“Make it pretty”
is the wrong prompt

Eight PM on a Sunday. I sent myself the link to the video-management page of our course platform, along with the message I had just received about it: "this page is ugly, and not professional at all."

I opened it and could not argue. An eight-column table showing raw ids no human reads. Dates rendered as 2026-07-06T02:14:33.000Z. Error messages sprawling across full rows. A bare file input for uploads. The page did everything correctly, and looked like an internal tool nobody loves.

My hands started typing the command everyone types: "make this page pretty and professional." This time I stopped, because I already knew where that prompt leads.

Part 1The night the admin page got called out

The page in question manages course videos: upload a file, wait for transcoding, attach the result to a lesson. Simple job. But its face was what we jokingly call "a debug table that accidentally shipped to production."

I used to fix pages like this the way everyone does: tell the AI to make it prettier, more professional. And I got back exactly what you would predict. A soft gradient here, floating cards with shadows, decorative icons on table headers, gentle pastels. It did look "nicer." But the person using it still had to ask the same questions on every visit: which lesson uses this video, which one finished transcoding, which one failed and what do I do about it.

Part 2Pretty, for whom?

Consider what the model actually receives from "make it pretty." No information at all. It does not know who uses the page, when, or to decide what. All it knows is "pretty", which has no definition.

With no information, the model does the only thing it can: it answers with the statistical average of millions of sites it has seen. Cream backgrounds, rows of rounded cards, a giant number in the hero with a purple-blue gradient, tiny uppercase labels above every section. None of that was chosen because it fits your work. It was chosen because it is the most frequent pattern in the training data.

That is why AI-designed sites all look alike, to the point where people can spot "AI made this" instantly. Beauty with no owner: put three such sites side by side and you cannot tell whose is whose.

The problem with "make it pretty" is not that AI cannot do it. It answers a question nobody defined, so you get the average's answer, not your work's answer.

Part 3Refusing to code until three questions are answered

This time I tried the other road: invoking the design skill installed on my machine (the one called impeccable) instead of prompting directly. The first thing it did was not open a CSS file. It checked whether the project had a file named PRODUCT.md. It did not, so the skill stopped and refused to continue until three things were answered.

1. Who uses this, and what are they doing when they open it? Not a glossy persona. Concrete: what time of day, are they in a hurry, what decision are they making.
2. Does design serve the task, or is design the product? This picks the register. An admin page the owner opens many times a day is product: familiarity is a feature. A landing page that has to sell is brand: memorability is a feature. Different rulebooks.
3. What must it never look like? This matters more than people expect, because what you refuse is always sharper than what you want.

Our answers that night became a PRODUCT.md that reads roughly like this:

Users: a solo owner managing courses, students and money in between other work; decisions must close on one screen · Thai learners, new to AI and code, mostly on phones
Register: product; design serves the task
Personality: quiet, real, trustworthy
Never like: loud gradient-heavy SaaS dashboards · marketplaces where discount badges shout over each other · government-style raw-table screens

You answer once and the file stays with the project. The next design job on the same site starts from this context instead of from zero, and every teammate, human or AI, sees the same logic.

Part 4Bans matter more than guidelines

Context alone is not enough. What sets this skill apart from generic advice is a ban list written as hard refusals: when a pattern appears, restructure it, do not soften it. Some entries:

• Gradient text. Pretty with no meaning; use one color and carry emphasis with weight.
• The giant-metric hero: one huge number, small label, supporting stats. The SaaS uniform.
• Identical cards in a grid: icon, heading, blurb, repeated forever.
• Tiny uppercase eyebrow labels above every single section.

Recognize the list? Every entry is one of the signatures of AI-generated design from Part 2. The skill knows its own averages and forbids itself from using them.

Product-register work gets its own extra rules, and these are my favorites: every interactive control ships with all its states (default, hover, focus, disabled, loading, error), not just the happy one. Motion conveys state, never decoration. Empty states teach the interface instead of saying "no data." The "professional" we got called out about lives in these details, not in the color scheme.

Part 5The result was never about color

With the strategy questions answered, the actual redesign took about an hour. And almost everything that changed was not aesthetics. It was the information a person decides with.

• Raw ids replaced by real thumbnails from the videos, which the system had all along but never showed
• A new line, "used in lesson: …", linking straight to the lesson editor. That data sat in the database the whole time; nobody had asked for it
• Status became human-language chips, with accent color only where attention is needed: transcoding and failed get the rust, everything else stays quiet
• Long error messages folded behind "show reason", next to a working retry button
• The upload input became a drag-and-drop zone that states accepted formats, size limit, and progress in plain words

The prettiness that did arrive was a byproduct of answering "what decisions does this screen exist for" and cutting everything that did not serve them. Most professional-looking software is not dramatically more beautiful than average. It answers its user's question faster.

Part 6Use this without the skill

The core of this story is not the tool. It is forcing yourself to answer strategy questions before giving any design order. Even without Claude Code you can start today: write one file with five sections.

1. Who uses it, and what are they doing when each screen is open. Concrete scenes, not slide-deck personas.
2. Register: does design serve the task (product) or is design the product (brand)?
3. Three personality words. Ours: quiet, real, trustworthy.
4. Three anti-references. Point at real sites or real patterns. The more specific, the better.
5. Three to five decision principles that can actually settle arguments. Ours includes "information people decide with beats decoration."

Attach that file to every design prompt, with any AI. The difference shows from the first job, because the model stops answering from the world's average and starts answering from your context.

Re-reading our PRODUCT.md from that night, I noticed something. The word "pretty" never appears in the file. Yet the page it produced is the one nobody calls unprofessional anymore.