How to Write AI Prompts That Work: A Formula, Techniques and Templates

The difference between a generic, useless answer and one that saves you an hour almost always comes down to the prompt: the instruction you write. This skill is called "prompt engineering," but there's nothing technical about it. It's learning to ask clearly, the way you would with a very capable coworker who doesn't know your situation.

1. The principle
2. The six-part formula
3. 7 techniques
4. Templates
5. Iterate and save

The fundamental principle

Imagine you hire a brilliant assistant on their first day. They don't know who you are, what the work is for, who will read it or what you consider a good result. If you say "write an email to a customer," they'll produce something correct but generic. If you give them the context, they'll nail it the first time. AI works exactly the same way: anything you don't tell it, it has to guess.

The formula: six parts of a good prompt

You don't always need all six, but the more you include, the better the result. Compare:

❌ Write a post about saving money.

✅ You're the social media manager for a credit union in Ohio.
Write an Instagram post for 20- to 30-year-olds who just landed
their first full-time job. Goal: get them to start saving 10% of
every paycheck. Friendly tone, no lecturing.
Format: a hook in the first line, 3 practical tips as bullets,
a closing question to spark comments and 5 hashtags.
150 words max.

The first only states the task. The second has a role (social media manager), task, context (a credit union in Ohio), audience and goal (first-job 20-somethings saving 10%), format (hook, 3 bullets, question, hashtags, 150 words) and a tone constraint.

Techniques that improve any prompt

1. Give examples (few-shot prompting)

If you want a specific style or format, show it. That works better than describing it:

Classify each review as Positive, Negative or Neutral.
Examples:
"Arrived fast and well packed" → Positive
"The color doesn't match the photo" → Negative
Now classify:
1. "Does what it says"
2. "Took three weeks to arrive"
3. "It's a regular box"

Expected result: 1. Positive · 2. Negative · 3. Neutral, in the same format as the examples. If you also want it to explain each decision, ask; otherwise, short answers are easier to paste into a spreadsheet.

2. Separate instructions from data

When you paste a long text, mark where it starts and ends so it doesn't get confused with your instructions. Triple quotes, dashes or tags all work:

Summarize the following contract in 5 key points for someone
without legal training, and flag any penalty clauses.

<contract>
[paste the text here]
</contract>

3. Ask it to reason step by step

For problems involving logic, numbers or decisions, ask it to analyze before concluding: "Walk through the pros and risks of each option step by step, then give your recommendation." Models with a reasoning mode do this on their own, but asking them to show their assumptions still helps.

4. Ask for several options

"Give me 5 different headlines: two informative, two as questions and one with a statistic." Picking from options is faster than fixing a single answer.

5. Iterate: the first result is a draft

Don't start over if you don't like something. Ask for specific changes: "shorter," "cut the adjectives," "the second paragraph sounds salesy, make it more neutral," "keep everything but change the ending." The conversation is part of the method.

6. Ask it to ask you questions

When the task is complex and you're not sure what context to give: "Before you start, ask me whatever you need to know to do this well." The AI will tell you what's missing.

7. Ask for a self-check

"Review your answer: is there any fact you can't confirm or any calculation error?" It often catches its own mistakes.

Ready-to-use templates

📧 EMAIL
Write an email to [recipient] with the goal of [goal].
Context: [situation]. Tone: [formal/friendly/firm but polite].
Maximum length: [X] words. Include a subject line.

📄 SUMMARY
Summarize this document for [audience] who needs to [decision or use].
Format: 1-sentence takeaway + 5 key points + risks or open items.
Use only the information in the text; if something isn't there,
don't make it up.

💡 IDEAS
Give me [N] ideas for [topic] for [audience/business] with
[constraint: budget, deadline, channel]. For each one: what it is,
why it could work and the first step to test it.

🎓 LEARN
Explain [concept] as if I were [level]. Use an everyday analogy
and a concrete example, and end with 3 questions to check that
I understood.

✅ REVIEW
Review this text as a [role: editor, lawyer, demanding customer].
Point out errors, confusing parts and improvements, then give me
the corrected version keeping my style.

Iterating in practice: an example in three turns

Here's what a real conversation looks like with technique 5. Elena runs a small tutoring center in San Antonio and needs a Facebook ad for her evening English classes.

TURN 1
You're a local marketing specialist. Write a Facebook ad for my
tutoring center in San Antonio: English classes for working
adults, evening schedule (7 to 9 p.m.), groups of 8 people max,
first class free. Goal: get people to text us. 80 words max.

TURN 2
I like it, but the opening is generic. Give me 3 different hooks
for the first line: one with a question, one with a fact about
the class and one about the fear of speaking English at work.

TURN 3
Use the third hook, remove the exclamation points and end with
this exact call to action: "Text us to save your free class."

Expected result: in turn 1, a correct but standard ad; in turn 2, options to choose from; in turn 3, copy in her own voice, ready to publish. Three short messages beat one "perfect" prompt written in a single shot, because each tweak builds on something you can already see.

A prompt that interviews you

I want to prepare [a proposal for a new client / my resume for an
analyst job / my end-of-year speech]. Before writing anything,
ask me the 5 most important questions to do this well, one at a
time. When I've answered them all, write the draft.

Expected result: a first question (not five at once) and, after your answers, a draft that fits far better than if you'd asked for the text directly. It's technique 6 turned into a template.

Common prompting mistakes

Custom instructions and projects

Most assistants let you save permanent instructions (about you, your work and how you want answers) or create "projects" with reference documents. If you always work on the same business or topic, set it up once and you'll stop repeating context in every chat. Here's a custom instruction you can adapt:

I'm an office coordinator at a food distributor in Chicago. I
mostly write emails to suppliers, meeting notes and reports for
management. I prefer direct answers in American English, with
bullet points when there are more than three ideas. If you're
missing information to do the task well, ask me before assuming.

Review these instructions every few months. If your role or tasks change, outdated instructions will quietly push answers in the wrong direction.

You're a Spanish teacher with experience teaching working adults
who speak English. Create a 12-week study plan for me.
Context: I'm a nurse in Houston at a beginner level (A2). I work
12-hour shifts three days a week, so I can study 30 minutes on
workdays and 2 hours on my days off. My goal is to handle basic
conversations with Spanish-speaking patients (around B1).
Format: a week-by-week table with the goal, activities and free
resources, and at the end 5 tips to keep me from quitting.
Before you create it, ask me up to 3 questions if anything is
missing.

Summary

• Anything you don't say, the AI has to guess: give it context.
• The formula: role, task, context, audience and goal, format, examples or constraints.
• Use examples, delimit your data, ask for reasoning, options and a self-check.
• Iterate on the draft instead of starting over.
• Save prompts that work and set up custom instructions for recurring work.

Next lesson: AI for Productivity at Work: Email, Meetings, Reports and Excel. You'll learn which tasks to delegate, see use cases with prompts and measure how much time you actually save.