Prompt Writing 101: How to Get Better Answers From AI

Two people ask the same assistant to write the same email. One gets a bland template. The other gets a draft worth sending. The model was identical, and so was the subscription. The whole difference sat in the twenty seconds before they pressed enter.

This is unfamiliar ground for most of us. Two decades of search engines trained us to type fragments, such as “weather geneva sunday”. An assistant tolerates fragments, but it rewards a brief. Writing that brief has acquired a grand name, prompt engineering. There is nothing esoteric about it. It is ordinary delegation, applied to an unusually fast colleague.

What the model can and cannot see

An assistant answers from two sources only: the patterns it learned during training, and the text sitting in front of it right now. That second part is called the context window. The mechanism behind it is unpacked in how AI chat assistants work.

The practical consequence is blunt. The model does not know your job, your reader, your deadline or yesterday's chat. Your prompt is very nearly its entire world. Anything you leave out, it will fill in with the most average guess available.

That is the real cost of a vague question. Ask something loose and the model answers the average version of it. It produces the reply that would suit the largest number of people who might have typed those words. Averages are rarely what anyone wants. Specificity narrows the field until the most probable answer is also the useful one.

There is a quieter effect too. The model mirrors the register it is given. A rushed, fragmentary prompt tends to draw a generic, listicle-flavoured reply. A considered brief in full sentences draws something closer to considered prose. Your prompt sets the topic and the standard at once.

Treat an assistant like a brilliant colleague on their first day: capable of almost anything, aware of almost nothing about you.

The four parts of a working prompt

Most effective prompts are built from the same four components. Not every request needs all four. When an answer disappoints, though, the missing ingredient is almost always one of these.

Context: what the model cannot know

Context covers who you are, what the situation is, and any material to work from. Pasting the actual text beats describing it every time. Give it the email thread, the product description, the paragraph you dislike. One pasted paragraph does more good than a page of adjectives.

• Your situation: the role, the company, the constraint you are working under.
• The source material: the text, data or draft the answer should be built from.
• The reader: who receives the output and what they already know.
• Previous attempts: what you tried and why it did not work.

Keep context relevant rather than exhaustive. A long dump of unrelated material buries the parts that matter. Three sentences of situation and one block of source text is a good default.

Task: one verb, one job

State the task as a single instruction built around a verb. Summarise, rewrite, compare, list, draft, critique, translate, classify. If you notice an “and also” chain forming, split the work across separate turns. A model asked to do five things at once tends to do all five thinly.

Constraints: length, tone, format and limits

Constraints do the same job as a tolerance on an engineering drawing. They tell the maker what counts as acceptable, a principle we explore in how precision engineering controls quality. “Under 150 words, warm but professional, as a bulleted list, no jargon” beats any amount of vague hoping.

Say what to avoid as well as what to include. Banned words, sections to leave untouched and facts that must survive editing all belong here. Constraints are also how you keep an assistant out of territory it should not enter. That boundary gets fuller treatment in data, privacy and good AI habits.

Examples: show the shape you want

An example is the fastest instruction of all. Language models are superb imitators, and one sample of good output usually beats a paragraph describing it. Researchers call this few-shot prompting. Paste last month's version, then ask for this month's in the same shape.

Eight moves that reliably improve an answer

The anatomy tells you what to include. These eight moves tell you what to do when a reply is nearly right. They work across assistants, because they act on the same mechanism. Add one at a time and watch what changes.

1. Name the audience. “Explain contactless payment to a curious twelve-year-old” and “explain it to a payments lawyer” produce entirely different, entirely appropriate answers.
2. Show, do not only tell. Paste an example of the layout, tone or structure you want. Copying beats describing.
3. Ask for structure. Request a table, numbered steps or labelled sections. Structured replies are easier to scan and far easier to check.
4. Invite reasoning first. For anything multi-step, ask the model to work through the problem before stating a conclusion. Research on chain-of-thought prompting found that eliciting intermediate steps improves accuracy on hard tasks.
5. Set a length budget. Word or sentence limits prevent both padding and thinness. “In exactly three sentences” is clarifying for the model and for you.
6. Give the negative case. Name what a bad answer would look like. “Do not open with a definition” saves a whole revision round.
7. Iterate instead of restarting. “Halve the length, keep the second paragraph, make the ending more direct” gets further than a new mega-prompt.
8. Ask for alternatives. Request three versions in different tones, or ask the model to argue against its own recommendation. The first answer is the most probable one, not the best one.

Weak prompt, strong prompt

The fastest way to absorb all of this is to watch weak prompts turn into strong ones. Each rewrite below adds specificity, not length.

Notice what the sharper versions share. Each names a reader, supplies the raw material, bounds the output and asks for something checkable. None runs longer than three sentences. Precision here is a matter of decisions, not word count. The meeting example works especially well on a clean transcript, which is why our best practices for meeting transcription pair naturally with prompt craft.

When the answer comes back wrong

Most disappointing replies have a diagnosable cause. Two patterns account for the majority.

It ignored half of my instructions

This usually means the prompt asked for too much, or asked for two things that quietly conflict. Brevity and exhaustive detail cannot both win. Very long prompts also bury requirements in the middle, where they compete with everything else.

The fix is mechanical. Split the work into turns. Put the single most important constraint in its own short sentence. Then ask for a targeted revision rather than a fresh answer.

It sounds right but the facts look shaky

Fluency is a style, not evidence. A model can produce a smooth sentence and a wrong number in the same breath. Open a source you trust and confirm every name, date, figure and reference against it. The pattern behind those errors is set out in why chatbots make things up.

Other symptoms have equally simple remedies.

• The tone is off. Give a sample of the voice you want instead of adjectives such as “professional”.
• It forgot something you said earlier. Long chats push early messages out of view, so restate the key facts.
• The output is the wrong shape. Name the destination: a slide, a spreadsheet column, a sixty-second read-aloud.
• Every answer feels generic. You have described the task but not the situation, so add context.
• It refuses a reasonable request. Explain the purpose and the setting, and ask again in plainer terms.

Turning technique into a habit

Prompting improves through deliberate practice, not through collecting tricks. Keep a small library of prompts that have worked for you. The briefing paragraph for your weekly report, the rewrite instruction that matches your voice, the format you always want for research notes. Reuse beats reinvention.

A simple regime makes the habit stick. For two weeks, take one real task a day and write the prompt twice. Write it as you naturally would, then write it again with the four-part anatomy. Compare the two answers side by side. Most people need only a few days before the fuller style becomes automatic.

Judge the habit by edits, not by feel. Count how much of a reply you keep before rewriting it. A prompt that lifts the keep rate from a quarter to three quarters has paid for the extra twenty seconds many times over. That is the only benchmark that matters at your desk.

Vendor documentation rewards a skim as well. Guides such as OpenAI's prompt engineering guide and Anthropic's prompting overview collect patterns that transfer between assistants. Read one, then test three of its ideas on work you actually have to do.

The deeper reward is not better AI answers, though those arrive within minutes. Writing a good prompt forces a moment of clarity about what you want: audience, purpose, form. That is the same discipline good writing has always demanded. A no-cost assistant such as ASKAI.FREE is a fine place to rehearse before it matters.

From here the habits branch in two directions. Applied to reading and note-taking they become a method, which we set out in using AI assistants for research and study. Applied to recurring chores they become a small personal system, covered in our guide to practical tasks an assistant can do today. Both start in the same place. Everything else we have written on the craft sits under using AI well and inside the AI Assistants hub.