How to Write AI Prompts That Actually Work

Search for prompt tips and you will find the same list everywhere. Tell it to act as an expert. Offer it a tip. Threaten it, politely. Say "take a deep breath."

Some of this worked on models from two years ago. Almost none of it is why a prompt succeeds today. The prompts that hold up in real work are not incantations — they are briefs. And the useful skill is not memorising phrases, it is being able to look at a bad output and say precisely what was missing.

This is a long piece because the subject is not actually a list of tricks. It is a way of thinking about instructions.

A prompt is a brief, not a spell

Imagine handing a task to a competent freelancer you have never met. They are capable, fast, and know nothing about your company, your customers, or what you already tried last week. What would you have to tell them?

That list is your prompt. Nothing mystical about it.

The reason "write me a blog post about marketing" produces something bland is not that the model is weak. It is that the request is genuinely under-specified. A human given the same brief would also return something generic, then ask three follow-up questions. The model does not ask. It guesses, confidently, and hands you the average of everything it has read on the subject.

That last phrase is worth sitting with. A vague prompt returns the average. If you want something that is not average, the prompt has to contain the information that makes your situation specific.

The four parts

Almost every prompt that works has these, in roughly this order.

Context

Who you are, who the output is for, and what the situation is.

We run a two-person accounting practice in Pune. Our clients are small business owners who find tax filing stressful and are not financially literate. They are not impressed by jargon and they worry about being judged for disorganised records.

Notice how much of that is about the reader, not about you. Context that describes your audience does more work than context that describes your company.

Instruction

The actual task, stated as one clear ask rather than five bundled together.

If you catch yourself writing "and also," split it into a second prompt. Models handle one task well and three tasks mediocrely. Two prompts run in sequence will beat one prompt doing both, almost every time.

Examples

The single highest-leverage thing most people skip.

One example of the output you want is worth three paragraphs describing it. Paste in a previous email you were happy with and say "match this tone." Paste in two, and the model starts to infer the pattern between them rather than copying one.

This works because tone is genuinely hard to describe. "Professional but warm, not stiff, not overfamiliar" means something different to everyone. A sample removes the ambiguity entirely.

Format

What you want back, structurally. Five bullet points. A table with these columns. A JSON object with these keys. Under 200 words.

Models are good at following format instructions and bad at guessing them. If you do not specify, you get prose, because prose is the average.

The same request, twice

Here is a real difference in quality that has nothing to do with clever wording.

Write a follow-up email to a client who hasn't paid.

You will get something stiff and faintly threatening, addressed to no one, that you would not send.

We're a two-person accounting practice. A client we've worked with for three years is 20 days late on a ₹40,000 invoice. They are usually reliable and I don't want to damage the relationship — I suspect they simply forgot. Write a short follow-up email, warm but clear, that makes it easy for them to pay without embarrassment. Under 120 words. Here's how I usually write to them: [paste a previous email].

The second is not a better-worded version of the first. It contains information the first one did not have. That is the whole difference.

Count what was added: the relationship history, the amount, the emotional read of the situation, the goal, the length, and a sample of voice. Six pieces of context, none of them clever, all of them things you already knew and did not think to say.

Why your prompt failed

When the output is wrong, resist the urge to rewrite from scratch. Diagnose first — there are really only three causes.

It didn't know something. The output is confidently wrong about your product, your market, your constraints. It invented a feature you do not have, or assumed a context that does not apply.

Fix: add context, not adjectives. People instinctively respond to a bad output by adding intensifiers — "make it really compelling" — when the actual problem is that the model does not know what your product does.

It didn't understand the task. The output is well-written but answers a different question. You asked for a summary and got an analysis. You asked for options and got a recommendation.

Fix: simplify the instruction. Usually you asked for two things at once and it picked one.

It didn't know what "good" looks like. The output is correct but wrong in tone, length, or shape. Factually fine, unusable as written.

Fix: add an example. Describing the tone you want almost never works as well as showing one.

Getting good at this diagnosis is most of the skill. People who seem to have a magic touch with AI are not writing more elaborate prompts — they are failing faster and correcting more precisely.

Change one thing at a time

The most common way people waste an hour: the output is wrong, so they rewrite the entire prompt, and the new one is wrong differently. Now they have learned nothing about which change mattered.

Treat it like debugging. Change one variable, run it again, note what moved. It feels slower and is dramatically faster.

