Treat a prompt as a short brief
A better prompt does not need secret wording or an elaborate fictional persona. It needs a clear task, relevant context, and a description of a usable result. Most improvements come from removing ambiguity and checking the response, not from making instructions longer.
Think of ChatGPT as a collaborator receiving a brief without the background in your head. If you ask “make this better,” it must guess whether you mean shorter, more persuasive, more accurate, or easier to understand. Those guesses often explain why an answer looks polished but misses the purpose.
Define success before writing
Write a one-sentence outcome: “I need a customer email explaining a delayed shipment without promising a delivery date.” Then identify the audience, available facts, and things the answer must not invent.
For this example, the audience is a customer waiting for an order. Known facts are that dispatch is delayed and a status update will follow tomorrow. Unknown facts include the arrival date and whether compensation is available. A useful answer must preserve those boundaries.
OpenAI's guidance emphasizes clarity, specificity, and iterative refinement. Its API guidance also recommends separating instructions from context and showing the requested output format. These principles are useful in ordinary chat as well. 1 2
Use five practical ingredients
| Ingredient | Question it answers | Example |
|---|---|---|
| Task | What should happen? | Draft a delay notification |
| Context | What does the assistant need to know? | Order has not shipped |
| Audience | Who will use the result? | A retail customer |
| Constraints | What boundaries apply? | Do not invent dates or refunds |
| Output | What should the result look like? | Subject plus a short email |
Not every prompt needs all five ingredients spelled out. A simple request can remain simple. Add detail when it changes the result, and remove instructions that merely sound impressive.
Rewrite a vague prompt
Weak prompt: “Write a professional email about a delay.”
Stronger prompt: “Draft a customer email about a delayed shipment. Facts: the order has not dispatched; we will provide another update tomorrow. Do not invent an arrival date, cause, tracking number, or refund policy. Use a calm, direct tone. Return one subject line and an email under 120 words.”
The stronger version limits the space for invention and makes review easier. You can now ask whether the draft includes every supplied fact, avoids unsupported claims, and fits the requested length. “Professional” alone would not give you those checks.
If essential information is missing, allow clarification: “Ask up to three questions if missing information would materially change the answer. Otherwise proceed and label assumptions.” This is more useful than inviting unlimited questions for a straightforward task.
Supply clean context
Separate the material to analyze from the instructions. For example, introduce a passage with “Source text begins” and end it clearly before adding your request. Tell the assistant whether to use only that material or whether external research is permitted and available.
When summarizing a document, say: “Summarize only the supplied excerpt. Do not infer what omitted sections contain. List any unanswered questions separately.” A summary should not quietly become a speculative completion.
Remove private information that is unnecessary for the task. Replace real customer details with placeholders when drafting a reusable message. Do not upload confidential material merely because a prompt template recommends providing more context. Relevant context and excessive disclosure are different things.
Show an example when format matters
If you need structured output, provide a small example using invented data. For a task tracker, specify fields such as task, owner, and due_date, and explain how missing information should appear. Without a missing-value rule, the assistant may fill gaps with plausible guesses.
A useful instruction is: “Return a table with columns Claim, Supporting passage, and Uncertainty. If the passage does not support a claim, write ‘Not supported in supplied text.’” That format encourages visible evidence rather than a confident narrative.
For machine-readable output, ask for the exact format and validate it with an appropriate parser. A request for JSON is not proof that the result is valid JSON, nor that its values are correct. Check structure and substance separately.
Break complex work into stages
For a substantial guide, first agree on the audience and scope. Then request an outline, identify missing evidence, draft a section, and review factual claims. Producing everything in one request can hide errors inside a large, fluent answer.
For planning, distinguish ideas from commitments. Ask the assistant to propose options, identify dependencies, and mark estimates as estimates. Do not let an invented deadline become an operational promise simply because it appears in a well-formatted plan.
You can request a concise explanation of assumptions and the checks used without asking for private internal reasoning. What matters is an auditable result: sources, calculations, limitations, and decisions you can examine.
Improve with targeted feedback
After reading the first output, identify the specific failure. “Too vague” is less useful than “Replace general advice with three actions a one-person shop can complete this afternoon.” “Make it warmer” works better when paired with “Keep the factual boundaries and do not add an apology that admits legal responsibility.”
Change one or two important variables at a time. If you simultaneously alter the audience, length, tone, and objective, it becomes difficult to tell which change helped. Save successful prompts with a note about the task they were designed for.
For recurring work, keep a tiny evaluation set: three realistic inputs and a checklist of acceptable behavior. Include an incomplete input to test whether the assistant invents missing facts. A reusable prompt is valuable only while it continues to pass those checks.
Know what prompting cannot solve
Clear instructions cannot guarantee factual accuracy, access to a source, current information, or compliance with a specialized legal requirement. Ask for uncertainty, but verify consequential claims independently. A longer prompt does not convert an unverified answer into evidence.
If the free tier lacks a tool you need, redesign the task around supplied material or use another appropriate resource. Compare options through ChatGPT Plus alternatives instead of assuming a paid subscription automatically solves the problem.
Before using the answer, check purpose, facts, privacy, format, and next action. The best prompt is not the one that sounds most technical. It is the one that reliably produces a useful draft you can inspect, correct, and responsibly use.