AI Prompt Builder

Choose cloud model or local model. Fill in the fields. Get one complete prompt.

The newest reasoning models from OpenAI, Anthropic and Google want the same thing from a prompt: the goal, not the steps. A small model on your own computer wants the opposite. This tool builds the right prompt for each.

Method checked against the vendors' own prompting guides on November 11, 2026. The sources are at the end of the page.

Cloud models: start with the goal

A cloud model runs on the vendor's computers. Claude, Gemini and GPT are cloud models. Their newest versions are reasoning models: they plan the steps themselves before they answer. So your prompt gives the result you want and the conditions for success. The model selects the steps.

PartWhat it doesExample
GoalThe result you want, and the reason you want it.Turn my rough draft into an issue that a busy manager can read in three minutes. Readers stop at long paragraphs, so short and clear matters more than complete.
ContextThe facts the model cannot know: your situation, your audience, your source text.I run a one-person consulting business. The readers are managers aged 40 and over. My draft is below. [paste the draft]
Done whenWhat must be true before the model gives you the answer.Each paragraph makes one point. Every example from my draft is still there. No fact is new.
LimitsWhat the model must not change, and what it must ask you first.Keep my wording where it is clear. Add no statistics. If a sentence is unclear, mark it and ask me.
OutputThe form of the answer: length, structure and tone.About 400 words in plain paragraphs, with one subheading per section. Then list your changes in three bullets.
Role (optional)One sentence on the job the model does. All three vendors keep it short.You are an editor for a business newsletter.

A complete cloud prompt

Role:
You are an editor for a business newsletter.

Goal:
Turn my rough draft into an issue that a busy manager can read in three minutes. Readers stop at long paragraphs, so short and clear matters more than complete.

Context:
I run a one-person consulting business. The readers are managers aged 40 and over. My draft is below.
[paste the draft]

Done when:
Each paragraph makes one point. Every example from my draft is still there. No fact is new.

Limits:
Keep my wording where it is clear. Add no statistics. If a sentence is unclear, mark it and ask me.

Output:
About 400 words in plain paragraphs, with one subheading per section. Then list your changes in three bullets.

Three rules for a cloud prompt

  • Give the goal, not the steps. List steps only when the order matters to you.
  • Do not write "think step by step". A reasoning model already does this. For a harder task, raise the thinking or effort setting.
  • Put a long document first. Anthropic and Google both say to put long source text before the question. Anthropic reports up to 30 percent better answers in its tests with the question last.

What OpenAI, Anthropic and Google say

The three guides cover GPT-5.6 Sol, the current Claude models up to Claude Fable 5.2, and Gemini 3 and later. They agree on most points.

AdviceOpenAIAnthropicGoogle
State the result"Describe the destination rather than prescribing every step.""Prefer general instructions over prescriptive steps.""Be precise and direct." Leave out persuasive language.
Say when the work is doneThe suggested prompt has "Success criteria" and "Stop rules" sections.Give clear success criteria. On most models, ask for a check against your criteria before the answer.The sample instruction has a "Validate" step: review the output against the task.
Give context and the reasonGive the important constraints and the available evidence.Explain why an instruction matters. The model then applies it to cases you did not list.For long input, give the context first and ask the question last.
Set the outputAn "Output" section: sections, length, format and tone.Be specific about the output format and the constraints.Answers are short by default. Ask for detail when you want it.
Keep the role shortA "Role" line: the function of the model and its context.One sentence in the system prompt makes a difference.Put the role first, or in the system instruction.
Drop old habitsRemove repeated rules and steps the model already does. Keep absolute words for true hard rules.Forceful wording can make new models overreact. Use calm, direct wording.Explicit reasoning steps in the answer are generally unnecessary.

The guides differ in two places. Google still recommends examples in every prompt, and its sample prompt still ends with a "think step-by-step" line. Anthropic says to remove self-check lines for Claude Opus 5, because that model already checks its work. Test a prompt on your own model before you depend on it.

Local models: be explicit and give an example

A local model runs on your own computer, for example through Ollama. It is much smaller than a cloud model, so it knows less and reads less text in one prompt. Google's web.dev guide calls any model under 30 billion parameters a smaller model. Its advice is the reverse of the cloud advice: write a more detailed prompt, give clear format instructions and add examples.

PartWhat it doesExample
RoleOne short sentence. A small model ignores parts of a long instruction.You are an editor for a business newsletter.
ContextOnly the text the task needs. A local model reads much less text in one prompt.Paragraph: "Our Q3 workshop series, which ran for six weeks and had 48 people, went well and most of them said they would come back, so we plan to repeat it in spring."
Task and stepsOne task, with the steps in order.1. Read the paragraph. 2. Rewrite it in at most three short sentences. 3. Keep every number.
ExampleOne or two samples of a correct answer.Input: "The report, which took two months and covers 12 clients, is done and we will send it Friday." Answer: "The report is done. It took two months and covers 12 clients. We will send it Friday."
FormatThe exact form of the answer, as the last line.Only output the rewritten paragraph. No introduction.

A complete local prompt

Role:
You are an editor for a business newsletter.

Context:
Paragraph: "Our Q3 workshop series, which ran for six weeks and had 48 people, went well and most of them said they would come back, so we plan to repeat it in spring."

Task and steps:
1. Read the paragraph.
2. Rewrite it in at most three short sentences.
3. Keep every number.

Example:
Input: "The report, which took two months and covers 12 clients, is done and we will send it Friday."
Answer: "The report is done. It took two months and covers 12 clients. We will send it Friday."

Format:
Only output the rewritten paragraph. No introduction.

Three rules for a local prompt

  • One task per prompt. Divide a large job into small prompts.
  • Steps and examples help. The web.dev guide got more accurate answers with examples and step-by-step reasoning.
  • Check each answer. A small model breaks the format more often, so read the result before you use it.

Cloud prompt and local prompt compared

QuestionCloud reasoning modelLocal model
What comes first?The goal and the conditions for success.One task with numbered steps.
Steps?Leave them out, unless the order matters.Write them out.
Examples?Optional. Add them to fix the format or the tone.Add one or two.
How much context?As much as the task needs. Long documents go first.Only what the task needs.
Format line?Length, structure and tone.The exact text to print, as the last line.

How to use the result

  1. Paste the prompt into a new chat, so no earlier conversation changes the answer.
  2. Read the answer against your "Done when" line or your Format line. If it fails, tell the model which condition it missed.
  3. When the prompt works, save it. A tested prompt is worth more than a new one.

For definitions, see prompt, reasoning model, system prompt and context window in the AI Encyclopedia.

Sources