design-helper — Claude Code plugin: compare CSS, install
The design-helper plugin puts CSS, layout, and motion options side by side in a browser, then applies your choice to the project. Use it before committing to a UI change.
MIT ★ 1,596 updated 2026-10-29
/plugin marketplace add hashgraph-online/awesome-codex-plugins
/plugin install design-helper@awesome-codex-pluginsWhat design-helper does
design-helper reads the current component values and the project’s style system, then builds a browser comparison sheet. Each round changes one property, such as shadow blur, spacing, or a border, while the other values stay fixed. Every candidate shows its value and the cost of changing it, and the sheet gives a recommendation with a reason.
The comparison sheet, local server, and screenshots live in a temporary folder outside the project. After you choose an option, the plugin checks what it will affect, edits the real code, and verifies the result. Tailwind classes and CSS variables use the project’s own values; SCSS and CSS-in-JS values are copied into CSS variables for the sheet. Motion is judged in the browser, while keyboard behavior and screen reader checks are outside its scope.
You need a browser and python3. Without python3, the sheet opens through file://, with no auto-reload or automatic temporary-folder cleanup. Playwright is optional for computed-style checks and screenshots. Replies and comparison-sheet text are in Korean.
What it helps with
- compare CSS values side by side in a browser
- test spacing, borders, or shadows before editing the UI
- see which code locations a chosen value will affect
- apply the selected candidate to the real component
- judge motion while it runs in the browser
Example prompts for the agent
- Show me 5 versions of the card’s shadow blur without changing project code.
- Compare several spacing values in the existing UI, changing only spacing.
- Go with G, check what it touches, and apply it to the card.
How to install design-helper
Plugin marketplace
/plugin marketplace add hashgraph-online/awesome-codex-plugins/plugin install design-helper@awesome-codex-pluginsStep by step, with commands for each platform: Claude Code plugins: marketplaces, installing, and building your own
Questions about design-helper
How do I install design-helper in Claude Code?
Inside a Claude Code session, run /plugin marketplace add hashgraph-online/awesome-codex-plugins, then run /plugin install design-helper@awesome-codex-plugins. Start a new session so the skill loads.
What does design-helper do?
It builds a browser sheet with side-by-side candidates for CSS, layout, or motion changes and leaves project code alone until you choose one. It then checks the impact and applies the selected value.
What does design-helper need to run?
You need Claude Code or Codex, a browser, and python3. Playwright is optional for computed-style inspection and screenshots, while replies and sheet text are in Korean.
Source text
design-helper · MIT
# design-helper
[](https://github.com/Dessert99/design-helper/commits/main)
[](https://github.com/Dessert99/design-helper/releases)
[English](README.md) | [한국어](README.ko.md)
> A Claude Code and Codex skill for seeing design options before changing your code.
> [!NOTE]
> Replies and every string on the comparison sheet are in **Korean**. See [Language](#language) to change that.
**1. "Show me five options": Compare them while your project code stays untouched.**

**2. "Go with C": Apply the chosen value to the code.**

## Why
Ask an agent to "show me five options" and it may edit your product code before you
have chosen one. Then you have to undo changes just to keep comparing. Judging CSS
values from numbers alone also leads to repeated edits and reloads.
design-helper **separates seeing options from applying one**. It creates comparison
files outside your project and puts the candidates side by side in your browser. Your
project code changes after you pick an option.
Use it to compare borders, spacing, shadows or layouts in an existing UI before you
commit to a change.
## What it does
A typical round:
```
you Show me the card's shadow blur in 5 versions
→ reads the project's style system and the card's current values
→ draws cards A–E with only the blur changed: 4 · 8 · 12 · 16 · 24px
(color, opacity, offset and the rest stay as they are)
→ labels each specimen with its value and cost:
토큰 그대로 · --shadow-md 고침 (23곳) · 새 토큰
(token as-is · edits --shadow-md in 23 places · new token)
→ under the cards, notes where the differences stop showing and which
one it recommends, and why
→ leaves project code untouched so far
you C is better, but a bit darker
→ adds an opacity ladder below, starting from C (F · G · H …)
you Go with G
→ checks what else it touches, applies it to the real card, verifies
the result, and reports back
you Done
→ stops the server and deletes the temporary files
```
Only one property changes at a time and everything else stays put, so any difference
you see comes from that property. Each value you pick carries into the next comparison.
