Prompt

This is the one part of the whole business you control completely, and the one skill on this list you can practise today without buying anything. The gap between a mediocre answer and a genuinely useful one is far more often the request than the software. The same assistant, on the same subject, will produce a shrug or a finished draft depending on how much of what you already know made it into the message. It is also the cheapest lever available. Better wording costs nothing, whereas a bigger model costs on every request forever.

Everything in the request is read as one continuous piece of text, then continued. There is no separate channel for instructions and material: your question, the document you pasted, the examples you gave and any rules the product set behind the scenes are all part of the same block. That is why a long attachment can drown a short instruction, and why putting the instruction after the material often works better than putting it before.

Specificity does most of the work. Naming the audience, the format, the length and the purpose removes guesses the assistant would otherwise make silently. Compare asking for a summary with asking for five bullet points a finance director could read in a minute, each naming the decision it affects. The second describes a finished thing; the first describes a direction.

Showing beats describing. Two or three examples of the output you want will usually outperform a paragraph explaining it, because the pattern is unambiguous where the description is open to interpretation. This is why pasting a previous piece of work and asking for another like it is such a reliable move.

Giving the assistant something to work from changes the failure mode. Asked about your refund policy with nothing attached, it will produce a plausible refund policy. Asked with the actual policy pasted in, it will answer from the document. Nothing about the assistant changed; the material available to it did.

Constraints are worth stating even when they feel obvious. Say what to leave out, what not to invent, and what to do when the answer is not in the material. A request that ends with an instruction to say so rather than guess when something is missing produces noticeably fewer confident inventions than one that does not.

Order matters more than it looks. What comes last tends to carry the most weight, which is why an instruction placed after a long attachment is followed more reliably than the same instruction placed above it. If a request contains several demands, the ones at the very end are the ones most likely to survive a long answer intact, so put the non-negotiable requirement there rather than in the middle.

Iterating is part of the method, not a sign the first attempt failed. Saying what is wrong with an answer gives the assistant your intent and a worked example of missing it, which is more information than the original request contained. Three short corrections usually get further than one heroic opening request, and they cost less thinking on your part because each one addresses something you can now see.

• Paste a rejected draft and the feedback it received, then ask for a revision that answers each point. The feedback is doing more work than any instruction you could write.

  ChatGPT

• Attach a long source document and put your instruction after it rather than before, so the request is the last thing read rather than the thing buried at the top.

  Claude

• Ask a research assistant a question worded as the sentence you want back, and see how much the shape of the question shapes the shape of the answer.

  Perplexity

• Save a request you have refined into a reusable instruction inside your notes workspace, so the wording that worked is not rebuilt from memory every time the task recurs.

  Notion AI

What makes a prompt good?

Enough context that the assistant is not guessing, a clear statement of what the output should be, and an example where one is available. Naming the reader, the format and the length removes most ambiguity. A useful test: if you handed the same text to a capable colleague with no other briefing, could they produce what you want?

Should I use a prompt template someone else wrote?

As a starting point, yes, particularly for a task you do repeatedly. Read what it actually asks for rather than pasting it blind, because most of a template's value is the specificity it supplies and specificity about someone else's job may not suit yours. Keep the structure, replace the substance.

Is it better to ask for everything at once or in steps?

Steps, for anything with several stages or where you will want to intervene. Asking for an outline, correcting it, then asking for the draft gives you a checkpoint before the expensive part. Single requests suit well-defined jobs where you can describe the finished thing precisely.

Does it matter which assistant I write the prompt for?

Less than people expect. The habits that help, context, examples, explicit constraints, transfer across products because they address the same underlying mechanism. Fine details differ, and a request tuned heavily to one product may need adjusting elsewhere, but a well-specified brief travels.

Do I need to be polite to an AI assistant?

It makes no difference to the mechanism, which has no feelings to bruise. Courtesy costs a few words and some people find it keeps their own tone consistent, which has its own value when the text may be read by a colleague later. Nothing about please or thank you reliably improves the answer itself.

Why did the same prompt give me a different answer the second time?

Most assistants introduce deliberate variation, so identical requests produce differently worded replies even when the substance holds. Where consistency matters, some products expose a setting that reduces it, and pinning the format in the request helps too. Otherwise treat the variation as a reason to review output rather than to expect one canonical answer.

• The prompt is everything you send in one request, instruction and attached material together.
• Specificity does the work: name the audience, the format, the length and the purpose.
• Showing two or three examples usually beats describing what you want.
• Attaching the relevant material changes the failure mode from invention to reading.
• Treat the first answer as a draft and correct it rather than rewriting the request.

• ChatGPT

  A ready place to practise this, where iterating on a draft takes seconds.

• Claude

  Large attachment capacity, so the instruction-after-material habit matters more here.

• Perplexity

  Research answers track the shape of the question closely, which makes wording effects visible.