Squeeze My Prompt

How to write a complete prompt

Published November 10, 2026. It will be updated as the AI makers' guidance changes.

This guide is for people who type or paste one message into an AI tool such as ChatGPT, Claude, Gemini, or a coding agent. It covers what to put in that message. It does not cover settings that only developers can reach.

It is built from the AI makers’ own documentation and from published studies. Numbers in square brackets point to the source list at the end. Where the sources disagree, or where something could not be confirmed, the guide says so.

1. The main idea: say it all up front

Most people find their idea by chatting. They send a short request, see what comes back, and correct it over several messages. Google reports that the prompts people try are usually under nine words, while its most useful prompts average about 21 words and include context [1].

A 2025 study by researchers at Microsoft Research and Salesforce tested what this costs [2]. It gave fifteen AI models the same instructions two ways: the whole instruction in one message, or the same instruction revealed bit by bit over a conversation. The models averaged 90% when given the whole instruction at once and 65% when the same instruction was revealed over a conversation. Measured task by task, performance fell by an average of 39% across six kinds of task and fifteen models. The study compared complete instructions with spread-out ones. It did not test how a prompt is laid out.

So the single most useful habit is to work out what you want first, then say it in one complete message. Treat the prompt as a brief you would hand to a capable colleague who knows nothing about your situation. Anthropic’s test is the same: if a colleague with little context would be confused by your prompt, the AI will be too [3].

2. What a strong prompt contains

The three largest AI makers agree on the core [3][4][5]. A strong prompt states:

  1. The goal. What you want, starting with a verb.
  2. The purpose. What the result is for and what you will do with it. Anthropic’s newest guidance is direct about this: give the reason, not only the request [6].
  3. Who it is for. The reader or user, and what they already know.
  4. Context the AI cannot guess. Your situation, what you have tried, the facts that matter.
  5. What “done” looks like. How you will judge the result. In software work, Anthropic’s Claude Code guide says that when the AI reports work as done and it is not quite right, the usual cause is that it had no way to check its own work [7].
  6. Limits, with reasons. What to stay within and what to leave out for now. A rule with its reason attached lets the AI apply it sensibly to cases you did not foresee [3].
  7. The shape of the answer. Format, length, and level of detail, described as what you want rather than what you don’t [3].

Two more are worth adding when you have them:

  • Your material. A draft, notes, data, or a document. Put it before the request (see section 3).
  • An example of what good looks like. Examples steer format and tone strongly. Be aware that the AI will copy them closely, so one example can narrow the result more than you intend [3][5].

More is not always better. A Carnegie Mellon study found that simply listing more requirements does not reliably help [8], and OpenAI advises starting from the smallest prompt that does the job [4]. Every line should tell the AI something it could not have guessed.

3. How to lay it out

Use clear section markers. Anthropic, Google, and OpenAI all accept XML-style tags such as <context> and <task> to separate the parts of a prompt [3][4][5]. Tags are plain text and paste the same way into any chat box. They keep the parts of a long prompt apart; they are not what makes a prompt work.

One layout works for every tool. A March 2026 benchmark compared XML, Markdown, and JSON formatting on four models. On three of them, including the two leading ones, the gap was 0.3 points or less. The fourth, a smaller model, scored 12.4 points lower with Markdown than with tags [9]. An earlier study, not yet peer-reviewed, found large swings on an older model and much more robustness on a newer one [10]. So the format you choose matters little on current leading models, and you do not need a different prompt for each AI. In these tests the choice of markup made little difference.

Put long material first and the request last. Anthropic and Google both advise this [3][5]. Anthropic reports that placing the question after long documents improved answer quality by up to 30% in its tests. That figure applies to very long inputs, of roughly 20,000 tokens or more, so it should not be read as a general gain [3].

A layout that satisfies all three makers:

One plain sentence stating the goal.

<context> who you are, who it is for, background </context>
<source_material> anything you pasted </source_material>
<examples> if you have them </examples>
<task> what to do </task>
<constraints> limits, and what is out of scope </constraints>
<output_format> the shape of the answer </output_format>

One closing sentence restating the request.

Leave out any section you have nothing for.

4. What to leave out

These are popular and no longer help on current models.

