Show it, do not describe it

TL;DR
- A paragraph about a layout leaves a hundred choices open. The layout itself closes them.
- Point at the real thing: a page you like, a file that works, a past piece of your own.
- Make "done" something you can check, and show it when you can.
I used to describe what I wanted. "A clean layout. Plenty of space. The photo on one side and the words on the other." Then I would get back something clean and spacious that was not what I meant.
The words were fine. They just left too much open.
Now I point at the thing itself. It works better almost every time.
Why a description falls short
A description is a summary. You hold a full picture in your head, and you write down the three parts you happen to notice. The AI gets the three parts and has to invent the rest.
Take "the photo on one side and the words on the other". Which side? How wide? What happens on a phone? Does the photo run to the edge of the screen, or sit in a box? Each gap is a guess. Each guess is a chance to miss.
A finished example answers all of those at once, including the ones you never thought to ask.
What I point at
When I brief an AI on a website now, I do not describe the layout. My standing rules say which page is the model, and the AI opens it and reads it. One line in my site rules reads, in full: "The financial template is the model." Behind that one line is a working page with every width, gap and break point already decided.
The same goes for code. When a new piece needs a setup file, my rules name an existing file that does it right: "The brochure app's config is the model." The AI copies a thing that works. It does not rebuild one from my memory of it.
When I build a draft website for a business, the AI starts from that business's own site. Every fact on the draft comes from there and nowhere else. The real thing is the brief.
You can do this without any code. Paste the email you liked last month and say "this tone". Share the report your boss praised and say "this shape". Attach the photo and say "this light".
Point at it, do not paste it all
There is a trap on the other side. If you paste every example you own into every request, the AI drowns.
Anthropic wrote about this when it rebuilt the instructions inside its own coding tool: "A common myth is that you want to make these a central repository for every known practice that you might run into... Instead, consider having a tree of files that can be loaded at the right time" (The new rules of context engineering, Thariq Shihipar, Anthropic, July 2026).
So name the example, and say where it lives. Load it when the job needs it. One good reference beats five average ones.
Show what done looks like
The most useful thing to show is the finish line.
Matt Shumer makes this point in a video on the newer models (November 2026): make "done" a test. If the AI can check its own work against something real, it stops guessing whether it is finished.
A test can be simple:
A finished example the result should match in shape.
A list of the questions the page must answer.
A file that must open without an error.
A number that must add up to the total on the invoice.
If you cannot show the finish line, at least say what you will look for. "I will check that every price on the page matches the price list." The AI can run that check before it hands the work back.
The better the reference, the better the work
One more thing follows from this. The AI matches the quality of what you point at. A sloppy example gets you a sloppy copy, made carefully. A fine example gets you something close to fine.
So it is worth a few minutes to find the best example you have. Your best page. Your clearest email. The report that did its job. Then point at that one.
The short version
Before you write a paragraph that describes something, ask yourself one question. Could I just show it?
If you can, show it.
If you can show the finish line, show that too.