- **Your code stays untouched while you compare.** The sheet, server, notes and
screenshots live in a temporary folder outside your project. Ask for more options
and the comparison continues there.
- **Your pick is the apply command.** Say "Go with G" and the skill checks the impact
before editing the code. You don't need to copy values yourself.
- **The sheet explains itself.** Each section starts with what to decide, where to look
and how to answer. Each specimen shows its value and cost, plus what it gains and
loses when that matters. A table of contents on the left lists every request, so you
can jump back to an earlier comparison.
- **You don't refresh.** When the sheet changes, the skill reloads the tab and brings it
to the front. If the agent can't control a browser, it says so and gives you the URL.
- **Your conversation stays light.** A subagent reads the code, draws the sheet and
checks it. The main conversation passes messages and applies your pick, so a long
comparison doesn't crowd out the rest of your work. In a test run the main agent grew
by about 5k tokens over four rounds. Where subagents aren't available, the main agent
does the work itself.
- **It draws in your style system.** Tailwind and CSS variables are drawn with your real
classes and tokens. The browser can't read SCSS variables or CSS-in-JS themes, so their
values are copied into the sheet as CSS variables. With no style system, it writes
values directly.
## What it doesn't do
- **It doesn't make up a design.** It shows a range of values to compare, not a finished
look. It recommends one option and says why, and notes what each option gives up,
but you make the pick.
- **It doesn't plan your next step.** No "let's look at color next" and no list of
remaining steps. You decide what to look at next.
- **It can't check behavior.** Focus traps, keyboard handling and screen readers can't
be judged by eye, so they're out of scope.
- **You judge motion.** Screenshots only capture a still frame, so animation is settled
by what you see in the browser and tell it.
- **It doesn't automatically check that code stayed untouched before your pick.** Its
instructions keep comparison files outside the project, but there is no `git diff`
check for unexpected product code changes yet.
## Requirements
- [Claude Code](https://claude.com/claude-code) or Codex
- A browser
- `python3` to run the local server that serves and auto-reloads the sheet. Without it
the sheet opens over `file://`, with no auto-reload and no automatic cleanup of the
temporary folder after two idle minutes.
- Playwright (optional) to read computed styles and take screenshots. Without it that
check is skipped.
## Install
### As a plugin (recommended)
This repository is a plugin marketplace for both agents.
```sh
# Claude Code: inside a session
/plugin marketplace add Dessert99/design-helper
/plugin install design-helper@design-helper
# Codex: from the shell
codex plugin marketplace add Dessert99/design-helper
codex plugin add design-helper@design-helper
```
Start a new session afterwards so the skill loads. To update, refresh the marketplace
(`/plugin marketplace update design-helper` in Claude Code,
`codex plugin marketplace upgrade` in Codex) and install again.
### From a clone
Use this if you want to edit the skill, for example to [change its language](#language).
Clone it anywhere and symlink the skill folder into your agent's skills directory. The
symlink points at the clone, so your edits take effect right away and there's no copy to
keep in sync.
```sh
git clone https://github.com/Dessert99/design-helper.git ~/src/design-helper
ln -sfn ~/src/design-helper/skills/design-helper ~/.claude/skills/design-helper # Claude Code
ln -sfn ~/src/design-helper/skills/design-helper ~/.codex/skills/design-helper # Codex
```
Both agents read the `description` in `SKILL.md` when a session starts, and the rest
(`SKILL.md`, `worker.md`, `references/`, `assets/`) when the skill runs. Edits to the body apply
immediately; a changed `description` needs a new session.
