A good AI prompt tells the model what to do, what it needs to know, and what a useful answer should look like. The reliable way to improve results is not to hunt for a hidden “secret,” but to define success, try a clear first draft, and revise it based on what the model actually returns. These habits apply across generative AI tools, though results can vary by model and version.
1. Name the task
Start with a direct action: summarize, compare, explain, classify, or draft. A model can respond to a question, perform an operation on a subject, or continue a supplied passage; make clear which kind of response you want. Google’s Gemini prompt design guidance recommends clear, specific instructions.
Instead of “Electric cars and hybrids,” try “Compare electric cars and hybrids for a driver who mostly commutes in a city.” The verb tells the model what to do; the rest narrows the job.
2. Say what success looks like
Include the purpose of the answer: who will use it, what they need to decide, or what problem they are trying to solve. “Explain cloud storage” could mean a beginner’s overview, a security briefing, or a buying comparison. “Explain cloud storage to a first-time user deciding whether to back up family photos” gives the response a clearer target.
#1 Best Overall
Anthropic recommends defining success criteria before drafting and testing a prompt. If you have not decided what a good answer must achieve, it is difficult to tell whether the model has done the job.
3. Provide the context the model needs
Supply relevant source material, facts, audience details, or constraints instead of expecting the model to infer them. If you want a summary of a report, include the report or the portion to summarize. If you want help replying to an email, provide the message and any relevant relationship or background.
Keep context focused. Extra material that does not bear on the task can make the request harder to follow. When working with an incomplete sentence or other partial input, explain whether the model should continue it, analyze it, or transform it; Google notes that context and examples can influence how a model continues a prompt.
4. Specify the constraints that matter
State meaningful boundaries directly: scope, length, tone, exclusions, or required content. For example: “Draft a polite reply under 150 words, acknowledge the delay, and do not promise a delivery date.” Such instructions steer the response, but they are not guarantees that every model will comply perfectly.
Avoid piling on arbitrary restrictions. Include a constraint when it helps make the answer more useful or easier to assess.
5. Add an example when words alone leave room for guesswork
If a desired pattern is hard to describe, show a short example of the input and the kind of output you want. An example can clarify tone, labeling, or how to handle a particular case. Google’s prompt-design guide demonstrates examples and output prefixes for structured tasks.
Use examples that represent the cases the model is likely to encounter. A narrow or misleading example may steer the answer in an unhelpful direction; it should illustrate the pattern, not substitute for a clear task.
6. Ask for the output shape
Say whether the response should be a list, table, set of steps, or another format. If you need information to fit specific fields, name those fields. For instance: “Return a table with columns for option, benefit, drawback, and best fit.” That makes the result easier to scan and check.
For applications that require machine-readable output, prose instructions alone may not be enough. Google recommends using an API structured-output feature for complex structured responses rather than relying only on a prompt to enforce the format.
7. Break complex work into ordered steps
For a task with dependencies, spell out the sequence so the model can address each part in order. For example: “First extract the dates from the notes. Then arrange them chronologically. Finally, write a short timeline and flag any missing dates.” This makes the requested workflow explicit and helps you see where an answer went off track.
Decomposition is an organizational aid, not a guarantee of better results. For a straightforward request, one concise instruction may be clearer than a long checklist.
8. Use role and style cues only when they add useful context
A role cue can suggest a perspective or voice, such as “Explain this as a patient writing tutor.” But it cannot replace the task itself, the facts the model needs, or the limits on the answer. Anthropic includes role prompting among the techniques discussed in its prompt engineering overview.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsPair a role with concrete instructions: “As a patient writing tutor, identify the three clearest ways to improve this paragraph and give one example revision.” A role without a specific job leaves the model guessing.
9. Check the answer against your criteria, then revise
Read the result against the goal you set, not just for fluency. Check whether it is accurate against the source or intended outcome, complete, in the requested format, useful to the intended reader, and consistent across repeated runs if stability matters. These are practical evaluation questions, not a standardized benchmark.
When a response misses, identify the specific failure and revise the relevant part of the prompt. If it omitted a case, name that case; if it used the wrong audience level, specify the reader; if the format drifted, make the required structure more explicit. Google describes prompting as iterative: “These guidelines and templates are starting points. Experiment and refine based on your specific use cases and observed model responses.”
Anthropic likewise advises testing prompts empirically against defined success criteria. It also notes that rewriting a prompt is not always the best fix: choosing a different model may be a more direct way to address cost or latency.
Recommended Free Tools
10. Recheck prompts when the model changes
Do not assume a prompt will behave identically across AI products or model versions. OpenAI’s prompt engineering documentation says prompting behavior can vary between model snapshots and recommends pinned versions and evaluations when consistency matters in an application.
If a workflow is important, keep a set of representative inputs and criteria, then check the outputs after changing the model or version. A prompt that worked once is not evidence that it will keep working unchanged.
A compact prompt pattern to try
For many everyday requests, a useful starting point is: “Task: [what to do]. Context: [relevant material and audience]. Success: [what a useful answer must accomplish]. Constraints: [scope, tone, length, or exclusions]. Format: [how to present the answer].” Add an example or ordered steps only when they clarify the task.
This pattern is a starting point, not a universal template. Adjust it to the job, then judge the response by whether it meets the stated goal.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