A rough order of what to try first:

1. Add a sample output. Fixes tone, length and shape problems in one move.
2. Add the missing context. Ask yourself what a new employee would need to know.
3. Split the task in two. If the instruction has an "and" in it.
4. Specify the format explicitly. Especially if you are going to reuse this.
5. Only then, reword. Rewording is what people try first and it helps least.

Getting output you can actually use

Everything above is about quality. This section is about reliability, which is a different problem and matters more the moment you use a prompt repeatedly rather than once.

Ask for structure explicitly. If the output feeds into a spreadsheet, a document template, or another tool, say exactly what shape you want:

Return only a JSON array. Each object must have exactly these keys: "name" (string), "category" (one of: enquiry, complaint, quote_request), "urgency" (integer 1–5). No commentary before or after the JSON.

That last sentence matters more than it looks. Models like to introduce their output — "Sure! Here's the JSON you asked for:" — which breaks anything trying to parse it.

Constrain the choices. "Categorise this email" invites invention. "Categorise this email as exactly one of: enquiry, complaint, quote_request, other" does not. Closed sets are far more reliable than open ones.

Give it an escape hatch. If the model has no acceptable answer, and you have not told it what to do, it will make one up. Add: "If the document does not contain this information, return 'not found' rather than guessing." This single line eliminates a surprising share of hallucinations in extraction tasks.

Test on the awkward cases. A prompt that works on your three cleanest examples will fail on the messy ones, which is most of real life. Deliberately test the short input, the input in the wrong language, the input with a typo in the name, the one with two orders in a single email.

When to let it think

For anything involving reasoning — a calculation, a judgement between options, a multi-step deduction — asking the model to work through it before answering measurably improves accuracy.

Work through this step by step, then give your final answer on a new line beginning with "Answer:".

Newer reasoning models do much of this internally, so the gain is smaller than it was. But it costs you one line, and it has a second benefit: when the answer is wrong, you can see where the reasoning went off, which tells you what context was missing.

The flip side: do not ask for reasoning on simple tasks. It slows things down and pads the output for no gain.

A template worth stealing

For anything you will run more than twice:

CONTEXT
[Who you are. Who this is for. What situation this sits in.]

TASK
[One instruction. If there's an "and", split it out.]

EXAMPLE OF GOOD OUTPUT
[Paste a real one. Two if you have them.]

FORMAT
[Length, structure, what to omit.]

IF UNSURE
[What to do when the input doesn't fit — usually: say so, don't invent.]

INPUT
[The actual thing to process.]

It is not elegant. It works, and it makes the gaps obvious — you notice you have nothing to put under EXAMPLE, which is exactly why the output has been disappointing.

Treat prompts as assets

The prompt you spent twenty minutes refining is worth keeping. Most people lose it in a chat window and rebuild it from memory next week, slightly worse.

Keep a plain document of the prompts you reuse. For each one, record:

• What it's for, in one line
• The prompt itself
• A known-good input and output, so you can tell when it stops working
• The date you last checked it

That third item is the one everyone skips and the one that saves you. Models update. A prompt that worked in March may drift in September, and without a reference output you will not notice — you will just start being quietly dissatisfied.

Teams get more from this than individuals. A shared prompt library means the person who worked out how to summarise client calls properly does it once, for everyone. It also means the quality of AI output across your team stops depending on who happens to be good at this.

What does not matter as much as you think

Politeness. Say please if you like. It does not measurably help.

Long system prompts. Past a point, more instruction dilutes rather than sharpens. If your prompt is 600 words and the output is still wrong, the problem is usually that you have buried the actual task somewhere in the middle. Models attend to the beginning and end of an input more reliably than the middle — put the instruction at one end, not buried.

Assigning it a persona. "You are a world-class copywriter" was genuinely useful on older models. On current ones it does very little that a clear description of the audience does not do better.

Chasing the newest model. A clear prompt on last year's model beats a vague one on this year's, most days. Upgrade when you hit a real ceiling, not on release day.

Where to go from here

Start with one task you do every week. Write a proper brief for it — context, instruction, example, format. Iterate until it works on five different inputs, not one. Then save it with a reference output.

That single exercise will teach you more than any list of tips, because it forces you through the whole loop: specifying, failing, diagnosing, and keeping.

If you want the structured version — techniques that hold up under real work, testing prompts against actual inputs, and fixing drift when it appears — that is what Prompt Engineering Mastery covers over three weeks. If you are earlier than that and still working out what these models are actually doing, AI Fundamentals for Everyone is two weeks, self-paced, and assumes nothing.