A symlinked skill has no plugin agent, so Claude Code hands the work to a general-purpose
subagent that reads `worker.md` instead. It works the same way.
To remove it, delete the symlink (`rm ~/.claude/skills/design-helper`). The clone stays.
## Usage
Talk as you normally would while building or changing UI. Requests like these trigger it:
"Show me," "compare," and "C is better, but darker" keep you in comparison mode.
Project code stays untouched until you choose a specimen.
- "Show me 5 versions of the button radius"
- "Which is better, A or B?"
- "Compare the card padding"
- "This header feels off, what should I do?" If it isn't clear what to change, it asks
one question at a time.
- "Lay out the dashboard." With no design yet, it starts with a few very different
combinations. The one you pick isn't written to code; it's noted as a starting point,
and properties are then settled one at a time.
Specimens in the browser are lettered A · B · C. You answer in chat.
| You say | It does |
|---|---|
| "Go with C" · "I like B" | Takes it as your pick, applies it to the code and checks it. A combination picked on a blank slate is only noted as a starting point |
| "C is better, but darker" | Takes it as a request to compare more and adds a ladder below |
| "Nice" | Asks which specimen you mean |
| "Done" | Stops the server and cleans up the temporary files |
Example conversations are in the [usage flows](docs/flows.ko.md) (Korean only). They
cover 11 situations, including not knowing what to change, comparing hover and focus
states, comparing animation, and changing a value you already picked.
## Compared with similar skills
Checked in September 2026 against the commits linked below. Most of these skills help
you decide **what to make**; design-helper helps you decide **what value a thing you've
already decided on should have**. They work well together.
**[frontend-design](https://github.com/anthropics/skills/tree/33375500bcea98d610eb30ce10ac4e59b89c390d/skills/frontend-design)** (Anthropic, official)
- What it does: has the agent think like a design lead and pick bold, deliberate styling over the usual AI look.
- Difference: it doesn't draw candidates to compare; the agent sets the taste. design-helper draws candidates and recommends one, and you pick.
- Use it when: you want a polished first design from a blank page in one go. design-helper won't come up with a direction for you.
**[playground](https://github.com/anthropics/claude-plugins-official/tree/aecd4c852f10b466245f18383fa6aad8c0b10d57/plugins/playground)** (Anthropic, official plugin)
- What it does: builds a single HTML file with controls on one side and a live preview on the other, and you copy the result out as a prompt to paste back into Claude. Besides design, it has templates for data, code structure, and document review.
- Difference: one preview carries sliders for several properties, so you see one state at a time. design-helper fixes candidates that change one property and puts them side by side, then applies your pick to the code and checks it instead of handing back a prompt.
- Use it when: you want to get a feel for a combination by playing with it, with no single right answer, or you need an explorer for something that isn't design.
**[impeccable](https://github.com/pbakaus/impeccable/tree/9d715cc4f5564a990ca8345abfdd5df6dc9b41c8)** (pbakaus)
- What it does: a design vocabulary, anti-pattern rules, and commands like `audit` · `polish` · `typeset` · `layout`. In live mode you pick an element in the running app, flip through variants in the browser, and the one you accept is written to source. It reads your CSS tokens and computed styles, and variants can have tuning sliders.
- Difference: the closest match. Its variants rebuild the whole element, so several properties change at once, and you see them one at a time. design-helper puts a ladder that changes one property on one screen, side by side, and labels each specimen with its token impact.
- Use it when: you want to pick directly in the real app, or run a design audit and polish with one command.
**[visual companion](https://github.com/obra/superpowers/blob/8ca22dba9a94f28898bbce59f2537ff4d87c747d/skills/brainstorming/visual-companion.md)** (part of the brainstorming skill in obra/superpowers)
- What it does: while you brainstorm before building, it shows mockups, diagrams and side-by-side comparisons in the browser, and you click to pick. It also helps with spacing and visual hierarchy.
