Skip to main content
skill CX Codex

design-style-picker — Claude Code skill: pick a style from images

Generates a batch of visual design directions and lets you choose from images instead of a written brief. Handy when a client cannot put the look into words.

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add daymade/claude-code-skills --skill design-style-picker -a claude-code
Open on GitHub

What design-style-picker does

The skill takes over the taste conversation. Rather than asking you to describe a style, it builds a matrix of options: one axis moves color intensity in large steps, another changes how the layout is organized. The images have to look clearly different, and near-duplicates get regenerated.

Before generating anything it inspects the current UI, reads design tokens, CSS variables, component names and screenshots. Those stay as the starting vocabulary instead of being thrown away. Color is treated as a system: broad zones for architecture, neutral components for routine UI, strong accents for brand and risk. Asking for less color usually means fewer competing focal points, not a black-and-white mockup.

The skill then opens the generated images itself and flags which ones look too colorful, too dead, or closest to the target. Implementation works from the chosen references — principles, not pixels. A full selection session is covered in references/selection-playbook.md.

What it helps with

  • choose an interface style when you cannot describe it in words
  • compare several design-system directions side by side
  • evolve an existing UI without dropping current tokens and components
  • find the too-much-color boundary on deliberate samples
  • turn picked references into an implementation

Example prompts for the agent

  • I can't describe the style in words — generate options and let me pick
  • Build a direction matrix for the landing page: color ladder plus different layout schemes
  • Evolve the design system but keep the current tokens and components

How to install design-style-picker

1Skills CLI (Vercel)

npx skills add daymade/claude-code-skills --skill design-style-picker -a claude-code

2Manually: put the files in place

~/.claude/skills/design-style-picker/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/design-style-picker/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about design-style-picker

How do I install design-style-picker in Claude?

Use the skills installer: npx skills add daymade/claude-code-skills --skill design-style-picker -a claude-code. The skill file then sits at ~/.claude/skills/design-style-picker/SKILL.md.

What does design-style-picker do?

It generates a set of visual directions and lets you choose one from the images. Then it implements the chosen style from those references.

What does the skill need to work?

Your current UI or screenshots, plus tokens and component names if you have them. It builds on existing assets rather than drawing unrelated concepts over them.

Source text

SKILL.md · MIT
---
name: design-style-picker
description: Batch-generate and compare visual design directions so a user can choose the style they actually want. Use when the user says they cannot describe an abstract visual style, asks for many style options, wants to choose from generated UI/design-system images, rejects outputs as too colorful/too dead/too generic, or needs an existing UI/design system evolved without discarding current assets.
---

# Design Style Picker

## Purpose

Use this skill to turn vague taste into concrete visual choices. The goal is not to guess one final design; it is to generate a structured set of options that exposes the user's taste boundary quickly.

## Core Rule

Do not ask the user to describe an abstract style if they already said they cannot. Generate comparable visual evidence, let them pick, then implement from the selected references.

## Workflow

1. **Restate The Real Target**
   - Say what the user is actually choosing: design-system style, business app surface, landing page, deck, component library, etc.
   - Distinguish choosing one direction, retaining several directions, and assigning directions to different visual roles. Multiple options do not imply mutual exclusion; retaining both does not decide how they coexist. Keep an unresolved adoption or fusion choice open while exploring the authorized directions.
   - Separate the primary artifact from validation samples. If the task is a design system, business screens are optional validation samples, not the main deliverable.
   - Preserve any existing UI, assets, tokens, layout, brand cues, and domain context unless the user explicitly asks to discard them.
   - Freeze the existing spatial structure, navigation, zoom/detail levels, and reading behavior outside this round's authorized axes. Name those axes before evolving a product; do not rearrange it into another page merely to make a style comparison easier.

2. **Collect Existing Assets First**
   - Inspect the current rendered UI or screenshots.
   - Before selecting a structural screenshot, match its actual surface and interaction level to the user's target. A local detail screen can anchor a local-screen task; it cannot stand in for a whole workspace or its overview-to-detail flow. Obtain the missing views or use the verified specification, labeling any part not visually observed.
   - Read design tokens, CSS variables, component names, key images, brand/domain references, and existing screenshots.
   - Treat current assets as the starting vocabulary. Do not generate unrelated "fresh" concepts over them.

3. **Generate A Matrix, Not Minor Variants**
   - Use at least two authorized axes when taste is unclear:
     - Vertical ladder: one dimension changes by large steps, such as color intensity 20/35/50/65/80.
     - Horizontal directions: different organization strategies, such as data-driven color, brand spine, warm product imagery, scenario modules, or governance-led layout.
   - Make options visibly different. If two images look like siblings, regenerate one with a clearer contrast.
   - Prefer batch generation. The user is waiting for selection, not watching one slow image at a time.

