Taste Is an Input. You Cannot Ask a Model For It.

You can spot it now. A landing page loads and something in the first half-second says a model made this — before you've read a word, before you could name what gave it away.

That reaction is not connoisseurship. It's pattern recognition, and it took about eighteen months to develop.

Which means your audience has it too.

What changed

Leonxlnx/taste-skill (MIT, created 19 February 2026) describes itself as "The Anti-Slop Frontend Framework for AI Agents." It reports roughly 73,000 stars.

Its entire stated purpose is preventing models from generating boring, generic output. Not making them more capable — making them less average.

And it doesn't do that with one house style. It ships sub-skills for named directions: brutalist, minimalist, editorial. Pick a direction, get that direction.

Two things are worth sitting with. First, that a tool with this job description needs to exist at all. Second, that 73,000 people agreed strongly enough to star it.

Why it matters

Here is the correction that changes how you work.

Generic is not a limitation. It's a default.

A model does not produce bland output because it can't do better. It produces bland output because bland is what you get when nothing constrains the generation. Ask for something underspecified and the model returns something like the average of everything it has seen that matches the description. The average of all clean modern landing pages is a clean modern landing page — and it is the one everybody else got, because they asked the same underspecified question.

That has a direct consequence, and it is the thesis:

The differentiator is no longer the model. It's whether you supplied a direction before generation, or let it fall back to the mean.

Everyone has access to roughly the same frontier models. Nobody's advantage lives there anymore. The advantage lives in the specificity of the constraint you provide — which is the part that is still entirely yours.

The part that stings

You cannot ask a model for taste.

"Make it beautiful." "Make it interesting." "Make it not look AI-generated." These read like direction and function as noise, because none of them narrow the choice set. The model has no way to act on them except by producing what it already considers good, which is the average, which is the thing you were trying to escape.

Taste is an input you supply, not an output you request. That is uncomfortable, because supplying it requires having a view — and having a view is slower than asking for one.

The alternative is genuinely worse. You review the output, don't like it, can't articulate why, and iterate toward the average from a different starting point.

Named directions beat adjectives

Look at what the skill actually offers: brutalist, minimalist, editorial.

Those are not moods. They are traditions — each with conventions, constraints, and a visual grammar the model has seen thousands of real examples of. Say "brutalist" and you have imported raw structure, exposed grid, heavy type, deliberate harshness. Say "bold" and you have imported nothing, because "bold" describes a feeling and not a set of decisions.

The practical version, which costs one line:

• "Clean and modern" → the page everyone else got.
• "Swiss International grid, single anchor colour, extreme type contrast" → a page with an argument.

Proof and limits

Almost everything above is self-reported. The framing, the anti-slop positioning, the sub-skill structure — all from the repo's own description.

The star count is real and public. It is also the weakest kind of evidence about quality. Stars measure that a lot of people agreed with the premise; they do not measure that the tool delivers on it. A repo can be enormously popular because it names a frustration precisely and still be mediocre in use.

So take the star count as evidence about the problem, not the solution. On the problem, 73,000 stars is a strong signal that this is broadly felt.

What to actually do

None of this requires installing anything.

1. Name the direction before you generate. Not after you review. A direction supplied up front costs one line; a redesign costs a day.
2. Use a tradition with a name. Brutalist, Swiss, editorial, maximalist, terminal, whatever fits — something with real precedent behind it.
3. Write down what you're rejecting. "Not a SaaS landing page" narrows the space nearly as well as a positive direction does.
4. Be specific in one place, not everywhere. One committed decision — a type treatment, a colour, a layout that breaks — carries more than ten hedged ones.
5. Then use the tell as a test. Generate it, look at it cold, and ask whether the first half-second says a model made this. If it does, nothing constrained it.

Four questions, honestly answered

What is confirmed. The repo exists under MIT, was created 19 February 2026, positions itself as an anti-slop frontend framework for AI agents, and ships sub-skills for named aesthetic directions.

What is still unknown. Output quality in practice, how well the directions hold across component types, and whether it stays maintained now that it has an audience.

Why it is worth watching. It is a load-bearing admission about where AI-assisted creative work actually stands: the models are good enough that the constraint is now the human's point of view, and the tooling is starting to reflect that.

How it fits the larger ecosystem. Third project this week built on the same idea — DESIGN.md, bounded layout vocabularies, named aesthetic directions. All three replace open-ended generation with a specified choice set. That is not a coincidence; it is the field figuring out that specification beats persuasion.

Disclosure: this is installed and running in the CKC build stack.

The part that stays true

The model is not the bottleneck. The absence of a point of view is.

What direction did you give it before it generated? If the answer is none, you got the average — on purpose, whether or not you meant to.

---

Sources

• Leonxlnx/taste-skill