“Prompt” is commonly used to mean an instruction sent to AI. Some advice makes prompts sound like spells: find the right template and you will always get a good answer. In practice, quality usually comes from explaining what you need, providing relevant information, and reviewing the output.
A short instruction can work well for a simple task. A long one can still fail if the goal is missing or unrelated requests are mixed together. Think of a prompt as a handoff written for a colleague who cannot know what you are thinking.
1. State the goal
Say what you want to accomplish and where the result will be used. “Write about the product” leaves out the purpose. “Draft an email to current customers explaining a change and where to find instructions” is more specific.
A goal helps AI choose detail, structure, and tone. If the goal is unclear, ask AI to clarify before producing the output.
2. Provide enough context
Context may include the audience, situation, success criteria, confirmed decisions, or assumptions to avoid. You do not need to explain everything—only what could change the answer.
For an announcement, say what recipients already know, what is changing, and whether the effective date is confirmed. If the date is unsettled, say so; do not leave AI to invent one.
3. Specify data and sources
If you want AI to work from particular documents, identify which ones are authoritative. If sources conflict, explain how to handle that. If you do not want external information added, say so.
For example: “Use only the attached meeting record. If it does not name an owner, write ‘not specified’; do not infer one.” This sets a boundary that can be checked.
4. Set limits
Limits may cover length, scope, tone, data that must not be included, or actions that are not allowed. “Do not make things up” is a good intention but difficult to check by itself. Make it concrete: “Do not create figures; mark [needs verification] wherever information is missing.”
Place important constraints before AI starts. If there are several, use separate lines instead of compressing them into one long sentence.
5. Describe the output
Specify a format to reduce rework: which columns a table needs, sections in a memo, how many categories, or who the answer is for. Choose a format that helps you use the result, not one that merely looks complete.
Instead of “analyze the feedback,” try: “Create a table with theme, representative comment, confidence, and question to verify. Keep quotes short; do not infer prevalence.”
6. Add an example when the standard is hard to describe
An example can show the desired level of detail, format, or tone. Include what you like and what to avoid if the distinction matters.
An example is not a universal rule. A sample may contain inaccurate information or be unsuitable for a new situation, so say whether it demonstrates style or content.
Put it together
For summarizing notes, you could write:
Goal: Prepare a briefing so the product team can decide what to do next.
Context: These notes come from three interviews; they do not represent all customers.
Data: Use only the attached files.
Limits: Do not infer frequency from examples; flag contradictions.
Output: Three issue groups, each with a quote, what is known, and an open question.
After the first draft, give focused feedback: “Keep the first group; split the third because its items lead to different decisions.” This is often more effective than rewriting the whole prompt.
Revise based on observed errors
If the answer is too long, set a length or ask it to prioritize decision-relevant points. If it guesses, ask it to separate sourced information from unknowns. If the tone is off, provide a short example and explain what to follow.
More rules do not always help. Each condition should address a specific need or risk. Try the revised instruction on similar input to see whether it improved the result or created a new issue. For complex work, you may first ask AI to summarize its understanding and assumptions; for a small task, that extra step may be unnecessary.
When should AI ask a question first?
If an important decision is missing, ask AI to list what it needs before proceeding. That is useful when different interpretations would produce different results.
For low-risk tasks, it may be faster to ask AI to state assumptions and draft an answer. Either way, confidence in the wording is not proof that the system understood correctly.
Key takeaway: There is no magic wording. State the goal, context, sources, constraints, and desired output, then treat the result as a draft to evaluate.
In summary
A good prompt is a clear handoff. The right level of detail depends on complexity and risk: the more conditions or consequences involved, the clearer the sources, assumptions, permissions, and review criteria should be.
Start with a real task, revise based on observed errors, and save a template only after trying it in the situations where it is meant to help.