How do you write an AI prompt that gets a good answer?
The new colleague: a short learning video. No voice, only titles and screens, with music.
What this video shows
Most of us have typed something like this: “You are a world-class expert. This is very important. Write a welcome email.” The answer comes back fine, and a bit flat. It could be any company’s email.
This video is about why that happens and what to do instead. I think of the AI as a brilliant colleague who started this morning. Fast and smart, with no idea who your customers are or what a good result looks like to you. Nobody would hand a new colleague a one-line task and a compliment. We would tell them the goal, who it is for, show them one email we liked, and say what shape the answer should take. The video does exactly that, line by line, on one ordinary task: a welcome email for a small online course.
It also shows a trap I fall into myself. The first answer is close, so I keep correcting it: shorter, warmer, no, not that. Each turn drifts further from what I wanted. The fix is to open a fresh chat and paste the whole brief in one message.
What you will learn
- Why “you are an expert”, capital letters, threats and tips do not change the answer. Researchers at the Wharton School tested them on hard exams across many models.
- The six parts of a brief the AI can act on: the goal, what good looks like, who it is for and why, one example, the shape of the answer, and what to do when unsure.
- Why a long thread of corrections makes the answer drift, and why one complete message in a fresh chat brings it back.
- A test before you press send: would a colleague who knows nothing about this task understand it?
Further reading
- Prompting best practices, Anthropic (living page, read 7 October 2026). The “brilliant but new employee” idea, and the test of showing your prompt to a colleague with no context.
- Prompting Science Report 3: I’ll pay you or I’ll kill you, but will you care?, Meincke, Mollick, Mollick and Shapiro, Wharton School, August 2025. Threats and tips did not change exam scores on average.
- Prompting Science Report 4: Playing Pretend: Expert Personas Don’t Improve Factual Accuracy, Basil and others, Wharton School, December 2025. “You are a world-class expert” gave no reliable gain.
- LLMs Get Lost in Multi-Turn Conversation, Laban, Hayashi, Zhou and Neville, May 2025. Models do much worse when the task arrives in pieces over many turns; put it in one message or start a new chat.
This is how my agent team gets its briefs every day. If you want your own team to work with AI this way, I join teams as a lead engineer and bring agents into how they build. Tell me about your team.
Transcript
Every line of text shown on screen, in order. Nothing is spoken. Lines in the chat windows are what the person types or what the AI answers.
Opening
- Logo: vai SoftLab
- The new colleague
- Learning video · prompt engineering
Scene one: the magic words
Typed into the chat window:
- You are a world-class expert.
- THIS IS VERY IMPORTANT.
- Write a welcome email for our new customers.
The answer card:
- Dear valued customer, (then grey lines: one long block of text)
- Labels under the answer: generic · could be any company
The first line of the prompt changes three times. The answer card comes back the same each time:
- Please, please, you are the best at this.
- Do this right or I will report you!
- I will tip you generously!
Title: Different magic words. Same answer.
Small line: tested on hard exams, many models, by university researchers
Scene two: the new colleague
The robot sits at an empty desk with a badge that says NEW.
Title: The AI is a brilliant new colleague. It started this morning.
Small line: it knows nothing about your company, your customers, or what good looks like
Scene three: the brief
Typed into an empty chat window, one line at a time. The gold label beside each line is in brackets.
- Write a welcome email for customers who just bought our online course. (the goal)
- Good means: short, warm, and one clear next step: log in and watch the first lesson. (what good looks like)
- They are busy parents who learn in the evening, so it must fit on a phone screen. (who it is for, and why)
- Here is an email of ours they liked: (one example) A small quote card with an email slides in beside the window.
- Give me a subject line and the body, nothing else. (the shape of the answer)
- If anything is unclear, ask me before you write. (when unsure, ask)
Send. The AI answers with a question first:
- Should I mention the money-back guarantee?
- Label: it asked first
The person types: yes
The email card:
- Subject: Welcome! Your first lesson is ready
- Hi, and welcome to the course.
- Each lesson is short and works on your phone.
- Next step: log in and watch the first lesson.
- Not happy? You get your money back.
- Labels beside the email: right length · right tone · one next step
Title: Clear beats clever.
Scene four: lost in the thread
The chat thread grows. The person keeps correcting, one short message at a time:
- shorter
- warmer
- no, not that
- add the guarantee
- take it out again
The answer cards get longer and greyer. The robot walks off the edge of the screen.
Title: A long thread of corrections makes it worse.
Small line: once it takes a wrong turn, it does not find its way back
A fresh, empty chat window opens on top. The whole brief from scene three is pasted in as one block, with the quote card, and sent. The same good email comes back:
- Subject: Welcome! Your first lesson is ready
- Hi, and welcome to the course.
- Each lesson is short and works on your phone.
- Next step: log in and watch the first lesson.
- Not happy? You get your money back.
Title: Start fresh. Paste the whole brief in one message.
Scene five: the test
One card in the middle of the screen:
- Before you send it, ask: would a colleague who knows nothing about this task understand it?
- If they would be confused, so will the AI.
Scene six: the line to remember
- Say what you want, why, and what good looks like.
Closing
- Logo: vai SoftLab
- vaisoftlab.com