Prompt checklist
| Element | What to write | Source |
|---|---|---|
| Goal | One sentence on the task and what "done" looks like | Anthropic: be clear and direct |
| Reason | Why the constraint matters | Anthropic: add context to improve performance |
| Audience | Who reads the output and what they already know | All three guides |
| Inputs | The documents or data, labeled and separated | OpenAI: Markdown and XML for boundaries |
| Examples | 2 to 5 varied input and output pairs | Google: always include few-shot examples |
| Format | Length, structure, fields, tone | All three guides |
| Order | Long material first, the question last | Anthropic and Google long-context advice |
| Test | 10 or more real cases with a pass rule | OpenAI and Anthropic evaluation guidance |
Seven steps to a better prompt
- Define success before you write. Anthropic's prompt engineering overview assumes you already have "a clear definition of the success criteria" and a way to test against it. Write two or three pass rules, such as "names the refund amount" or "under 120 words."
- Say exactly what you want. Anthropic's prompting best practices describe the model as "a brilliant but new employee who lacks context on your norms." Name the deliverable, the reader and the bar.
- Explain the why. The same guide marks "Your response will be read aloud by a text-to-speech engine, so never use ellipses" as more effective than "NEVER use ellipses." Reasons let the model generalize the rule.
- Show, do not only tell. Google's prompting strategies guide says "We recommend to always include few-shot examples in your prompts," and that clear examples can replace some instructions. Vary them so the model copies the pattern, not one surface detail.
- Separate instructions from material. OpenAI's prompt engineering guide recommends Markdown headers and XML tags to mark "logical boundaries" between instructions and context data.
- Order long prompts carefully. Put documents at the top and the question at the end. Anthropic reports that queries at the end can improve response quality by up to 30% on complex multi-document inputs.
- Test, then pin. Run the prompt on at least 10 real inputs, score each against your pass rules and fix the worst failure first. OpenAI recommends pinning production apps to specific model snapshots, because different snapshots "could produce different results."
Google's Gemini prompting guide, checked 17 October 2026, recommends always including few-shot examples and says prompts without them are likely to be less effective.
Where the techniques came from
Two research papers explain why examples and step-by-step reasoning help.
- Few-shot prompting. The GPT-3 paper (Brown and colleagues, 2020) showed a 175 billion parameter model could perform new tasks from demonstrations "specified purely via text interaction," with no fine-tuning.
- Chain-of-thought prompting. Wei and colleagues (2022) found that giving a 540B-parameter model just eight worked examples with reasoning steps reached state-of-the-art accuracy on the GSM8K math benchmark at the time.
Many current models now reason internally. Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 3.8 Flash all expose an effort or thinking setting, so on those models you usually raise the setting rather than write "think step by step." Ask for the reasoning in the output only when you need to audit it.
Worked example: a customer reply
Before:
Write a reply to this customer complaint.
The model has to guess the tone, the policy, the length and what the company can offer.
After:
You are drafting an email reply for a support agent at an online furniture store. The agent will edit and send it.
Goal: acknowledge the problem, state what we will do, and give one clear next step.
Policy: damaged items qualify for a free replacement or a full refund within 30 days of delivery. We cannot offer extra discounts.
Format: under 120 words, plain text, no bullet points, sign off as "The Support Team".
Why: customers read these on phones, and long replies get escalated.
Example input: "My lamp arrived with a cracked base." Example output: "Thank you for letting us know, and sorry your lamp arrived damaged. You can choose a free replacement or a full refund. Reply with your choice and a photo of the base, and we will arrange it today. The Support Team"
Customer email: <email>{customer_text}</email>
Write the reply.
Testing it: run 12 real complaints, including one outside the 30-day window and one asking for a discount. Pass rules: policy stated correctly, under 120 words, one next step. If the out-of-window case fails, add an example that shows the right refusal. To check claims in any factual answer, use our guide to verifying AI citations.
Common mistakes
- Vague goals. "Make it better" gives the model nothing to aim at.
- Rules without reasons. Bare capital-letter commands generalize poorly compared with a one-line explanation.
- One example only. A single example invites copying its length and wording; use two to five varied ones.
- Mixing data and instructions. Untagged pasted text can be read as new instructions.
- Testing on one input. A prompt that works once can fail on the fifth case.
- Changing model without retesting. A new model version can change results; rerun your test set.
- Overloading one prompt. Split long jobs into steps such as extract, then draft, then check.
How we researched this
We read OpenAI's prompt engineering guide, Anthropic's prompt engineering overview and prompting best practices, Google's Gemini prompting strategies, and the GPT-3 and chain-of-thought papers on 17 October 2026. The checklist combines points the vendors document; the worked example is our illustration of those points. For choosing a model to prompt, see our AI chatbots ranking and the ChatGPT vs Claude comparison. If prompting is becoming your job, read the prompt engineer career guide, and for long documents, context windows explained.
Frequently asked questions
What is the most important part of a prompt? A clear goal with a testable definition of success. Without it, you cannot tell whether a change helped.
How many examples should a prompt include? Usually two to five varied examples. Google recommends always including few-shot examples, and variety stops the model copying one example's quirks.
Should I still write "think step by step"? Rarely on current reasoning models. Raise the model's effort or thinking setting instead, and ask for visible reasoning only when you need to check it.
Do prompts work the same across ChatGPT, Claude and Gemini? The core techniques transfer, but details differ. OpenAI notes even different snapshots of one model can respond differently, so retest when you switch.
Is it better to use capital letters for important rules? Explaining the reason usually works better. Anthropic's guide labels a rule with its reason as more effective than a bare "NEVER" instruction.
Where should the question go in a long prompt? At the end, after the documents. Anthropic and Google both recommend this order for long inputs.
How do I know if my prompt is good enough? Run it on 10 or more real cases and count passes against rules you wrote first. Fix the most common failure, then rerun.