Anyone can get a fluent paragraph. That stopped being impressive the week the models learned to sound like a competent junior. Fluency is the default. The scarce move is knowing whether the paragraph is true, useful, and worth sending to a human who pays.
Most of what is sold as prompt engineering is the first half of that sentence: how to ask. Libraries. Better models. A new syntax every quarter. None of that is wrong. It is just not the part that fails in public. What fails in public is the page you almost shipped because it sounded finished.
The draft that almost ships
On the Method page there is a reconstruction of that failure. A designer selling a two-week homepage sprint. The first useful draft from the model is not garbage. That is the trap.
I help SaaS founders turn confusing homepages into clear ones. In two weeks you’ll have a new homepage that converts. Book a call to see if we’re a fit. 50+ founders have used this process.
A tired client will nod. It has a who, a time, a promise. It also cannot survive a stranger. “Converts” to what. “Book a call” is not the $900. “50+ founders” appeared because models like proof and will invent it if you did not forbid it. “I help” is a job description, not a decision. This is the copy that dies in a sales call, not in the chat.
Write the list before you generate
Evaluation is not a vibe you apply at the end. It is a list you write before anyone prompts. For that landing the list was short on purpose:
- A stranger can say who it is for, what they buy, and the price, in five seconds.
- The headline can be argued with. If it cannot be wrong, it is decoration.
- One action. The paid one. Not a menu.
- No number the client did not give you.
- If a competitor can paste their name in and the page still works, it dies.
Run the draft against that list and the work is obvious. You do not need a certificate in prompt engineering. You need the nerve to reject a paragraph that would have gotten a polite email and no money.
“You still need good prompts”
Yes. Prompts are how you reach a draft worth stressing. They are downstream of the list. If you cannot say how you will reject the answer, you were never doing the work. You were collecting text. A better prompt without a kill-list just gets you to a more expensive brochure.
This is why prompt packs expire and named methods do not. The pack is a starter. The list is the skill. The Loop puts the list in Structure and the refusal in Stress. Skip Stress and you are decorating a guess with better adjectives.
Taste is not the whole list
People hear “evaluation” and say it is just whether you like it. Taste is real. It is not the whole list. Claims that must be true are not taste. A price that must appear is not taste. A number you did not earn is not taste. Taste decides between two drafts that already passed. Most people never get that far because they accept the first fluent one.
Make the model hostile when you can. Ask it to attack the page as the customer who will not pay. Then run that output against the same list. The model is a cheap adversary. It is not the judge. You are.
What this has to do with getting paid
Clients do not buy prompts. They buy a page, a brief, a decision. Evaluation is the step that turns chat into one of those. If your AI education never asks you to throw work away, it is a tutorial. Tutorials do not raise rates. That is the other essay.
Starting prompts so you reach something worth stressing: Prompt Secrets, $4. The list, and the Monday, stay free on the Method page.