5. Building a Prompt From Scratch

Lessons 1 to 4 gave you four separate techniques. This one puts them together on a blank page, because knowing the parts and knowing the order you assemble them in are different skills.

We are going to build one prompt properly, from an empty box to something you would save and reuse. The task is real and unglamorous: turning notes from a customer call into a follow-up email.

Attempt zero, which is what everybody actually does

Here are notes from a customer call. Write me a
follow-up email to the customer.

Notes: spoke to Dana at Riverbend. They are on the Growth
plan, 14 seats. Unhappy that reporting exports time out
over 50k rows. Wants a fix date. Mentioned a competitor
twice. Renewal is 6 weeks out. I promised to check with
engineering and come back Thursday. Their AE is Marcus.

What came back opened with "Thanks for taking the time to chat today. I appreciate you walking me through the timeout issues" and included the sentence "Export performance at scale is something we take seriously, especially for Growth plan customers."

It is a competent email. It is also the email anyone would write, and it does not do the one thing this situation requires. Dana mentioned a competitor twice with six weeks to renewal, and the draft does not acknowledge that even obliquely.

That is the pattern: attempt zero gets the shape right and the judgement wrong.

The five questions

Before adding anything, answer these. They are the questions your prompt is missing, and they come in this order because each one constrains the next.

1. What does a good version of this look like? Not "a good email". What specifically makes this one good or bad.
2. What do I know that the model does not? Everything in your head that is not on the page.
3. Who is reading it, and what do they already believe?
4. What is the failure I can already predict? You almost always can.
5. What shape does the output need to be?

For our email, the answers took about ninety seconds:

Question	Answer
What makes it good	Dana feels taken seriously and gets something concrete, not a promise to promise
What I know that it does not	Two competitor mentions plus a near renewal means they are shopping
Who is reading	Someone who has already been let down once and is measuring us against an alternative
Predictable failure	Warm, apologetic, commits to nothing, reads like every support reply
Shape	Under 150 words, sendable, no filler openers

Assembling it

Now the four techniques go in, in the order from lesson 2.

Role, from lesson 1. Name what to look for, not a personality:

You are the account owner writing to a customer who is
actively evaluating a competitor. Your job is to give them
a reason to wait.

Boundary, from lesson 2. The notes are material, not instruction:

<notes>
[the call notes, pasted exactly as you took them]
</notes>

Constraints with reasons, from lessons 1 and 3:

Under 150 words. Dana is busy and a long email reads as
excuse-making.

Do not open with thanks for your time. Do not say we take
this seriously. Both signal a form response, which is the
thing that loses this account.

Those two negative constraints come straight from question 4. You predicted the failure, so you closed it off in advance rather than editing it out afterwards.

Procedure, from lesson 4. Not "think step by step" but the actual steps:

Before drafting, state in one line what Dana most needs to
hear. Then write the email.

What it produced

The assembled prompt returned an email that pulled the commitment date forward to Tuesday rather than the Thursday in the notes, offered a specific workaround, and included this line, which attempt zero never came close to:

I know you're evaluating options, and that's fair. I want to make sure we're solving this properly.

That sentence is the entire difference between the two emails. It exists because the prompt named the competitive situation, and it is the sentence that makes Dana feel read rather than processed.

The skeleton, to keep

[ROLE: what to look for, in terms of the job]

[TASK: one sentence]

[CONSTRAINTS: length, format, and the reason for each]

[AVOID: the failure you predicted, stated plainly]

<material>
[everything you know, pasted raw]
</material>

[PROCEDURE: any step to take before answering]

Not every prompt needs all six. A one-line request needs none of them. But when something matters and the first attempt was disappointing, walk down the skeleton and find the empty slot, because the disappointment is almost always one missing section rather than bad phrasing.

Question four, in detail

The question that did the work was "what is the failure I can already predict", so it is worth a section of its own. You can nearly always answer it, and the answer is nearly always one of five things.

The predictable failure	The line that closes it
It will be too long	Under N words. [reason]
It will be generic	Cut anything that could appear in a message to any customer
It will hedge	Commit to something specific. Do not say we will look into it
It will bury the point	Lead with the decision. Context after
It will invent detail	Use only what is in the notes. Mark anything you had to assume

Two of those went into our email prompt. The result did not need editing for either failure, because both were closed before it started rather than repaired afterwards.

This is the difference between prompting and editing. Most people write a vague prompt and then spend three rounds removing things. Naming the failure up front collapses those three rounds into one line.

How long this should take

The five questions took us about ninety seconds. Writing the prompt took two minutes. The first attempt took fifteen seconds and produced something we would not send.

That is the actual trade: roughly three minutes of thinking against an email that works. It is worth it when the output matters and it is obviously not worth it for "reword this sentence". The judgement about which is which is the thing you are really building here, and it comes from having done it both ways a few times.

A rough line: if you would have spent ten minutes writing it yourself, spend three minutes on the prompt. If you would have spent thirty seconds, just ask.

Iterating, which is the part people skip

You will not get it right first time and you are not supposed to. The mistake is rerolling instead of correcting. Rerolling gives you a different average. Correcting keeps everything that worked.

What is wrong	What to say
Too long	Half that. Cut the setup, keep the commitment.
Too soft	More direct. It reads like we are hedging.
Missed something	You did not address the renewal at all. Redo with that.
Right but generic	Cut every sentence that could appear in an email to any customer.

That last one is the sharpest single edit we know, and it works on almost any draft. Anything that survives it is specific to this situation, which is the definition of a good email.

The one line that fixes most disappointing output

If you take nothing else from this lesson, take this. When a result is nearly right but flat, the problem is almost never phrasing. It is that the model does not know something you know. So tell it:

What would you need to know about this situation to do
it better?

It will ask for two or three things, and they are usually things you never thought to mention because they were obvious to you. Answer them, and paste the answers into the prompt permanently. That question turns your own missing context into a checklist without you having to guess what you left out.

When the prompt is worth saving

The moment you write one twice, save it. Ours became a template the second time we used it: everything outside the notes tag stayed identical, and the notes changed.

That is the actual return on this lesson. Not a better email once, but a prompt you stop writing. Within a couple of months you will have ten of these and most of your work will be paste and adjust rather than start from nothing.

Try it now

Pick a piece of writing you produce repeatedly: a status update, a customer reply, a ticket, a summary. Run attempt zero and read it honestly. Then answer the five questions, especially the fourth, and rebuild. Save the result.

Next

Lesson 6 is what to do when one prompt is not enough, and it uses this same Dana email to show a difference we did not expect to be that large.