pixel-anti-slop — the taste gate (O-HS1..7 + O-F1 + impeccable)

Stance: ~40–45% of generated UI is slop until it survives this gate. pixel's moat is the gate, not the generator (OPINIONS O7). Run this on your OWN output too — an agent that emits AI-purple and a rounded-icon-tile-above-every-heading and calls it done has lost the plot.

trigger

• generating or reviewing ANY web UI (a hero, a card, a page, a component)
• a fresh v0 / Lovable / Bolt draft landed (untrusted — O7)
• "does this look AI-generated?" / "is this hero slop?"
• before you call a mockup "done" (the §0-of-taste check)

The stance this encodes (named, not generic)

The machine-detectable slop tells — BLOCK on sight (impeccable flags these deterministically):

1. Pure/near-black #000/#09090b → use warm-black #141312 (O-HS1). Untinted black is a top slop tell.
2. Inter-for-everything → pair a serif display + sans body (Instrument Serif + Poppins) (O-HS2/O-F6). Reading-page override, October 2026: on long pages people read (a brain, a plan, a report, an approach page), no serif or stretched display face; use Instrument Sans alone on light paper (pixel-brain-page). The pairing holds for product interfaces only. Serif+sans is the single highest-impact anti-AI change. Inter is "the black t-shirt of fonts" (Mantz) — neutral-by-design, not safe-by-default.
3. AI purple/indigo (or purple→blue gradients) → warm-black + ONE non-purple OKLCH accent (O-HS4/O-F1). This is NOT a taste choice the model makes — it's a training-data artifact: Wathan made every Tailwind UI button bg-indigo-500, it saturated the corpus, the model now emits it as "statistically correct." O-F1 HARD-OVERRIDES energy react.md's "Purple theme #6b21a8".
4. Symmetric card grids → asymmetric / bento (O-HS3). Symmetric = the AI pattern.
5. Cards-nested-in-cards, dark glows, bounce/elastic easing, a rounded-square icon tile above every heading → all literal impeccable rules (O-HS4).
6. Placeholder gray boxes / pure-white text on dark → use a gradient/mask-*/icon instead of a gray box; zinc-100/200, never #fff on dark (O-HS5).
7. A loading spinner with no context → stream a rich card ("Searching restaurants…") (O-HS7). Generative UI is the signature.

The restraint rules:

• ONE brand color-moment per page, OKLCH, on Radix 12-step scales with a purpose-per-step (1-2 bg · 3-5 component · 6-8 borders · 9-10 solid · 11-12 text) (O-HS6).
• Commit to ONE aesthetic direction within a surface, match intensity to purpose (marketing = spectacle ok; dashboard/a11y-first = restraint). NEVER ship averaged or unmotivated maximalism (O-F8).

steps

1. Run the deterministic gate FIRST (keyless, exit-code, $0):

   npx impeccable detect --json <dir>/src   # exit 0 = 0 slop violations; non-zero = blocked

   This is the cheapest, most objective check — it encodes the 41 rules above. A non-zero exit is a BLOCK, not a warning.
2. Apply the human-eye pass the linter can't fully cover: is there ONE color-moment (not three)? Is the layout asymmetric? Is the type a serif+sans pair? Are tool-states rich cards, not spinners?
3. If any tell is present → fix it, re-run. Do not rationalize "it's close enough" — O11: NEVER accept "7/10"; either the dimension meets the bar or the fix is explicit.

output

• impeccable detect exit code (0 = pass) + the list of any slop rules hit.
• For each violation: the OPINIONS stance it breaks (O-HS#/O-F#) + the one-line fix.
• Verdict: SHIP (0 violations + human pass) or DRAFT — must fix (the named tells).

verify

npx impeccable detect --json <dir>/src exits 0 AND no #000/#09090b/bg-indigo/font-inter-as-sole-face in the emitted tokens.

sources

brain/OPINIONS.md O-HS1..7, O-F1, O-F6, O-F8, O7, O11 · brain/KNOWN-PITFALLS.md §7.4 · impeccable (npm, keyless, 41 rules).