Technique What the evidence says
“Think step by step” On reasoning models, a Wharton study measured changes of +2.9%, +3.1%, and −3.3%, with responses taking 20–80% longer [11]. OpenAI says to avoid step-by-step guidance unless the exact path matters [4].
Asking the AI to write out its reasoning Anthropic’s documentation says that on its newest models this “might be declined” [3]. Asking for a short explanation of the answer is still fine.
ALL-CAPS, “CRITICAL”, “you MUST” Anthropic’s documentation says its Opus 4.5 and 4.6 models over-react to forceful wording, and advises plain phrasing [3]. OpenAI says to keep absolute words for true invariants [4].
Offering tips, making threats, emotional appeals A Wharton study across five models found no meaningful effect on accuracy, and unpredictable swings on individual questions [12].
“You are a world-class expert” Two studies found expert personas do not improve factual accuracy [13][14]. A role is still useful for setting tone and voice [3][5].
“Double-check your work” Results are mixed and depend on the model. Self-review helped on 2023 models for style tasks [15], failed to help reasoning without outside feedback [16], and Anthropic advises removing such lines for one current model because it over-verifies [17]. Stating what “done” looks like works better.
Long rulebooks and prescribed procedures All three makers now advise describing the destination and trusting the model with the route [3][4][5].

5. What does help

Technique Evidence
Stating everything in one message Study: 39% average drop when spread over turns [2].
Giving the reason behind a rule Maker documentation [3][4][6].
Stating what is out of scope Maker documentation. Phrase it as a boundary (“not in this version”) rather than a list of prohibitions [3][17].
Permission to say “I don’t know” Anthropic says this can sharply reduce false information [18]. OpenAI and Google give equivalent advice [4][5].
Giving today’s date for time-sensitive questions Google advises telling the model to pay attention to the current year [5].
Asking the AI to question you first, when it matters OpenAI: ask only when the missing information would materially change the answer [4].
Having the AI interview you A study found model-led questions matched or beat prompts users wrote themselves, with less effort [19]. Finish by having it restate the whole brief, since scattered instructions cost quality [2].
Asking for many ideas that differ from each other AI-assisted brainstorming produces less varied ideas [20]. Asking for several distinct ideas with structure restores variety [21][22].
Framing your own view as something to test Models tend to agree with the user [23]. Ask for the strongest case against your view.
Examples Strong effect on format and tone. See the disagreement in section 7.

6. By type of task

Building software. Tool makers’ guides ask for the same four things: what you are building, who uses it and in what situation, the specific things it must do, and how you will check it works [7][24][25][26]. Name a look and feel, because without one the result defaults to a generic style [27][28]. Say who is allowed to see what data, and name anything sensitive: in Veracode’s 2025 tests of 80 coding tasks across more than 100 models, 45% of the AI-written code contained a well-known type of security flaw [29]. Ask the AI to plan and raise questions before it writes code [7][30].

Writing and editing. About two-thirds of writing requests ask the AI to change text the person supplies [31]. Paste your draft, say who will read it, what it should make happen, and how heavy an edit you want. List the facts that must appear and mark any you are unsure of, so the AI does not invent specifics [4].

Researching a topic. State the question, what you will do with the answer, the scope (period, place, limits), the sources you trust, and the output you want [32][33][34]. Deep research tools build a plan from whatever you give them, so these details decide what gets searched.

Analyzing data. Describe the data: what one row is, which columns matter, what period it covers. Define your terms (“repeat customer means two or more orders in twelve months”). Say what decision the analysis informs. Ask for a data-quality check first [35].

Analyzing a market or a stock. No AI maker publishes guidance specific to this. The evidence points to dates and sources as the main risk: in one benchmark an earlier model with document search answered 81% of questions about company filings wrongly or not at all [36]. Name the exact subject and period, give an as-of date, say which sources to use, and ask for the case for and against. State your own view as something to be challenged [23].

Planning for a business. Describe the business, the current situation and what you have tried, the limits any plan must respect, and what you want back [1]. Generic input produces textbook advice.

Learning something. Say what you already know and where it gets unclear, and why you are learning it. An Anthropic analysis comparing countries and US states found that the education level of people’s prompts closely tracks the education level of the replies they get [37].

Getting advice or deciding. Give the facts, the exact question, and what matters most to you in order.

