below tested older models, so we name them.
Does saying "please" help?
Not with accuracy, on current models. An October 2026 test on Claude Opus 5.5, Claude Haiku 5.5, GPT-6.1 Sol and Gemini 3.8 Flash found "accuracy didn't move: the current models got 93–96% right whatever I said to them" (Arize; a vendor test, not peer reviewed). A thank-you added up to 6% more tokens. Be polite if you like; it just doesn't change the answer much.
Does being rude work better?
No. The "rude prompts work better" headlines came from a 50-question test on one model, GPT-4o (Dobariya and Kumar, 2025). A larger study of GPT-4o mini, Gemini 2.0 Flash and Llama 4 Scout (2025, revised 2026) found them "broadly robust to tonal variation in typical mixed-domain use" (Cai et al.), though the Penn State authors' own follow-up found tone effects "highly model-dependent" (Dobariya and Kumar, 2026).
Tips, threats and "my job depends on this"
No reliable effect. Wharton researchers tested offers of $1,000 and a trillion dollars, threats and emotional stories on five OpenAI and Google models from 2024 and 2025 (including GPT-4o and Gemini 2.0 Flash) and found "threatening or tipping a model generally has no significant effect on benchmark performance" (Meincke et al., 2025). Individual answers swung both ways unpredictably, and one framing made two Gemini models reply to the made-up email instead of answering the question.
"Act as a world-class expert"
It changes the tone, not the correctness. Across 162 roles, "adding personas in system prompts does not improve model performance" (Zheng et al., EMNLP 2024, on four open model families), and a Wharton study of six OpenAI and Google models found expert personas don't "reliably improve accuracy" (Wharton, 2025). Roles can still set a voice. Saying who the answer is for ("explain this to a new hire") usually does that job better.
"Think step by step"
Mostly redundant now. Wharton found it gave reasoning models (o3-mini, o4-mini and Gemini 2.5 Flash) "only marginal benefits despite substantial time costs," with answers taking 20% to 80% longer (Wharton, 2025). Ask for the steps only when you want to check the work.
ALL CAPS and "IMPORTANT"
Rarely helps, and can backfire (OpenAI says GPT-6 Astra "could take it too seriously"), and it costs more: our capitalized sentence in Example A was 16 tokens against 14 in lowercase. OpenAI and Anthropic both now advise calm, plain rules.
Are longer prompts better?
Relevant beats long. When researchers padded the same question with about 3,000 tokens of extra text, average accuracy across five models (including GPT-4, GPT-3.5 and Gemini Pro) fell from 92% to 68% (Levy et al., ACL 2024). Include what the AI needs; cut what it doesn't.
ChatGPT prompt tips (and how Claude and Gemini differ)
- Use the thinking mode instead of asking it to think. ChatGPT, Claude and Gemini all have one.
- Expect different default lengths. OpenAI's top model leans long and formatted; Gemini leans short. Always say the length you want.
- Edit your message instead of piling on follow-ups. Each follow-up re-sends the whole chat. Fixing the original prompt and regenerating is cleaner and cheaper.
- Start a new chat for a new task. Old context can pull answers off course and costs tokens on every message.
- Ask the AI to improve your prompt. OpenAI suggests asking ChatGPT to help write the prompt; Google suggests typing "Make this a power prompt:" in Gemini. Then check that the result kept your facts.
Prompt engineering for beginners: a 10-minute practice routine
- Write the prompt with the four building blocks.
- Run it in a fresh chat.
- Read the answer as a critic: what's wrong, missing or too long?
- Change one thing in the prompt and run it again in a fresh chat.
- When it works, save it as a template, with [brackets] for the parts that change. Our guide to saving and reusing prompts shows where to keep it.
Prompts for business: make them repeatable
- One task per prompt. Several jobs in one prompt get thinner answers.
- Use templates with named blanks, like [customer name] and [order number], so the facts change and the instructions don't.
- Mark pasted text clearly, so the AI knows which part is the customer's email and which part is your instruction.
- Tell it to write "Not specified" rather than invent a missing detail.
- Never paste passwords or data you're not allowed to share.
- Start from a tested template: our prompt library has 116 free ones for about 30 industries.
Better prompts cost less
If you pay per use, prompting is a budget decision. The answer costs about five times more per token than the prompt, so a clear length limit is the single cheapest improvement you can make. There's one catch: on reasoning models, telling the AI to be brief can increase its hidden thinking. Arize measured hidden thinking nearly doubling on Opus 5.5 and GPT-6.1 Sol when they were told not to add extra text, so check your real bill, not only the visible answer.
The other big saving is the model itself. A clear prompt often lets a small, cheap model do a job you'd otherwise send to an expensive one. See how to reduce your AI costs, or count your own prompts with the free token counter.
