Type “Write a thank-you email to a client” into ChatGPT, Claude or Gemini and a polished email comes back in seconds. It thanks the client for their continued partnership and looks forward to working together in the future. Nothing in it is wrong. Nothing in it is useful either, because it could go to any client of any company.
That email is not the model failing. It is the model doing exactly what it does with a short prompt, and once you see why, the rest of this course follows from it.
What the model has to work with
A language model has read an enormous amount of text, so it knows how thank-you emails, status updates and job ads usually go. When you send a prompt, it writes the response that most plausibly follows your words, a little at a time. It cannot see your inbox, your client or last week’s meeting. Apart from what your app adds, such as custom instructions you saved or a file you attached, your prompt is all it knows about the job.
So when the prompt leaves something out, the model does not leave a blank. It fills the gap with the most likely answer, which is usually the most common one. The most likely thank-you email is the average of every thank-you email it has seen. That is why first answers so often sound generic: the request was generic, and the answer matched it.
It cannot ask, so it guesses
Hand the same one-line request to a new colleague and they would come back with questions. Which client? Thanks for what? Do you want something from them? A model in a chat window usually answers first. Every gap becomes a decision it makes on your behalf, without telling you. Here are the guesses hiding in that one line:
- Who the client is, and how well you know them.
- What they did that deserves thanks.
- How formal to be, and how long to write.
- Whether you want something next, such as a meeting or a renewal.
- What to stay away from, such as a price or a promise you cannot make.
Each guess is a small coin toss. Stack enough of them and the result drifts far from what you meant, even though every sentence reads well. That is the trap: a guess sounds exactly as confident as a fact.
What the numbers show
When we scored 319 everyday work requests written as one line, 261 of them scored 1 out of 10 on the ten checks this course teaches. Almost none gave the model the facts it needed, and not one said what to do when a fact was missing. Two short additions, one saying who the answer is for, how long and what shape, and one saying what to do when something is missing, lifted the average from 1.2 to 6.8.
Write a thank-you email to a client.
Write a thank-you email to a client who runs three dental clinics. Last month they referred two other practices to us, and both signed up. Thank them for the referrals specifically, and offer them one free month of our scheduling software as a thank-you. Under 120 words, from me, first name only.
The after answers the guesses that matter most: which client, what they did, what to offer and how long to write.
Information beats magic words
Plenty of prompt advice is about wording: special phrases, grand job titles, capital letters for emphasis. Most of it matters far less than plain information. Anthropic’s own guidance compares Claude to a brilliant but new employee who has no context on your norms or ways of working. The question to ask of any prompt is the one you would ask of a brief for that new hire: what would they need to know?
Each lesson in this course takes one of the ten checks, shows what a model does without it, and ends with an exercise. The box scores your prompt as you type, with the same free estimate the Prompt Grader uses. The estimate looks for telltale words, so it can be fooled in both directions. Write what you would really send, and treat the score as a hint.
Read your prompt as if you were the model: you know a lot about the world and nothing about your job. Every question you would have to ask is a guess you are leaving it to make.
Your turn
You will be on leave from June 3 to June 14 and back on June 17. You need an out-of-office reply for clients, and anything urgent about an invoice should go to your office manager on 555-0142.
Done when your prompt covers what to make, who it's for and background facts, and scores 3 or more out of 10. Edit the prompt below; the score changes as you type.
A hint
Say who will get the reply, then add the facts the model would otherwise invent: the dates, when you are back and who to contact.
Show one good answer
Write an out-of-office reply for clients who email me while I'm on leave from June 3 to June 14. Say I'll reply after I'm back on June 17. For anything urgent about an invoice, they should call our office manager on 555-0142. Keep it under 60 words. Most of these clients have worked with me for years, so it can be warm and informal.
It scores 5 out of 10. Yours does not need to look like it: any prompt that passes is right.
More practice
The same idea in 2 other kinds of work. Each one is scored the same way, and each is saved in this browser once it passes.
Practice 1
You run payroll for a 30-person company. Because of the July 4 holiday, timesheets are due at noon on July 2, two days earlier than usual, and anyone who misses the deadline is paid a week late. You post reminders in the team chat.
Done when your prompt covers what to make, who it's for and background facts, and scores 3 or more out of 10. Edit the prompt below; the score changes as you type.
A hint
Say who will read the reminder, then add the facts it would otherwise invent: the new deadline, why it moved and what happens if someone misses it.
Show one good answer
Write a short reminder to post in our team chat for the 30 staff who submit timesheets. Because of the July 4 holiday, timesheets are due at noon on July 2, two days earlier than usual. Anyone who misses the new deadline will be paid a week late. Under 50 words, and friendly rather than stern.
It scores 4 out of 10. Yours does not need to look like it: any prompt that passes is right.
Practice 2
You manage a neighborhood cafe. Over the holidays you close at 2pm on December 24, stay closed on December 25 and 26, and reopen at 7am on December 27. Most of your Instagram followers are regulars who stop in before work.
Done when your prompt covers what to make, who it's for and background facts, and scores 3 or more out of 10. Edit the prompt below; the score changes as you type.
A hint
Say who reads your posts and what they need from this one, then give the exact days and times.
Show one good answer
Write an Instagram post about our cafe's holiday hours. Most of our followers are regulars who stop in before work, so put the days we're closed first. We close at 2pm on December 24, stay closed on December 25 and 26, and reopen at 7am on December 27. Under 40 words, with one line thanking them for a good year.
It scores 4 out of 10. Yours does not need to look like it: any prompt that passes is right.
Quick check
3 questions. Pick an answer to see whether it is right, and why.
-
You send “Write a thank-you email to a client.” What will the model most likely do?
A model in a chat window usually answers first. It fills every gap with the most common answer, so you get the average thank-you email: smooth, and not much use.
-
Which change to “Write a thank-you email to a client” would make the biggest difference to the answer?
Information beats magic words. A grand job title or a plea for effort tells the model nothing about your client, so it still has to guess the facts that matter.
-
True to the lesson or not: if the answer reads smoothly, the model probably did not have to guess.
The model fills gaps with plausible text, and plausible text reads well. That is why invented details are easy to miss, and why you look for gaps before you send.
Remember. A model fills every gap with the most common answer, so the gaps you leave decide how generic the result will be.