2. The Same Request, Written Two Ways

This is the lesson that changes your results more than any other. We sent the same request twice, once the way most people write it and once with four extra lines, and kept both real outputs.

The task

Email your landlord about a broken heater. Real situation: it died 3 days ago, it is winter, you have called twice with no reply, and you want a repair date in writing.

Attempt 1: how most people write it

Write an email to my landlord about the broken heater.

What came back, trimmed:

Dear [Landlord's Name],
I'm writing to inform you that the heater in my unit [apartment number/address] has stopped working. I discovered this on [date]... I would appreciate it if you could arrange for a qualified technician to inspect and repair the heater at your earliest convenience... I'm available [mention your availability]...

Usable, and you now have homework. It is 135 words with 8 blanks to fill in. It never mentions the two ignored phone calls, because you never told it. And "at your earliest convenience" asks for nothing you could hold anyone to.

Attempt 2: the same request with context

Write an email to my landlord about a broken heater.

Context: it stopped working 3 days ago, it is winter,
I have reported it twice by phone with no response,
and I want a repair date in writing.

Tone: firm but not hostile. I want to stay on good terms.

Length: under 120 words. No greeting fluff.

What came back:

I'm writing to formally document that my heating system has been non-functional since September 1st. I've reported this issue twice by phone without receiving a response or repair timeline.

As the temperature continues to drop, this is a habitability concern that requires immediate attention...

I need you to confirm a specific repair date in writing within 48 hours.

The difference, measured

	Vague	With context
Words	135	97
Blanks you must fill	8	1
Mentions the 2 ignored calls	No	Yes
Asks for a specific deadline	No	Yes, 48 hours
Ready to send	After 10 minutes of editing	After adding your name

Four extra lines turned a template into a letter.

The four lines, and why each one works

Every one of these maps to something Anthropic tells developers to do, and none of it is technical.

1. Context. The facts only you know. The two ignored calls are the entire reason this email has any force, and it could not invent them.

2. Tone. "Firm but not hostile. I want to stay on good terms." Without this it picks a register at random.

3. Length. "Under 120 words." It came back at 97. Length instructions work and almost nobody uses them.

4. What you actually want to happen. "A repair date in writing" is the outcome. Give it the outcome and it writes toward it, which is how "48 hours" appeared without being asked for.

The template

[What you want]

Context: [the facts only you know]

Tone: [how it should sound]

Length: [a number]

What I want to happen: [the actual outcome]

Keep it somewhere you can paste from. You will use it for the rest of your life with this tool.

One thing to check in that output

Look again at the good version: "non-functional since September 1st." We said "3 days ago". It did the arithmetic and produced a date.

That is helpful and it is exactly the kind of thing to verify before sending, because a specific date in a complaint is a fact you are asserting. Useful, not automatically correct.

The one instruction that upgrades almost any prompt

Separate from the template, this clause is worth memorising:

...and tell me what information you would need
to answer this properly.

It converts a confident guess into a checklist. We used it on a phone contract question in lesson 1 and it came back with 6 specific things to find out, then a rule for deciding once you had them, rather than an opinion it had no basis for.

Use it whenever the honest answer is "it depends", which is most decisions worth asking about.

Two more instructions that reliably change the output

Explain why a constraint exists. Anthropic's own guide gives the example of "never use ellipses", which works less well than "this will be read aloud by a text to speech engine, so never use ellipses because it will not know how to pronounce them". Given the reason, it generalises to problems you did not think to mention.

Ask for the format you want. "As a table." "As 5 bullet points." "As a numbered checklist I can print." Most people accept whatever shape they are handed and then reformat it themselves.

A fully loaded prompt, and what it gave back

Here is the template used properly, on something with real stakes. Note how much of it is context:

Help me prepare for a job interview.

The role is operations manager at a 40 person logistics
company. I have 6 years of experience but all of it at
one company, and all of it in customer support, not
operations. The interview is with the founder. My worry
is that they will see the support background as a
weakness.

Give me the 5 questions they are most likely to ask me,
and for the career-change worry, give me a specific
answer I can adapt. Do not give me generic interview
advice.

Six sentences, and 5 of them are facts about the person. The result was 463 words and none of it was the usual list about firm handshakes. The 5 questions it predicted:

1. Why the shift from support to operations. It called this "the elephant" and said a founder will ask it directly.
2. A time you coordinated across teams to solve a customer problem, because logistics operations is orchestration.
3. What you know about our operation now and what you would change first.
4. Six years at one company, so how do you learn new things.
5. A process you built or improved from scratch.

Question 4 is the one worth noticing. Nobody said "I have only worked at one place" as a concern. The prompt said 6 years at one company as a neutral fact, and it worked out on its own that a founder would read that as a risk and probe it. That inference is only possible because the fact was there.

The last clause matters too. "Do not give me generic interview advice" is a negative constraint, and negative constraints are underrated. They cost 6 words and they close off the failure mode you can already predict.

Five mistakes that keep prompts vague

1. Describing the task instead of giving the facts. "I need a difficult email" tells it nothing. The difficulty lives in the facts.
2. Leaving out the awkward detail. The delay was our fault. I have never done this job. The number is bad. That is the detail the whole output should be built around, and it is the one people omit.
3. Asking for "a summary" with no audience. A summary for your boss and a summary for a client are different documents. Say which.
4. No length. With no limit it writes to a comfortable middle length that is almost always too long.
5. Stacking 4 unrelated requests into one message. It will do all 4 shallowly. Ask for one, refine it, then move on.

The 20 second check before you hit send

Anthropic's own guidance offers a test that is better than any checklist: show your prompt to a colleague who knows nothing about the task and ask them to do it. If they would be confused, so is the model.

You will not actually go and find a colleague. But you can run it in your head in about 20 seconds, and it catches the same 3 things every time: a missing fact, an unstated audience, and an assumption you did not know you were making.

Try it now

1. Think of a message you have been putting off.
2. Write the lazy version of the prompt. Send it. Notice the blanks.
3. Now use the template. Send that.
4. Compare which one you could actually send.

Next

Lesson 3 puts this to work on the tasks people use AI for most: rewriting, shortening, and changing tone, with the exact prompts for each.