Stop Saying "Write Me a...": The Four Ingredients of a Prompt That Works

Try a quick experiment. Open any AI chat and type:

Write me a leave request email.

You'll get an email back. Polite tone, medium length, vague reason. You'll read it, sigh, and rewrite most of it yourself.

Now try the same request written differently:

Write me a leave request email.
Recipient: my direct manager. Keep it professional but not stiff.
Reason: a family member is hospitalized and I need to be there for two days.
Dates: out August 4-5, back in the office August 6.
Handover: two ongoing projects are covered by a colleague — I'll send handover notes the evening before I leave.
Requirements: under 100 words, one paragraph, no filler opening, and close by saying I'll stay reachable by phone.

Most people can copy the second output, change a word or two, and send it as is.

Same model, same price. The difference: the first prompt is a wish. The second is a work order.

Why AI needs you to specify the format

An AI model works by predicting, from the text you give it, what should come next.

Type "write me a leave request email" and there are tens of thousands of reasonable versions in front of it — a student's, a manager's, formal, casual, three lines, three hundred words. It has to guess which one you meant. The odds of it guessing the one in your head are not great.

Here's a useful comparison. Walk into a cha chaan teng — Hong Kong's classic local diner — and just say "give me a rice dish." The server won't scold you for it, but they will ask: which rice dish? Coriander or no coriander? Spicy? Small or large? If they don't ask, you get whatever's most common that day, and it might not be what you wanted.

A prompt is just you answering, in advance, the questions the server would have asked.

Once you've answered them, something changes: the AI stops guessing and starts executing. Most of the improvement in output quality comes from this step — not from switching to a more expensive model.

Four ingredients: Role, Context, Format, Constraint

You don't need a formula. A good prompt is usually just four things spelled out.

Ingredient one: Role — who is speaking

Telling the AI its role sets the vocabulary, depth, and stance in one line.

• Weak: Explain what an insurance excess is.

• Strong: You're an insurance agent explaining "excess" (the amount the policyholder pays before the insurer covers the rest) to a client with no insurance experience and only basic financial literacy.

Same question, but the second version won't bury the answer in jargon — because you've already told it who's listening.

Ingredient two: Context — what it needs to know

The AI doesn't know your situation unless you write it down. Background is usually what decides whether the output is usable as-is or needs a rewrite.

Four things are worth covering: who will read this, what setting, what facts are already known, what outcome you're after.

• Weak: Write a product description.

• Strong: Write a product description for the first paragraph of an online store listing. The product is a ceramic pour-over coffee dripper. The selling points are better heat retention than plastic and easier cleaning. Buyers are mostly office workers just getting into pour-over coffee — budget-conscious, and worried they won't know how to use it.

Ingredient three: Format — the most overlooked, highest-payoff one

This is the single most valuable ingredient of the four. If you don't specify a format, the AI defaults to what it's used to: paragraph after paragraph of prose. What you actually want is often a table, a list, or something you can paste straight into Excel.

Specifying format is straightforward:

Output as a table with three columns: Task, Owner, Deadline.

Give me five bullet points, each under 15 words. No explanation — conclusions only.

Give me two versions: Version A formal, Version B casual, separated by a divider.

The most advanced — and most practical — move is to draw the template yourself and hand it over:

Using this exact format, turn the meeting notes below into a to-do list:

[TODO] <one-line description>
Owner: <name>
Deadline: <date, or "TBD" if not mentioned>
Source: <the exact sentence from the notes>

--- (separate each item with this line)

Once you've handed over a template, consistency jumps immediately — item ten looks like item one, because the model has something concrete to copy. This matters most if your work is repetitive: the same template can serve you for a year.

Ingredient four: Constraint — limits and red lines

Constraints aren't there to restrict the AI. They're there to save you editing time later.

The common categories:

• Length: Under 100 words, one line per point, five points max

• Language and register: formal written English, no casual phrasing, keep technical terms in English

• Prohibitions: no filler opening, no buzzwords like "empower" or "disrupt", don't invent numbers that aren't given

• What to do when unsure: if a field is missing information, write "TBD" — don't guess

That last one is worth writing every time. AI models tend to fill in blanks even when they don't actually know the answer. Telling it explicitly to leave a gap rather than guess blocks a large share of the errors you'd otherwise have to catch by hand.

