anti-slop-ajonai

Build web interfaces that don't read as machine-generated. Use when creating or reworking any frontend — landing page, marketing site, app UI, dashboard, component, or artifact — and whenever someone

SKILL.md

GitHub

Anti-Slop (AJONAI)

The one idea

A tell is not ugliness, and not popularity. A tell is a choice whose frequency is invariant to the brief — the thing that shows up at the same rate whether the page is a mortgage calculator, a techno label, or a hospice.

That reframes the whole problem. Design is a sequence of decisions. Wherever a brief is silent, a generator inherits a default and a designer makes a call. Every tell marks a place where a decision was skipped. Untouched defaults are damning precisely because a designer would have touched them.

So the goal is not to avoid a list of banned things. It is that every visible property be derivable from this brief and no other. A page passes when you can point at any value on screen and say why it is that value here and would be different elsewhere.

Why ban-lists fail

They work once, then become the next tell. This is documented, not theoretical:

Was the escape Became the tell
"Don't use pure white, use a warm off-white" Cream/beige page grounds are now a detector rule
"Escape Inter with an editorial serif, set the hero italic" Oversized italic serif hero is now a detector rule
"Add a tracked uppercase eyebrow for editorial hierarchy" The eyebrow is now the single most-banned element
"Stop shipping heavy shadows, use a 1px hairline" Models did both — hairline + wide shadow is now a named defect

Advice followed uniformly produces uniformity. A rule that says "avoid X" manufactures "everyone avoided X the same way." Treat every specific value in references/ as perishable, and the process below as durable.

Non-negotiable: decide before generating

The only intervention with evidence behind it is front-loading constraints into the first generation. Correcting afterwards measurably underperforms starting from a stronger instruction, and generic output is the mode of the distribution rather than an unlucky draw — so retrying, regenerating, and raising temperature do not escape it. Only changing the conditioning moves it.

Before writing any markup, write the decision sheet. Five axes, because these are exactly the slots a brief leaves open:

SUBJECT     one sentence: what this is, who it's for, its single job
GROUND      the page's world — materials, instruments, vernacular of the subject
PALETTE     4-6 named hex values, each with a role. State the neutral's hue bias.
TYPE        2+ roles (display / text / utility), named faces, and why these
SPACE       two registers: intra-component values and section rhythm — different numbers
SHAPE       one radius value, one elevation policy (tonal step OR shadow, not both)
MOTION      a count of animated moments and a duration band, or "none"
SIGNATURE   the one element this page is remembered by

Then critique the sheet before building: work through a similar prompt in your head and see if you arrive somewhere similar. Any line that would be identical for a different subject in the same category is a line you inherited. Revise it and say what changed.

Read references/00-decision-sheet.md for how to fill each line, with worked examples.

Weight interaction states above appearance

The counterintuitive finding, and the one that most separates this skill from others: when a large pool of candidate tells was adversarially audited, motion and interaction states had the highest survival rate; hero composition and iconography had the lowest.

A screenshot of a generated page is usually fine. The giveaway is that it is inert. Missing :active, no :focus-visible, no disabled styling, no loading, empty, or error state, untouched ::selection and caret-color, a scrollbar nobody looked at.

So: ship the states that are absent before restyling the states that are present. Cheapest, highest-signal work available. See references/04-states.md.

Reference map

Load only what the current task needs.

File Load when
references/00-decision-sheet.md Always, before generating
references/01-color.md Choosing or fixing a palette, gradients, dark mode, contrast
references/02-type.md Choosing faces, scale, tracking, leading, measure
references/03-space-surface.md Spacing, elevation, radius, borders, dividers
references/04-states.md Any interactive element — read this by default
references/05-motion.md Any animation or transition
references/06-assets.md Icons, imagery, charts, placeholder data
references/07-audit.md Reviewing existing UI, or self-checking before shipping
references/08-dated-defaults.md The current monoculture. Check the date at the top.

The checker

A dependency-free static checker ships with this skill. It reads source, not renders, so it is honest about its limits.

node scripts/check.mjs <paths...>          # files or directories
node scripts/check.mjs src/ --json         # machine-readable
node scripts/check.mjs src/ --only=color   # one category
node scripts/check.mjs src/ --strict       # advisory findings count as failures

Findings are error, warn, or advisory. Advisory never fails a run without --strict, because those rules have real false-positive rates. Suppress a checked-and-accepted finding with /* deliberate-ignore rule-id */ on the line above, or <!-- deliberate-ignore rule-id --> in HTML. A suppression is a decision, which is the point.

What this skill refuses to do

Other anti-slop rulesets contain real defects. Do not reproduce them:

  • No blanket typeface bans. Banning Helvetica and Arial outlaws two of the most-studied faces in the history of the discipline. The rule is share-and-context: is this face doing the display work of a page whose brief never asked for it? A system stack for body text is fine.
  • No total em-dash ban. Prescribing a hyphen where an en dash is correct damages typography to evade a classifier. Optimize for the reader, never for the detector.
  • WCAG large text is 24px, or 18.67px bold (18pt and 14pt at 96dpi). Several published skills state 18px, which lets 18–23.99px text pass at 3:1 where AA requires 4.5:1. That is an accessibility regression introduced by an anti-slop rule.
  • Don't treat human tells as AI tells. Off-scale magic numbers, a double space, sub-pixel misalignment, a novelty display face — these are evidence against machine authorship. A rubric that scores them as positives misfires on hand-built amateur sites, which are exactly the pages most likely to be falsely accused.
  • Don't claim a single element proves authorship. Individual properties are weak evidence; co-occurrence is strong. When three independent judges rated a large pool of candidate tells, under 10% were unanimously high-signal. Say "this reads as inherited," never "this is AI."

Working rules

  1. Pin the subject first. If the brief doesn't say what the thing is, decide and state it. Distinctive choices come from the subject's own world, not from a style library.
  2. Spend boldness once. One signature element carries the page; everything around it stays quiet. Uniformly-distributed effects read as generated.
  3. Two spacing registers minimum. One gap value everywhere asserts a flat, equal-weight structure the content does not have. Proximity is meaning.
  4. Commit in a direction. Detectors are built so that pure white passes and a committed warm ground passes — only the timid middle fires. Timidity is the actual failure.
  5. Real content, real assets. Placeholder services and invented companies are the most identifiable artifacts in generated work.
  6. Check the brief's own words first. If the user asked for a purple gradient, build a purple gradient and build it well. Their instruction always outranks this skill.