Brainstorming. State the challenge as a question, how you will judge ideas, what is fixed, and what you have already tried. Ask for a number of ideas and for them to differ from each other.

7. Where the sources disagree

  • Examples. Google recommends always including them [5]. Anthropic recommends three to five varied ones [3]. OpenAI’s newest guides do not mention them, and its older guidance says to try without first [4]. A reasonable reading: include an example when format or tone matters, and expect it to be copied closely.
  • How ambitious to be in a first software build. Lovable advises keeping the first version very small [25]. Anthropic advises asking for ambition, because the model otherwise builds the minimum [7]. Decide deliberately and say which you want.
  • Saying what not to do. Replit advises describing what you want rather than what to avoid [30]. Bolt, Lovable, and v0 all advise stating what must not change [24][25][38]. Both hold if exclusions are written as boundaries with reasons.
  • Where context goes. OpenAI’s general guide places context near the end of a developer message [39]. Anthropic and Google place long material first [3][5]. A goal sentence at the top with the full request at the bottom satisfies both.

8. What a pasted prompt cannot control

Much developer guidance concerns settings a chat user cannot reach: reasoning effort, structured output schemas, and system prompts. If a task is hard, choose your tool’s “thinking” or higher-effort mode in its own menu. Wording in the prompt is not a substitute.

9. Limits of this guide

  • Guidance changes with each model generation. Most maker documentation pages show no date. Everything here was checked on November 10, 2026, and this guide will be updated as the AI makers’ guidance changes.
  • Several studies cited used models that are no longer current [11][36]. They are included for the direction of their findings.
  • Some maker help pages blocked automated reading during research, so a few points in the underlying research files rest on search summaries. Those points were marked in the research notes and have been kept out of this guide where possible.
  • Not found in any maker’s documentation: a statement that flattering role descriptions hurt results. The studies cited show they do not help accuracy.

Sources

All checked November 10, 2026.