Three before/after rewrites

Enough theory — three everyday examples. Left is how most people phrase it. Right is the same request with all four ingredients added.

1. Summarizing meeting notes

• Before: Summarize these meeting notes for me.

• After: You're a project coordinator. Turn the meeting notes below into two sections: (1) Decisions — as a bullet list; (2) Action items — as a table with columns "Item, Owner, Deadline." Only include what's explicitly stated in the notes. If no owner is mentioned, write "TBD" — don't guess.

2. Reviewing a CV

• After: You're a hiring manager screening applications for a "Customer Service Team Lead" role. Below is my CV and the job description. Do three things: (1) identify the three pieces of experience I should move to the top, and explain why; (2) flag two things that are irrelevant to this role and should be cut; (3) rewrite the first bullet of my work experience in two versions — one emphasizing team size managed, one emphasizing customer satisfaction results. Use formal English. Don't flatter me.

3. A document you don't understand

• After: Below are the terms of my phone contract. List them in a table: key clause, what it actually means for me, anything I should watch out for. Explain it in plain English a non-lawyer can follow, and put a plain-language gloss in brackets the first time any technical term appears. If something isn't stated in the contract, don't guess at it.

What all three have in common: none of it is politeness. Every line closes off a guess the AI would otherwise have to make.

Three common mistakes

Mistake one: thinking longer is better. Length isn't the point — information density is. The leave request example above was six lines, and that was enough. To check whether your prompt is complete, ask yourself: if I handed this exact text to a new hire, could they produce the result without asking a follow-up question? If not, that gap is what's missing.

Mistake two: asking for too much at once. "Analyze this, summarize it, translate it, and turn it into a social post" — cram four jobs into one prompt and the AI will usually do two of them sloppily. Split the requests, or number them explicitly.

Mistake three: retyping the whole thing when the output is wrong. This is the most time-wasting habit of all. When the output misses, don't start over — give a follow-up correction instead:

Direction is right. Three fixes: paragraph two is too long, cut it to two lines; make the tone more formal; replace "as soon as possible" with an actual date.

I made this mistake constantly when I started — three full rewrites' worth of time, solved by one follow-up correction. The biggest advantage of a conversational AI is that you can close in on the right answer turn by turn. Typing everything from scratch each time throws that advantage away.

A template you can copy right now

Lay the four ingredients out as a fill-in-the-blank, and it covers about nine out of ten everyday cases:

You're a <role>.
Context: <who will read this, what setting, known facts>
Task: <the one thing to do>
Format: <table / list / template — spell out columns or line count>
Constraints: <word count, language, prohibitions, what to do if information is missing>

The first few times, it'll feel like extra work. After three or four uses it becomes automatic — and a prompt that works is worth saving and reusing. The same "summarize meeting notes" template can keep working for you a year from now.

What comes next

Once your prompts are good, the next problem shows up on its own: re-pasting the same role and rules every single time gets old fast. There's a fix for that — system prompts and long-term memory files let the AI hold onto your preferences so you stop repeating yourself. That's the next article.

As for actually running Claude, GPT, or similar models from Hong Kong or China: official sign-up is a dead end before you even get to prompting. Anthropic officially supports 175 countries and regions, and neither Hong Kong nor China is on the list. Essevin's job is to clear that one step — no VPN, no need for a foreign credit card. One account gives you access to multiple leading models, billed by usage, top-up through FPS or PayMe, with Alipay and WeChat Pay also supported for users in China.

If you don't want to install anything, sign up and paste the template from this article straight into the site's "AI Chat" — running the same request before and after makes the point better than any explanation.

Information in this article is current as of 25 July 2026 and is provided for general reference only; it does not constitute advice of any kind. Third-party product features, pricing and policies are subject to their official announcements. Essevin service details are as shown on essevin.com and in the console.