amplifier-bundle-taste  ·  v0.3.0

Teaching AI Taste

Anti-slop frontend design, made grep-testable

The Problem

Ask any LLM for a frontend, and you get the same slop

It isn't a bug — it's the model's statistical priors leaking out. The taste bundle (an Amplifier port of Leonxlnx/taste-skill) exists to rewire those defaults.

And that slop is not random — it has a fixed fingerprint.

Proof — the fingerprint

The bundle names eight repeat tells, verbatim

The README lists the recurring slop patterns explicitly — the same eight fingerprints show up again and again.

If it's this predictable, why doesn't a better prompt just fix it?

The Complication

"Make it nicer" doesn't move a hardcoded prior

A generic prompt asks the model to try harder, but the tells are baked into its defaults. It slides straight back to the same slop.

So the fix can't be a nicer request — it has to be a checkable rule.

Mechanism — rules, not vibes

Every ban is checkable in 5 seconds by eye or grep

The rewire names each ban as a hard rule you can verify instantly. The context file is explicit: "No vibes — only checkable rules."

And if a ban is checkable, it can be tuned by a switch.

# ai-tells-design.md Every ban below is testable in 5 seconds by visual inspection or grep over the source. No vibes — only checkable rules. # and, at the bottom: If a rule cannot be tested in 5 seconds, it does not belong on this list.
Mechanism — three dials

Three dials are switches, not vibes: cross a threshold, a ban fires

DESIGN_VARIANCE, MOTION_INTENSITY, and VISUAL_DENSITY (1–10, active defaults 8 / 6 / 4) are global switches that drive hard threshold bans.

So what does the model actually produce once the dials are on?

The Turn

Same prompt, same model, completely different output

The bundle's before/after example proves it. The naive baseline packs 8 banned tells into one component; with the bundle, the defaults change.

And because each ban is grep-testable, that difference is machine-checkable.

Naive baseline

  • centered hero
  • h-screen
  • purple→blue gradient
  • Inter font
  • "Elevate" / "Seamlessly"
  • = 8 banned patterns

With the bundle

  • asymmetric 12-col grid
  • min-h-[100dvh]
  • Cabinet Grotesk + Newsreader
  • off-black
  • concrete copy
The Payoff — the grep gate

The ban compiles into a CI script that exits 1 on slop

check-tells.sh is pure bash — 9 Critical + 3 Warning regexes, exit 1 on Critical. Run against a synthetic slop hero, it flagged the tells. Reproduced independently: EXITCODE=1.

Taste is now an enforceable, reproducible gate — not a suggestion.

$ bash scripts/check-tells.sh src/ [CRITICAL] hero.tsx: h-screen [CRITICAL] hero.tsx: from-purple-500 to-blue-600 $ echo EXITCODE=$? EXITCODE=1
The Takeaway

Name the failure modes. Make each checkable. Then enforce.

The portable pattern: turn subjective "taste" into named rules a machine can gate on — checkable in seconds, enforceable in CI.

Good taste stops being a vibe and becomes an enforceable rule.

Sources

Research Methodology

Data as of: HEAD 4cf2f37 on main (2026-05-05) · bundle v0.3.0 · Status: Released and self-tested

Repo: git@github.com:michaeljabbour/amplifier-bundle-taste.git — 6 commits, sole author Michael J. Jabbour (Amplifier port of Leonxlnx/taste-skill, MIT)

Commands run against a fresh ground-truth clone:

Gaps / caveats: The repo's own "≈ 5.3M combinations" figure is arithmetically wrong (the stated formula = ~25.7M) and is not claimed here. Aesthetic outcome quality was not verified — only the presence and enforceability of the rules. The exact violation count is not byte-reproducible; the material behavior (Critical detected, exit 1) is.

Primary contributor: Michael J. Jabbour (6/6 commits, sole author). Original rules by Leonxlnx (upstream, MIT).

More Amplifier Stories