  1. Google, “Google Workspace with Gemini prompting guide” (edition for small businesses and startups), September 2024. https://services.google.com/fh/files/misc/google_workspace_prompting_guide_abridged_smbs_startups_september2024.pdf
  2. Laban, Hayashi, Zhou, Neville, “LLMs Get Lost In Multi-Turn Conversation”, arXiv 2505.06120, May 9, 2025. https://arxiv.org/abs/2505.06120
  3. Anthropic, “Prompting best practices”, no date shown. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
  4. OpenAI, “Prompt guidance” (GPT-5.5 version), no date shown. https://developers.openai.com/api/docs/guides/latest-model?model=gpt-5.5
  5. Google, “Prompt design strategies”, no date shown. https://aistudio.google.com/docs/prompting-strategies
  6. Anthropic, “Prompting Claude Fable 5”, no date shown. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5
  7. Anthropic, “Best practices for Claude Code”, no date shown. https://code.claude.com/docs/en/best-practices
  8. Yang et al., “What Prompts Don’t Say: Understanding and Managing Underspecification in LLM Prompts”, arXiv 2505.13360, May 19, 2025. https://arxiv.org/abs/2505.13360v1
  9. Systima, “The Delimiter Hypothesis: Does Prompt Format Actually Matter?”, March 5, 2026 (vendor blog, not peer-reviewed). https://systima.ai/blog/delimiter-hypothesis
  10. He et al., “Does Prompt Formatting Have Any Impact on LLM Performance?”, arXiv 2411.10541, November 15, 2024. https://arxiv.org/abs/2411.10541
  11. Meincke, Mollick, Mollick, Shapiro, “The Decreasing Value of Chain of Thought in Prompting”, Wharton, June 8, 2025. https://gail.wharton.upenn.edu/research-and-insights/tech-report-chain-of-thought/
  12. Meincke, Mollick, Mollick, Shapiro, “I’ll pay you or I’ll kill you — but will you care?”, Wharton, August 1, 2025. https://gail.wharton.upenn.edu/research-and-insights/techreport-threaten-or-tip/
  13. Zheng et al., “When ‘A Helpful Assistant’ Is Not Really Helpful”, Findings of EMNLP 2024, arXiv 2311.10054. https://arxiv.org/abs/2311.10054
  14. Wharton, “Playing Pretend: Expert Personas Don’t Improve Factual Accuracy”, December 7, 2025. https://gail.wharton.upenn.edu/research-and-insights/playing-pretend-expert-personas/
  15. Madaan et al., “Self-Refine: Iterative Refinement with Self-Feedback”, arXiv 2303.17651, 2023. https://arxiv.org/abs/2303.17651
  16. Huang et al., “Large Language Models Cannot Self-Correct Reasoning Yet”, ICLR 2024, arXiv 2310.01798. https://arxiv.org/abs/2310.01798
  17. Anthropic, “Prompting Claude Opus 5”, no date shown. https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5
  18. Anthropic, “Reduce hallucinations”, no date shown. https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations
  19. Li, Tamkin, Goodman, Andreas, “Eliciting Human Preferences with Language Models”, ICLR 2025, arXiv 2310.11589. https://arxiv.org/abs/2310.11589
  20. Meincke, Nave, Terwiesch, “ChatGPT decreases idea diversity in brainstorming”, Nature Human Behaviour, 2025. https://mackinstitute.wharton.upenn.edu/2025/new-in-nature-chatgpt-decreases-idea-diversity-in-brainstorming/
  21. Meincke, Mollick, Terwiesch, “Prompting Diverse Ideas: Increasing AI Idea Variance”, arXiv 2402.01727, 2024. https://arxiv.org/abs/2402.01727
  22. Zhang et al., “Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity”, arXiv 2510.01171, 2025. https://arxiv.org/abs/2510.01171
  23. Sharma et al., “Towards Understanding Sycophancy in Language Models”, arXiv 2310.13548, 2023. https://arxiv.org/abs/2310.13548
  24. Esteban Suárez, “How to prompt v0”, Vercel, December 15, 2025. https://vercel.com/blog/how-to-prompt-v0
  25. Lovable, “Best practice”, no date shown. https://docs.lovable.dev/tips-tricks/from-idea-to-app
  26. GitHub, “Best practices for using Copilot to work on tasks”, no date shown. https://docs.github.com/en/copilot/tutorials/cloud-agent/get-the-best-results
  27. Anthropic, “Improving frontend design through skills”, November 12, 2025. https://claude.com/resources/articles/improving-frontend-design-through-skills
  28. OpenAI, “Designing delightful frontends with GPT-5.4”, no date shown. https://developers.openai.com/blog/designing-delightful-frontends-with-gpt-5-4
  29. Veracode, “2025 GenAI Code Security Report” press release, July 30, 2025. https://www.veracode.com/press-release/ai-generated-code-poses-major-security-risks-in-nearly-half-of-all-development-tasks-veracode-research-reveals/
  30. Replit, “Effective prompting”, no date shown. https://docs.replit.com/learn/effective-prompting
  31. Chatterji et al., “How People Use ChatGPT”, NBER Working Paper 34255, September 2025. https://www.nber.org/papers/w34255
  32. OpenAI Forum, “Exploring Deep Research: Three tips for better AI-assisted inquiry”, April 29, 2025. https://forum.openai.com/public/blogs/exploring-deep-research-three-tips-for-better-ai-assisted-inquiry-2025-04-28
  33. OpenAI Cookbook, “Introduction to deep research in the OpenAI API”, June 26, 2025 (archived). https://developers.openai.com/cookbook/examples/deep_research_api/introduction_to_deep_research_api
  34. Perplexity, “Prompt guide”, no date shown. https://docs.perplexity.ai/docs/agent-api/prompt-guide
  35. Anthropic Claude Academy, “Visualize program data”, no date shown. https://academy.claude.com/use-cases/visualize-program-data
  36. Islam et al., “FinanceBench: A New Benchmark for Financial Question Answering”, arXiv 2311.11944, November 20, 2023. https://arxiv.org/abs/2311.11944
  37. Anthropic, “Anthropic Economic Index report: Economic primitives”, January 15, 2026. https://www.anthropic.com/research/anthropic-economic-index-january-2026-report
  38. Bolt, “Prompting effectively”, no date shown. https://support.bolt.new/prompting/prompting-effectively
  39. OpenAI, “Prompt engineering”, no date shown. https://developers.openai.com/api/docs/guides/prompt-engineering