A useful AI prompt is a testable brief: it states the outcome, supplies the information the model needs, defines important constraints and explains what a satisfactory answer must contain. It does not need a magic phrase, a fixed length or an elaborate persona.
This guide was audited against current first-party guidance from OpenAI, Anthropic and Google on 12 August 2026. Prompt behaviour varies by model, version, interface and task, so the techniques below are starting points to evaluate—not guarantees of accuracy.
Decision Summary
Start with the smallest prompt that clearly defines the outcome and acceptance criteria. Add context, evidence, examples or structure only when the task needs them. Test revisions on the same representative cases before concluding that a technique works.
- Specify: outcome, audience, relevant inputs, constraints, evidence and output format.
- Verify: facts, calculations, citations and high-impact recommendations outside the model.
- Measure: correctness, completeness, format compliance, latency and cost across multiple cases.
The Prompt Framework
| Prompt element | Question to answer | Use when |
|---|---|---|
| Outcome | What must the response accomplish? | Every task |
| Context and inputs | What information may the model rely on? | The task depends on private, current or supplied material |
| Constraints | What boundaries, exclusions or approval rules apply? | Length, policy, risk or scope matters |
| Evidence | What sources, calculations or uncertainty must be shown? | The answer must be checked or defended |
| Output contract | Which fields, order, length and style are required? | The response feeds a workflow or publication |
| Success criteria | How will a reviewer decide whether it passed? | Repeated or high-value work |
Start with the Outcome and Success Criteria
Begin with the job to be done, not a persona. “Review this contract” is underspecified; “identify clauses that create renewal, termination or data-retention risk, cite each clause and separate facts from questions for counsel” gives the model an observable target. For a simple task, that may be the whole prompt.
Success criteria make revision possible. A support reply might need to answer every customer question, avoid promising a refund, stay below 180 words and finish with one next step. A research summary might need direct source links, dates, uncertainty labels and a list of unsupported claims. If the criteria cannot be checked, it is difficult to tell whether a longer prompt is actually better.
OpenAI's current model guidance recommends outcome-focused prompts that state the goal, relevant context, constraints, required evidence, success criteria and output format. It also advises keeping prompts lean and removing repeated instructions one group at a time while rerunning the same evaluations.
Provide Context, Evidence and Boundaries
Give the model the information that changes the answer: source material, audience, date, jurisdiction, product rules, brand examples or prior decisions. Separate supplied material from instructions with headings or tags so quoted text is less likely to be mistaken for a command.
More context is not automatically better. Irrelevant documents, duplicated rules and contradictory examples consume attention and tokens. Tell the model which sources are authoritative, whether outside knowledge is allowed and what to do when the supplied evidence is insufficient. For time-sensitive work, require current sources and source dates rather than assuming the model already knows the latest facts.
Define boundaries in operational terms. “Do not make external changes” is clearer than “be careful.” For an agent, distinguish safe reads and local validation from purchases, publication, deletion or messages to other people. For factual work, require the model to label inference and avoid inventing missing values.
A bounded example
Weak: “Write an email about our new reporting feature.”
Stronger: “Draft an announcement email for existing UK small-business customers. Use only the feature notes below. Explain the customer benefit without claiming time savings we have not measured. Include a subject line, 120–160 words and one CTA. Success means every factual claim is supported by the notes and the CTA is unambiguous.”
Use Examples, Roles and Structure Selectively
Examples
Examples are valuable when they encode a hard-to-describe format, tone, classification boundary or edge case. Make them representative and internally consistent. Do not add examples merely to make a prompt look sophisticated: they increase input size and can cause the model to copy accidental patterns.
Google's current prompt design guidance recommends clear instructions, context, constraints and consistent few-shot formatting. For complex machine-readable responses, it points developers towards structured-output features rather than relying only on prose instructions.
Roles
A role can focus perspective or tone—such as “review as a UK data-protection officer”—but it is not mandatory. Prefer a specific role tied to the task over grand claims such as “you are the world's best expert.” Always include the actual evidence and rules the role needs; a persona does not grant current knowledge or professional authority.
Structure
Use Markdown headings, XML-style tags or clear delimiters when a prompt mixes instructions, reference material, examples and variable input. Anthropic's prompting best practices specifically recommend clear, direct instructions and XML tags for separating complex prompt components. Simple questions do not need ceremonial markup.