- Difference: it doesn't make one-property ladders, applying the pick to code, or checking it afterwards part of the process. In design-helper those are the default.
- Use it when: you need visuals that aren't UI, like architecture diagrams, or you're setting direction before building.
**[design-shotgun](https://github.com/garrytan/gstack/blob/2a113ae7e623f590095bcaaa0cc581c9a10a6632/design-shotgun/SKILL.md.tmpl)** (garrytan/gstack)
- What it does: generates 3–8 different mockups as images, shows them on a comparison board, collects ratings and notes, and generates again. It remembers the taste you approved.
- Difference: the output is images (PNG), not code, so another skill has to apply it, and it needs an image-generation tool. design-helper draws in CSS with no extra tools.
- Use it when: you don't know yet what you want and want to see many directions.
**What design-helper does differently**: it keeps comparison files outside the project until you choose. It shows one-property ladders side by side with impact labels and a recommendation, then applies and checks your pick.
**Where design-helper falls short**: it doesn't propose a finished design direction, doesn't detect anti-patterns, compares on a separate sheet rather than in the live app, and replies only in Korean by default.
## Language
The instructions are in English, but chat replies and everything on the comparison
sheet are in Korean. Code, class names, file names and commit messages stay in English.
To change it, install [from a clone](#from-a-clone) and edit the `## Language` section in
`skills/design-helper/worker.md`. A plugin install is replaced on every update, so edits
there don't last. The Korean phrases quoted in the instructions (`토큰 그대로`, `끝났어`
and so on) need to change too.
## Docs
- [`SKILL.md`](skills/design-helper/SKILL.md): what the main agent reads. It hands the work to a subagent and passes messages between it and you, so the main conversation stays small. Without subagents, the main agent does the work itself
- [`worker.md`](skills/design-helper/worker.md): the full instructions the subagent follows. The work unit, drawing rules, reading replies, checking, finishing
- [`agents/worker.md`](agents/worker.md): the Claude Code plugin agent that runs `worker.md`. Codex plugins can't ship agents, so Codex spawns a plain subagent instead
- [`references/`](skills/design-helper/references/): detailed instructions read when needed
- [`clarification.md`](skills/design-helper/references/clarification.md): asking about a vague request
- [`sweeping.md`](skills/design-helper/references/sweeping.md): candidate values, captions and the recommendation, background and size, starting from a blank slate, wireframes
- [`writing.md`](skills/design-helper/references/writing.md): how every Korean sentence the user reads is written: short sentences, no symbol shorthand, attached particles, the caption layout
- [`sheet.md`](skills/design-helper/references/sheet.md): the sheet page's request dividers, table of contents, layout and style
- [`stylesystems.md`](skills/design-helper/references/stylesystems.md): detecting the style system and drawing in each one
- [`controls.md`](skills/design-helper/references/controls.md): optional sliders: when to add one and the rules it has to follow
- [`motion.md`](skills/design-helper/references/motion.md): comparing animation
- [`liveview.md`](skills/design-helper/references/liveview.md): the temporary folder, local server, browser tabs, cleanup
- [`assets/`](skills/design-helper/assets/): fixed files the worker copies instead of writing them each time
- [`sheet.html`](skills/design-helper/assets/sheet.html): the empty comparison sheet with its style, table of contents and auto-reload
- [`serve.py`](skills/design-helper/assets/serve.py): the local server that cleans itself up after two idle minutes
- [`check.cjs`](skills/design-helper/assets/check.cjs): reads computed styles and takes a screenshot with Playwright in one run
- [`docs/flows.ko.md`](docs/flows.ko.md): usage flows by situation (Korean)
- [`scripts/bump-version.sh`](scripts/bump-version.sh): sets the version in every plugin manifest, then commits and tags the release
## License
[MIT](LICENSE)
© Dessert99 · MIT · hashgraph-online/awesome-codex-plugins