uxfix

Original：🇺🇸 English

Translated

2 scripts

Fixe automatiquement les défauts UI détectés : boucle détecter → corriger → revérifier jusqu'à un score de 99/100. Contrastes WCAG, halos colorés, easings rebond, interlettrage, palettes IA, copy marketing.

3installs

Sourceblackangel242/skill-market

Added on2026-10-03

NPX Install

npx skill4agent add blackangel242/skill-market uxfix

SKILL.md Content

View Translation Comparison →

UxFix: autonomous UX fixing loop

Purpose

Don't just report UI defects: fix them alone, re-verify, and loop until the score reaches 99/100. The

impeccable

detector provides the analysis (findings + severity); UxFix applies the mechanical corrections and iterates. The agent supervises; the machine does the rounds.

Tooling

Engine:

<skill-dir>/engine/uxfix.js

(Node, no dependencies). Detector: the

impeccable

binary (installed globally by

market install uxfix

).

bash

node <skill-dir>/engine/uxfix.js <fichier-ou-dossier> [--target 99] [--max-iterations 6] [--dry-run] [--json]

IMPECCABLE_BIN

env overrides the detector path.

--dry-run

previews fixes without writing.

--json

emits a machine-readable report.

Workflow

Run
```
uxfix
```
on the built UI files (single HTML/CSS file or directory). It scans
```
.css
```
files and
```
<style>
```
blocks in HTML.
Read the report: score progression per iteration, fixes applied, manual remainder.
Handle the manual remainder yourself: findings with no auto-fixer (unknown rules, brand palette choices the engine neutralized, copy the dictionary missed) need human judgment.
Originals are safe: before the first write, everything is copied to
```
.uxfix-backup/<timestamp>/
```
next to the target.

The 99% contract

Score = 100 − 1 per unique warning − 5 per unique error. Duplicates of the same finding count once. Advisories are informational and don't count (same as the detector's own failure semantics).
99+ means zero primary findings, or a single trivial warning left.
The loop keeps fixing while the fixers make progress (max 6 iterations), then reports what's left instead of churning.

Auto-fix catalog

Finding	Automatic fix
`low-contrast`	Recomputes the offending color with real WCAG math (binary search toward black/white) until the required ratio is met; touches only the `color` / `background` declarations involved
`gradient-text`	Replaces the gradient with its darkest stop as a solid color; removes `background-clip: text`
`dark-glow`	Zero-offset chromatic `box-shadow` → neutral elevation shadow; chromatic `text-shadow` → `none`
`wide-tracking`	`letter-spacing` above 0.05em → `0.02em` ( `0.04em` for uppercase labels)
`cream-palette`	Flagged warm page background → `#ffffff`
`ai-color-palette`	Violet/purple hues (265–300°) rotated toward blue, saturation/lightness kept
`bounce-easing`	`cubic-bezier` with overshoot → `cubic-bezier(0.4, 0, 0.2, 1)`
`pulsing-dot`	Infinite pulse animation on the flagged dot → `animation: none`
`marketing-buzzword`	Dictionary rewrite of EN/FR buzzwords in HTML text nodes (scripts, styles, code blocks untouched)

Details and limits per rule:

references/fix-catalog.md

. Scoring details:

references/scoring.md

.

Operating rules

Run on the real built files, not on mockups: the detector resolves computed styles.
Palette changes (
```
ai-color-palette
```
,
```
cream-palette
```
) are applied with safe neutrals but must be flagged for human review: brand colors are a choice, not a bug. The report lists every one.
Copy rewrites use a fixed dictionary; nuanced marketing copy still deserves a human pass.
A 100/100 detector score is defect evidence, not proof of good design. Keep taste in the loop: run the loop, then look at the page.
Never run with
```
--target 100
```
as a hard gate in CI unless a human reviews the diff: the last point is sometimes a judgment call.