Zekaj tools

Prompting · OpenAI · Anthropic · Google

Prompt kit

Build a prompt step by step, or check one you already have. Every tip links to what OpenAI, Anthropic and Google publish about their own models, and to the research behind it. Where they disagree, you'll see that too.

This tool needs JavaScript. The short version: say what to do, why and for whom, what the answer should look like, mark off any pasted material, and allow "I don't know". The principles and sources below work without it.

12 principles

What the guides of the three biggest AI labs agree on, condensed. The chips link to the sources.

  1. 01Write for a smart newcomer

    Anthropic's test: show your prompt to a colleague with minimal context. If they'd be confused, the model will be too. OpenAI and Google both ask for specific, detailed instructions.

    AnthropicOpenAIGoogle

  2. 02Explain the why

    Purpose and audience help more than extra rules. Anthropic's example: instead of "never use ellipses", explain that the text will be read aloud by a speech engine that can't pronounce them. The model then generalises.

    AnthropicAnthropic blog

  3. 03Say what to do, not only what not to do

    Tell Claude what to do instead of what not to do, says Anthropic. OpenAI's help centre gives the same rule. "Don't use jargon" works worse than "write for a reader aged 12".

    AnthropicOpenAI

  4. 04Make it measurable

    OpenAI's advice: Reduce ‘fluffy’ and imprecise descriptions: "fairly short" becomes "3 to 5 sentences". Name the format too: table, bullets, paragraphs, JSON.

    OpenAIGoogle

  5. 05Show, don't only tell

    Examples are the strongest way to steer format and tone. Make them varied, keep their format consistent, and mark them clearly so they aren't mistaken for instructions. How many is debated (see below).

    AnthropicOpenAIGoogleResearch

  6. 06Separate instructions from material

    Wrap pasted text in tags like <document>, or use ### and Markdown headings. All three labs recommend clear delimiters so the model knows what to do and what to work on.

    AnthropicOpenAIGoogle

  7. 07Mind the order with long material

    Models use the start and end of a long prompt better than the middle. Anthropic reports that putting the question at the end improved response quality by up to 30 percent in their tests.

    AnthropicGoogleResearch

  8. 08Leave room for "I don't know"

    Models tend to fill gaps with plausible guesses. Explicitly allowing them to say information is missing, or to ask a question first, reduces made-up answers.

    Anthropic blogAnthropic

  9. 09One job per prompt

    Split big tasks into a chain: research, then draft, then review against your criteria. Each step is easier to check and fix. Each prompt should do one thing well.

    Anthropic blogAnthropicGoogle

  10. 10Let it think, don't micromanage

    Reasoning before answering improves hard tasks. But today's reasoning models do this themselves; OpenAI and Anthropic both advise high-level goals over hand-written step-by-step plans.

    ResearchOpenAIAnthropic

  11. 11Remove contradictions

    "Be brief" plus "explain in detail" forces the model to guess. OpenAI found contradictory or vague instructions can be more damaging to newer models. Longer prompts are not automatically better.

    OpenAIAnthropic blog

  12. 12Test, then iterate

    Decide what a good answer looks like before you start, try the prompt on a few real cases, and change one thing at a time. Rephrasing or reordering alone can change results.

    AnthropicOpenAIGoogle

Where the guides disagree

Advice depends on the model, the year and the task. Three places where the sources point in different directions, and what to do with that.

Instructions first, or question last?

  • OpenAI Help CentrePut instructions at the beginning, then separate the context with ### or """. O2
  • OpenAI API guideIdentity, instructions and examples first; context is usually best near the end. O1
  • Anthropic · GoogleWith long inputs: documents at the top, your question or instructions at the very end. A1G1

In practice: for short prompts it barely matters; lead with the task. With long material, put the material first and repeat the task at the end. The "lost in the middle" research explains why: models use the start and end of long contexts best. R3

How many examples?

  • GoogleWe recommend to always include few-shot examples in your prompts. Prompts without them are likely less effective. G1
  • Anthropic docsInclude 3 to 5 relevant, varied examples for best results. A1
  • Anthropic blog · OpenAIStart without examples or with just one. Add more only if the output still misses. A3O2

In practice: start with zero or one. When tone or format must be exact, add a few varied examples. Watch for copying: with a single example the model may mimic it too literally, and Google warns that too many examples can make it overfit. R4

"Let's think step by step": still needed?

  • Research, 2022Asking for reasoning steps, or simply adding "Let's think step by step", clearly improved maths and logic results. R1R2
  • OpenAIReasoning models work better with high-level guidance, like a senior colleague you give a goal. O1
  • AnthropicUse built-in thinking when available. "Think thoroughly" often beats a hand-written step-by-step plan. A1A3

In practice: with a reasoning or "thinking" model, describe the goal and what good looks like, and let it plan. With a fast, non-reasoning model, asking it to reason first still helps on tricky tasks.

Sources

Official guides from the labs themselves, plus the research papers they build on. Checked October 2026. These guides change with every new model, so check the date.

  1. A1Prompting best practices ↗Anthropic · Claude Developer Platform docs · living document
  2. A2Prompt engineering overview ↗Anthropic · on success criteria and testing before you tune
  3. A3Best practices for prompt engineering ↗Anthropic · Claude blog · November 2025
  4. O1Prompt engineering guide ↗OpenAI · API documentation
  5. O2Best practices for prompt engineering with the OpenAI API ↗OpenAI · Help Centre
  6. O3GPT-5 prompting guide ↗OpenAI · Cookbook · August 2025 (archived, written for GPT-5)
  7. G1Prompt design strategies ↗Google · Gemini API documentation
  8. R1Chain-of-thought prompting elicits reasoning in large language models ↗Wei et al. · NeurIPS 2022
  9. R2Large language models are zero-shot reasoners ↗Kojima et al. · NeurIPS 2022 · the "Let's think step by step" paper
  10. R3Lost in the middle: how language models use long contexts ↗Liu et al. · TACL 2024
  11. R4Language models are few-shot learners ↗Brown et al. · NeurIPS 2020 · the GPT-3 paper that showed learning from examples in the prompt