keepcalculator

06 Numbers and Powers

How to Talk to AI: A Guide to Better LLM Prompts

Card No. 06·78324 · Updated · Standard formula, shown below

Clear prompts help a language model produce answers that are more relevant, usable, and appropriately detailed. You do not need special software or technical knowledge: start with a task, give the context the model needs, and refine the result if it misses the mark.

What You’ll Need

  • Access to a language-model chat or writing tool.
  • A clear goal, plus any source material or relevant details the answer should use.
  • A way to review the output, especially if accuracy matters.

Step-by-Step

  1. Decide what a successful answer looks like. Before typing, name the task: explain a concept, summarize a document, draft an email, compare options, or generate ideas. Think about what you will do with the result. “Help me with my presentation” is broad; “Create a five-slide outline for a 10-minute presentation introducing recycling to a middle-school class” gives the model a target. If you are unsure, ask it to help identify the audience, scope, or questions that need answering.
  2. Give the essential context. Include information that would change the answer: intended audience, purpose, subject, relevant background, and constraints. For example, say whether the writing is for customers or coworkers, whether the reader is new to the topic, and what facts or source text to rely on. Do not assume the model knows your private circumstances or has access to files, websites, or earlier conversations unless you provide them or the tool explicitly supports that access.
  3. State the task with a direct verb. Use an instruction such as “summarize,” “rewrite,” “compare,” “classify,” or “draft.” Specify the material the instruction applies to. For a summary, say which text to summarize; for a rewrite, paste the text and explain what should change. A direct request reduces ambiguity and makes it easier to spot when the response has gone off course.
  4. Set useful boundaries. Explain what to include and avoid. Boundaries might cover length, tone, reading level, format, date range, or which sources to use. Ask for a table if you need comparable columns, numbered steps for a procedure, or plain language for a general audience. Keep constraints realistic and prioritize the ones that matter most; a long list of competing requirements can make the task harder to interpret.
  5. Show a model or structure when precision matters. If you have a preferred style, provide a short example and say what to imitate, such as its level of formality or sentence length. You can also give a simple template: “Issue: … / Cause: … / Next step: …”. Label examples as examples so the model does not mistake their details for facts to reuse. Avoid sharing confidential or personal information unless the tool and your organization’s rules permit it.
  6. Ask for an appropriate response format. Tell the model how to organize its answer and what level of detail you want. For instance: “Give me three options, each with one advantage and one drawback,” or “Return a short email with a subject line.” If you need assumptions identified, request a separate assumptions list. If you want sources, ask for verifiable references and check that they support the claims; a request for citations does not guarantee that citations are genuine.
  7. Review the first answer against your goal. Check whether it followed the task, used the supplied context, respected limits, and addressed the audience. Treat the output as a draft, not proof. Verify important facts, calculations, dates, quotations, and citations using reliable sources, particularly for consequential decisions. If the answer is useful but incomplete, keep the useful parts and specify the gap rather than restarting without explanation.
  8. Refine with one clear follow-up at a time. Point to the exact change: “Keep the first two recommendations, but add a low-cost option and explain the trade-off in one sentence.” If the answer is too broad, narrow the scope; if it is too long, set a target length and say what to retain. For a complex job, split it into stages—for example, outline first, review the outline, then draft—so you can correct direction early.

Common Mistakes

  • Being vague: “Write something good” leaves quality, audience, and purpose undefined. State what the output is for and what it should accomplish.
  • Combining unrelated tasks: Asking for research, a polished article, a translation, and a critique in one tangled request can lead to omissions. Separate distinct tasks or label each deliverable.
  • Giving conflicting instructions: “Be exhaustive” and “use no more than 50 words” may not both be possible. Decide which requirement takes priority.
  • Assuming the model knows hidden context: Previous chats, local files, current events, and personal preferences may not be available. Include the necessary material and specify a relevant time period.
  • Accepting confident wording as evidence: Fluent responses can still be mistaken, incomplete, or invented. Check facts and ask for uncertainty or assumptions to be made explicit.
  • Sharing more data than needed: Remove passwords, account details, confidential records, and identifying information unless you have a legitimate reason and permission to share them.

Tips

  • Put the main request near the beginning, then add context and constraints.
  • Use concrete details, such as the audience, intended use, and a specific length or format, instead of subjective words like “better.”
  • For documents, name the text clearly and separate it from instructions with labels such as “Source text” and “Task.”
  • Ask the model to list assumptions or ask you clarifying questions when missing details could change the result.
  • Request alternatives when there is no single best answer, and ask for the trade-offs behind each option.
  • For recurring work, save a reusable prompt template, but update its context and check that its instructions still fit each task.

Troubleshooting

The answer is too general: Add the intended audience, situation, and a concrete example. Ask for a narrower scope or specific points to cover.

The format is wrong: State the required structure explicitly and ask for a revision in that format. For example, specify headings, number of bullets, or table columns.

The response ignores a constraint: Repeat the highest-priority constraint in a short follow-up, explain what it violated, and ask for a corrected version. If several constraints conflict, simplify them or rank them.

It includes unsupported claims: Ask which statements rely on supplied material and which are assumptions. Check key claims independently; remove anything you cannot verify. Do not treat invented or uncheckable references as reliable evidence.

You do not know what to ask: Describe your goal and ask what information is missing. The model can suggest clarifying questions, but you decide which answers are accurate and appropriate to share.

Frequently Asked Questions

Do I need to use special prompt words?
No. Plain, direct language is usually enough. Specific context and clear instructions matter more than magic phrases.

How long should a prompt be?
Long enough to include the task, relevant context, and important constraints. Remove details that do not affect the answer, but do not omit information the model needs.

Should I ask the model to think step by step?
You can ask for a concise explanation, assumptions, or a checkable outline. For many tasks, a clear result and brief rationale are more useful than a long internal-style reasoning trace.

Can I trust an answer if it sounds certain?
Not automatically. Verify important claims, calculations, and sources. Confidence and polished writing do not establish accuracy.

Is it safe to paste a document into a prompt?
Only if you are allowed to share it with that service. Check applicable workplace or service rules, and remove sensitive details that are not necessary for the task.

What if the model gives different answers each time?
Make the instructions and inputs more specific, then ask for the same format and criteria. Some variation is normal, so compare the answers and verify facts rather than assuming one is correct.

Related

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.

More cards in this drawer

The whole Numbers and Powers drawer →