The prompt tax is why your AI product feels generic

You added twenty words to your prompt and the result got worse. That's not a model bug — it's tax.

I wrote that design taste is the new moat for this reason: when everyone uses the same models, the difference isn't who prompts more. It's who pays the least prompt tax.

The prompt tax is the hidden cost of re-describing taste each time instead of writing it once where agents can read it — and longer prompts raise it.

What is the prompt tax?

The prompt tax is the time, drift, and review cost you pay every time you describe taste in prose instead of storing it in a file.

You pay it three ways. First, tokens — you rewrite "muted palette, 12px radius, 8px grid, no neon" every session. Second, attention — you read a long output to find where it ignored you. Third, consistency — the next screen forgets what the last one followed, so twelve screens feel like twelve products.

Figma removed that re-brief cost. With Figma's canvas is now open to agents through MCP, an agent reads your components and tokens directly on the canvas instead of you pasting them. The file became the brief.

For solo founders that matters. You don't have design ops — you have a markdown file agents check before they write. Every sentence you keep in your head is tax you pay again tomorrow.

Why do longer prompts make AI products worse?

Because language is a poor container for taste, and more language gives the model more to get wrong.

Last month a 300-word prompt for "premium, trustworthy, not corporate" gave me a nice hero and a generic table. A 40-word prompt plus a DESIGN.md with five token rules gave me a table that matched last week's dashboard. The short prompt won because taste lived in the system, not the sentence.

Berkeley researchers tracking prompt work found six breakdowns — vocabulary gaps, execution gaps, a convergence trap toward polished averages, tacit ceilings, authorship disconnect, and freeze when steering. Berkeley study documenting six breakdowns in prompt-mediated design Each one shows up in the last 20% where your product stops looking like a template.

The Designer Fund 2026 report on AI in design showing output quality as the top blocker says the same from the other side: weekly AI use hit 91%, designers use seven tools on average, and the top complaint is still unreliable output quality. More words hide the few constraints that matter. One file with radius, spacing, and two forbidden moves beats "make it feel premium."

How do you replace prompting with a taste file?

You move taste out of prompts into a place agents can read without you.

I copy the pattern Meta proved last week with Astryx. On September 13 Meta open-sourced a 150+ component React system where CLI docs and an MCP server are the real interface. An agent runs astryx component --list and pulls structured props instead of guessing hex values. That's agent-readable taste. I wrote about it as your design system isn't agent-readable — the token file and the check matter more than the prompt.

Do it this week in one repo:

• Write the taste file. One page: palette, type scale, spacing, radius, and two things you never do. Not "be clean" — "12px radius on cards, 8px grid, muted blues, no gradients, no centered hero."
• Generate the agent docs. Add an AGENTS.md with component names, naming rules, and a ban on raw divs. Tell the agent to read it before it writes UI and to name tokens in the output.
• Add a check that fails. A linter that flags a magic color or radius mismatch. Without a check the file is a suggestion. With it, it's a contract.
• Review against the last two screens, not alone. Screenshot the new page next to the last shipped one. Fix one drift per ship and add the rule to the file.

When is a prompt still worth writing?

When you don't yet know what good looks like.

Exploration needs a destination. Brainstorming a direction or wrestling with a first draft benefits from a wide surface where you can be messy. That's when a longer prompt helps — you are learning, not shipping.

Production needs proximity. Filing a doc, updating a component, or generating the tenth variant should not need a new paragraph. The agent already knows the palette and spacing. You point it at the file and it acts.

A simple rule: if you know where the output lives, use the file. If you don't yet know what it should be, use the prompt. The tax hides until week three, when nothing matches.

Cut the tax this week. Write one page of taste and keep it readable to agents. Your prompts get shorter, your product gets more specific.