Ask for Verifiable Work, Not Hidden Reasoning
Do not treat a request to “show your chain of thought” as a reliability control. Current reasoning models may reason internally, and a fluent explanation can still be wrong. Ask for outputs you can inspect: a concise rationale, assumptions, calculations, direct quotations, citations, test results or a checklist showing how the answer meets the criteria.
For difficult tasks, request the outcome and validation rather than prescribing every mental step. For example: “Calculate the total, show the formula and inputs, then independently check the result.” For research: “Cite a primary source for every time-sensitive claim and list any claim you could not verify.” For code: “Implement the change, run the named tests and report failures accurately.”
Break a workflow into stages when later work depends on an earlier decision or when a human needs to approve a boundary. However, splitting every request into many conversational turns can add latency, cost and opportunities for context drift. Choose the simplest workflow that passes the evaluation.
Test and Improve Prompts
A prompt that succeeds once may fail on a different input. Build a small evaluation set containing normal cases, difficult cases and inputs that should trigger uncertainty or refusal. Run the same cases against each prompt version and, when relevant, each model or reasoning setting.
- Define the rubric. Score factual correctness, completeness, evidence, format, tone and boundary compliance.
- Establish a baseline. Save the current prompt, model, settings, outputs, latency and token usage.
- Change one thing. Add an example, clarify a constraint or remove repeated instructions.
- Rerun the same cases. Compare the whole set, not the best output.
- Review failures. Decide whether the cause is the prompt, missing context, model choice, retrieval, tools or the task design.
- Version the winner. Record the prompt, model and date so future model changes can be retested.
Anthropic's prompt engineering overview begins with success criteria and empirical tests, and notes that some quality, latency or cost problems are better solved by model selection than prompt changes. That is a useful discipline across providers.
Reusable Prompt Template
Outcome
[State the result the response must achieve.]
Context and inputs
[Provide only relevant material. Name authoritative sources.]
Constraints and boundaries
- [Required scope, date, audience or policy]
- [Actions or claims that are not allowed]
- [What to do when evidence is missing]
Evidence and validation
[Specify citations, calculations, tests or checks.]
Output
[Required format, fields, order, length and tone.]
Success criteria
[List the conditions a reviewer will check.]Delete any heading the task does not need. For a one-line factual question, the outcome alone may be enough. For a production workflow, keep stable policy and tool rules at the system or application layer and pass task-specific inputs separately. The exact instruction hierarchy and persistence behaviour depend on the provider and interface.
Sources and Limitations
This audit used first-party guidance available on 12 August 2026: OpenAI's current model and prompting guidance, Anthropic's prompt engineering overview and Claude prompting practices, and Google's Gemini prompt design strategies.
Vendor guidance describes intended product behaviour and recommended starting points. It does not prove that one technique improves every task. Interfaces, instruction hierarchies, context handling, model controls and recommended practices change; retest material workflows after a model or platform update.
Frequently Asked Questions
What should a good AI prompt include?
Include the desired outcome, relevant context, hard constraints, required evidence or inputs, success criteria and output format. Omit sections that do not help the model complete the specific task.
Does every prompt need a role or persona?
No. A role can help when domain perspective or tone matters, but it is optional. A concrete task, relevant context and acceptance criteria are usually more important than an elaborate persona.
Should I ask an AI model to show its chain of thought?
Do not depend on hidden chain-of-thought as evidence. Ask for a concise rationale, calculations, cited sources, assumptions or verification steps that you can inspect, while allowing the model to reason internally.
How long should an AI prompt be?
There is no reliable universal length. Use the shortest prompt that includes the information and controls needed for the task. Longer prompts can help when they add relevant context, but repetition and conflicting instructions can reduce quality and increase cost.
How do I know whether a prompt is better?
Define measurable success criteria and test prompt versions on the same representative cases. Compare correctness, completeness, format compliance, evidence quality, latency and cost rather than judging one impressive response.
Do prompt techniques transfer across ChatGPT, Claude and Gemini?
Core practices such as clear goals, relevant context, examples and explicit output requirements transfer, but model behaviour and platform controls differ. Check current vendor guidance and retest prompts when changing models or versions.