Prompting Basics

Why vague prompts get generic answers, and the four things every good prompt includes: the task, the context, the shape of the answer, and the input.

12 minLesson 1 of 8
On this page
  1. A prompt is a brief for a capable stranger
  2. See the difference
  3. The four parts of a good prompt
  4. Treat the first answer as a draft
  5. Try it yourself
  6. Common mistakes
  7. Check yourself
  8. What to practise next

You paste a long article into an AI assistant and type "Summarize this." You get back three tidy paragraphs. They are accurate. They are also useless to you, because what you actually needed to know was whether the article changes your team's plans, and the summary spends most of its words on background.

The model did not fail. It did exactly what you asked, and it had to guess everything you did not say. Prompting is the skill of not making it guess.

A prompt is a brief for a capable stranger

Picture handing the same task to a smart contractor who started today. They write well and know a lot about the world. They know nothing about you: who the summary is for, what decision it feeds, how long it should be, or what you already know.

A language model is in the same position. It has read a huge amount of text, so it is good at producing what a typical answer to your request looks like. If your request is vague, you get the typical answer. If your request carries your situation, you get an answer for your situation.

See the difference

Here is the "Summarize this" request, and a typical reply.

Vague promptIllustrative example. Real replies vary.
You
Summarize this article. [2,000-word article about moving a company's web app from one database to another]
Model
The article discusses a company's migration from its original database to a new system. It begins by describing the history of the original setup and the reasons it became difficult to maintain. The authors then outline the planning process, including stakeholder meetings and timelines. Finally, they reflect on lessons learned, noting that communication and testing were key to success.

Now the same article, with a prompt that says who is asking, why, and what shape the answer should take.

Specific promptIllustrative example. Real replies vary.
You
I lead a team of four engineers. We are deciding next month whether to make the same database move described in the article below. Summarize the article for that decision: - 5 bullet points at most - Focus on cost, time taken, and what went wrong - Skip the company history - End with one line: the biggest risk for a small team <article> [2,000-word article] </article>
Model
- Took 7 months against a 3-month plan; the overrun came from rewriting reports that queried the old database directly. - Two engineers worked on it full time; the article does not give a total cost. - Data copying itself was quick. Most of the problems were in code that assumed the old database's date formats. - They ran both databases side by side for six weeks to compare results before switching. - One outage happened during the switch, caused by a connection limit nobody had tested. Biggest risk for a small team: running two databases in parallel for weeks, which is what made their switch safe.

Nothing clever happened in the second prompt. It answers the questions a good colleague would have asked before starting.

The four parts of a good prompt

Most strong prompts contain four things. You do not need headings for them, but you should be able to point at each one.

  1. The task. What to do, as a direct instruction: "Summarize", "Rewrite", "List", "Find the bug in". Put it near the top.
  2. The context. Who it is for, why you need it, and what you already know. "For a decision my team makes next month" changes what counts as important.
  3. The shape of the answer. Length, format, tone, and what to leave out. "5 bullets, no history" is something the model can actually follow.
  4. The input. The material to work on, clearly separated from your instructions.

The fourth part matters more than it looks. If you paste an email into a prompt and the email says "please reply by Friday", the model can confuse that with an instruction from you. Wrap the material in clear markers, like <article> … </article> tags or a line saying "Text to summarize:", so there is no doubt about which words are yours.

Treat the first answer as a draft

Even a good prompt rarely gets a perfect answer on the first try. That is normal. The useful habit is to read the reply, name exactly what is wrong with it, and change the prompt to fix that one thing.

  • Too long? Give a number: "under 120 words".
  • Wrong focus? Say what matters: "I care about the cost, not the timeline".
  • Wrong tone? Say who reads it: "this goes to a customer who is already annoyed".
  • Made something up? Ask it to use only the provided text and say "not stated" when something is missing.

Changing one thing at a time tells you what actually helped. If you rewrite the whole prompt every time, you learn nothing about why the new version worked.

Try it yourself

Try it: Fix the landlord email

Here is a prompt someone sent to an AI assistant. The reply was a stiff, generic letter that missed the important facts. Rewrite the prompt so the assistant has what it needs. Use the four parts: task, context, shape, input.

Show a sample answer

One stronger version:

Rewritten promptIllustrative example. Real replies vary.
You
Write a short email to my landlord asking them to fix the heating. Context: the heating in my flat has not worked since Monday (3 days). It is around 12°C inside at night. I reported it by phone on Monday and was told someone would come "this week", but nobody has contacted me. I have a good relationship with the landlord and want to keep it. Shape: polite but firm, under 150 words. Ask for a specific repair date by Friday. Mention that I can give access any weekday after 4pm. Sign it: Priya, Flat 3B

Why it works:

  • Task is specific: not "about the heating" but "asking them to fix it".
  • Context gives the facts the email must contain: how long, how cold, what was already promised. Without these, the model invents vague filler.
  • Shape sets tone, length and the one concrete request, a repair date, so the email ends with an action.
  • Small details, like the sign-off and access times, save you an edit afterwards.

Common mistakes

Check yourself

Quick check

You ask for "a summary of this report" and get a generic overview. What is the most useful first change?

Which instruction is the model most likely to follow exactly?

Why wrap pasted material in markers like <article> tags?

What to practise next

Take a prompt you used this week that gave a disappointing answer. Check it for the four parts and rewrite whichever is missing. Then continue with Be Clear and Specific, which goes deeper on wording, and Show Examples, the fastest way to get a precise format.