Prompt optimizers and prompt improvers: do they work?
A prompt optimizer rewrites your prompt to be clearer and more specific. There are two kinds: one-click rewriters for chat users, and tools for developers that test versions of a prompt against graded examples. The second kind has much stronger evidence, because it measures results.
The companies' own tools, and the ones going away
- OpenAI's dataset-backed prompt optimizer is being retired with its Evals platform, which becomes read-only for existing users on October 31, 2026 and is scheduled to shut down on November 30 (OpenAI).
- Anthropic has retired its Console Workbench, where it launched its prompt improver in 2024; the new playground "doesn't support saving prompt history or evaluating prompts" (Anthropic). Anthropic hasn't said what happened to the improver.
- PromptPerfect, once a popular consumer optimizer, is reported to have shut down on September 1, 2026. That's from competitors' pages; we found no official notice.
What the evidence says
For one-click rewriters on today's chat models, independent evidence is thin. The only cross-tool comparison we found, from BenchLM in July 2026, included BenchLM's own tool and found the tools "tied on fully accepted outputs." It didn't compare the rewritten prompts with the originals, so it can't say whether rewriting helped at all (BenchLM).
Test any optimizer yourself in 10 minutes
- Pick a real prompt you use.
- Run the original in a fresh chat and save the answer.
- Run the rewritten version on the same model in another fresh chat.
- Compare the two answers, and the length of each.
Where PromptSaver fits
PromptSaver is our prompt optimizer, so we're not neutral. It rewrites a prompt using the building blocks above (it keeps every fact and limit, adds a length when the answer has a natural size, removes filler), shows what the prompt will cost, and names the cheapest model at Anthropic, OpenAI and Google that suits the task. Your prompt is sent to our AI provider to do the rewrite, and PromptSaver doesn't save its text. Test PromptSaver the same way you'd test any other tool.
Frequently asked questions
What is the best way to write an AI prompt?
Start with the result you want and who it's for, add the facts the AI can't guess, then say the format and length. Add one or two hard limits, like "draft only" or "don't invent numbers." OpenAI, Anthropic and Google all recommend these building blocks for their current models.
Does saying "please" to ChatGPT make a difference?
Not to accuracy on current models. An October 2026 test on Claude Opus 5.5, GPT-6.1 Sol and Gemini 3.8 Flash scored 93% to 96% whatever the tone. Politeness adds a few tokens, and being rude doesn't reliably help either.
Should I tell the AI to "act as" an expert?
It won't make answers more correct. Studies covering 162 roles and six OpenAI and Google models found expert personas don't reliably improve accuracy. Roles can change tone and depth, but saying who the answer is for usually works better.
Do I still need to say "think step by step"?
Usually not. Today's reasoning models already think before answering, and Wharton researchers measured only marginal gains with 20% to 80% slower answers. Turn on the app's thinking mode for hard problems instead.
How long should a prompt be?
Long enough to include what the AI can't guess, and no longer. Padding a question with extra text lowered accuracy sharply in one study, and in OpenAI's internal coding tests, leaner system prompts scored roughly 10% to 15% higher.
Does offering a tip or threatening the AI work?
No. A 2025 Wharton test of tips up to "a trillion dollars," threats and emotional stories on five models found no significant overall effect, with individual answers swinging both ways.
Should I include examples in my prompt?
For format and writing style, yes, and one good example is often enough. For reasoning tasks on reasoning models, examples can hurt. Google says always include them, Anthropic suggests 3 to 5, and OpenAI suggests trying without first.
Do ALL CAPS or words like "IMPORTANT" help?
Rarely, and they can backfire. OpenAI says to save words like ALWAYS and NEVER for true must-follow rules, and Anthropic says its newer models can overtrigger on tool instructions written with "CRITICAL" and "MUST." Capitals also use more tokens.
Can ChatGPT or Claude follow a word count?
For short lengths, mostly yes. A NeurIPS 2025 benchmark of 26 models found they reasonably follow short length instructions but fall short on very long targets. Ask for a range or a structure, like "5 bullets."
What is a prompt optimizer, and do they work?
A prompt optimizer rewrites your prompt to be clearer and more specific. Developer tools that test rewrites against graded examples have the best evidence; for one-click rewriters, independent evidence is thin. Test any tool by comparing answers from your original and the rewritten prompt.
How we researched this
We read the current prompting guides from OpenAI, Anthropic and Google, the peer-reviewed and preprint studies linked above, and the pages of the optimizer tools mentioned, on October 30, 2026. Token counts are our own, using OpenAI's o200k tokenizer. Most published studies tested older models, so we name them. PromptSaver makes a prompt optimizer, and we've said so where it matters.