Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Design & UI

charles-polish

Push a working interface from fine to exceptional: hill-climb it through levels, fix the invisible details, and make motion correct frame by frame. Use when a screen or component works but feels off or unfinished, when asked to polish, refine, take it further, or make it feel great, when adding or reviewing animation, or when reviewing UI against a high craft bar. Triggers on polish, feels off, make it better, level up, hover and press states, focus rings, tabular numbers, layout shift, hit areas, optical alignment, easing, springs, duration, transitions, choreography, reveal, morph, flash, jank, stutter, sound and haptic feedback. If the screen still looks AI-generated, run charles-unslop first.

LemonFace0309 · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add LemonFace0309/charles-skills --skill charles-polish

About this skill


name: charles-polish description: "Push a working interface from fine to exceptional: hill-climb it through levels, fix the invisible details, and make motion correct frame by frame. Use when a screen or component works but feels off or unfinished, when asked to polish, refine, take it further, or make it feel great, when adding or reviewing animation, or when reviewing UI against a high craft bar. Triggers on polish, feels off, make it better, level up, hover and press states, focus rings, tabular numbers, layout shift, hit areas, optical alignment, easing, springs, duration, transitions, choreography, reveal, morph, flash, jank, stutter, sound and haptic feedback. If the screen still looks AI-generated, run charles-unslop first."

Polish

Built from Emil Kowalski's UI polish and animation skills, Josh Puckett's Interface Craft, and feedback from Charles Liu's "Removing AI Slop" series.

Most work stops at level 3 of 10, because it works and it's easy to stop. The difference between fine and great is rarely the idea; it's how far the idea gets pushed. Most "something feels off" problems are one wrong note, so look, measure and name it before changing anything.

Run it

  1. Render it, at its real size and at the smallest size it will be seen. Capture each state and each motion as frames. Measure in the pixels people see: on a scaled stage, a 42 px button may be 37 px on screen.
  2. Score it against the anchors below and name the weakest thing on screen in one sentence.
  3. Structure: fix hierarchy and containers before details. Polishing the wrong structure is wasted work.
  4. States: map every state, including empty, loading, error and success, and design the unhappy ones.
  5. Details: run the details table.
  6. Motion: walk the three decision trees, then storyboard.
  7. Verify frame by frame (below). Nothing is done until you've looked at the frames.
  8. Climb: run the loop until two passes in a row find nothing that changes the score.

If you were asked to review and not to change anything, stop before step 8 and report each finding as severity, what's wrong, the evidence (frame, number or file and line) and the fix. High means a state is broken or a promise the screen makes isn't kept. Medium means people notice. Low means only a designer does.

Score anchors

ScoreIt means
3It works, and looks like everything else
5Designed at rest, but unhappy states or motion are default
7Every state is designed and motion is correct; nothing is memorable yet
8One thing people would screenshot, no visible defect at the viewing size
9That thing holds up in every state and input path, and a rival would copy it
10Don't award this to your own work

The climb

Think in levels. Early levels fix what's broken, and you can see what to do. Later levels are invention, and you can't, so use a move:

Stuck becauseMove
You don't know what better isZoom in. Take one element at 3×. What if this were the only thing on screen?
It's busyRemove something. Did you actually need it?
It feels offName it in one sentence, even before you know the fix
It's fine but plainReference the best. Find the world-class version and list what it does that yours doesn't
You keep tweaking one ideaGenerate more. Make three genuinely different versions, compare at real size, keep one
You can't see it anymoreCritique. Get a fresh reader with no context, or re-read your own work as a rival would

Rules for climbing:

  • Each level is one nameable move with a reason. "Add a spring" isn't a reason; "it can be redirected mid-flight, so it never blocks you" is. If you can't say why in a sentence, the change isn't honest.
  • Make the first leap the biggest. Going from default to designed must be obvious in a still. If before and after look alike, the default wasn't default enough or the change wasn't structural.
  • Each level is strictly better than the last. Never ship a deliberately broken step; a demonstrated bug looks like a bug.
  • Name the thing people would screenshot. If there isn't one, that's the next job.
  • Push the concept into the material, not effects on top of it. If the idea is a pressure gauge, the needle is the loading state, not a spinner beside it.
  • Undo your own decisions when they stop holding up. Restraint counts as a level.
  • Land success as the peak. The end of a flow is what people remember. If success is only a label swap, push the concept into it: the thing the screen has been holding back happens now.
  • Every input path reaches the same ending. Pointer, keyboard, a slow gesture and reduced motion may travel differently, but all of them end with the same state and the same best detail.

Details

Absence is what gets noticed, so run all of these. If you're short on time, start with form behavior, hit area, contrast and layout shift, because those are the ones people feel. Express them in the project's own styling system.

CheckRuleWhy
Fonts-webkit-font-smoothing: antialiased; subset filesCrisper text, nothing unused shipped
Field and buttonA resting button and a field never share one fill. The field stays the brightest; a waiting button sits one tone downTwo identical white slabs hide which one you type in
Cast lightWindow shadows and light patterns stop at an object's edgeBars running unbroken from the table across paper read as seeing through the paper
Dynamic numberstabular-numsDigits that change width make neighbors jiggle
WeightNever change font-weight on hover or selection; use colorIt reflows text
Layout shiftFixed dimensions or aspect-ratio for skeletons, images, async content; reserve space for messagesJumping content reads as broken
Typography… ’ “ ”; text-wrap: balance on hea
Read the complete source ↗