4. **Use Color As A System**
   - "Less colorful" does not mean black-and-white. It usually means fewer competing focal points.
   - Keep the product palette alive, but assign color roles:
     - Broad zones and section bands for architecture.
     - Data visualization and evidence systems for multi-color semantics.
     - Brand/risk colors for rare, high-signal emphasis.
     - Neutral components for routine UI.
   - Include explicit upper-bound samples when the user is tuning color: safe, middle, high, and overload boundary.

5. **Review Before Presenting**
   - Open generated images yourself.
   - Check target coverage before taste: can the user inspect the requested surface, scale, states, and interaction levels, and do the candidates retain the stated relationship? Repair missing coverage before judging colors or character details. Static images illustrate interaction levels; they do not prove interaction behavior.
   - Mark which are likely too dead, too colorful, too generic, too business-system-like, or closest to the target.
   - Present file paths and a short decision note for each useful candidate.

6. **Implement From Selected Images**
   - Extract principles, not pixels: color roles, layout density, focal hierarchy, component treatment, image use, governance/data placement.
   - Fuse selected references explicitly when that combination is approved. Example: "Use H02 for color placement and V04 for palette intensity." If several directions are only retained for exploration, keep them available without treating retention as approval to fuse or assign roles.
   - Keep implementation scoped to the existing UI unless the user asks for a new artifact.
   - Run rendered visual QA after implementation.

## Prompt Pattern

When generating images, include:

```text
This is an evolution of the existing UI/design system, not a replacement.
Preserve these assets: <tokens, imagery, sections, components, brand cues>.
Target surface and levels: <requested view, scale, states, overview/detail coverage>.
Reference coverage: <what each screenshot proves; missing views or specification-only parts>.
Frozen behavior: <spatial structure, navigation, zoom, reading>.
Direction relationship: <choose one / retain several / approved role assignment; unresolved choices>.
Axis: <vertical ladder or horizontal direction>.
Variant name: <clear label>.
Color/visual rule: <specific budget or organization method>.
Primary focal point: <one thing>.
Avoid: <known rejected styles from the user>.
```

## Lessons To Preserve

- A user saying "not colorful" may mean "no dozens of equal-weight small color chips", not "remove all color".
- A user saying "more weight" may mean visual authority and hierarchy, not dark-mode control room.
- For design-system work, do not replace the system with a business dashboard. Business screens can validate style, but should not become the answer.
- Always create deliberate boundary samples. They make "too much" visible and speed up selection.
- After an approved combination, fuse the chosen references and name what each contributes.

## References

- Read `references/selection-playbook.md` when running a full style-selection session or when the user gives taste corrections during image exploration.

© daymade · MIT · daymade/claude-code-skills

Similar skills

Everything for Design →

From the same set64

marketplace-health-check Runs a parallel audit of a skills marketplace repo: script safety, docs, PII, open PRs and issues, manifest integrity. CX Codex +1 389installs llm-eval-harness This skill probes a third-party LLM endpoint across six dimensions: availability, request fidelity, speed, concurrency, protocol compliance and quality regression. CX Codex +1 372installs github-sensitive-data-cleanup Scans and removes secrets, tokens, and personal data from Git history, then verifies and pushes safely. Use it after accidentally leaking credentials to a GitHub repository. CX Codex +1 353installs claude-switch-models-setup Sets up isolated profiles to run multiple providers like Kimi, DeepSeek, and Anthropic in separate terminal windows. It stops configuration drift while sharing common skills and hooks. CX Codex +1 344installs notify-wecom Sends a single one-off message to a WeCom (Enterprise WeChat) group bot. Handy when the team needs a quick ping without templates or setup. CX Codex +1 341installs setup-notifications-via-wecom Sets up a WeCom group bot webhook and sends technical status messages through it. Useful when a backup, sync or cron job should report on its own. CX Codex +1 341installs local-codex Runs the local OpenAI Codex CLI as a non-interactive sub-agent using ChatGPT OAuth. Ideal for offloading code generation and reviews without API keys. CX Codex +1 333installs wps-doc-scraper This skill archives public WPS/KDocs documents and embedded ProcessOn mind maps as raw data, SVG or PNG, and Markdown. Useful when a link has to be saved without logging in. CX Codex +1 334installs

Project sponsors

Companies that keep this analytics open
Gold sponsor

Impressions, clicks and code copies — open statistics. We keep the numbers public for the same reason NeBlask itself is public. Become a sponsor