Taste Skills for Anti-Slop UI: Claude, ChatGPT, Kimi & GLM

Taste skills are Agent Skills that encode aesthetic judgment — layout rules, typography discipline, and banned AI-slop patterns — so Claude, ChatGPT, Cursor, Codex, Kimi, GLM, and Gemini stop shipping purple-gradient mediocrity by default. Pair one taste skill with a DESIGN.md and a template from the Agent's Design gallery; do not stack five overlapping "make it beautiful" skills.

What is a taste skill?

A taste skill is a SKILL.md whose job is visual quality control for agents. It does not replace a design system. It tells the model how to decide when generating or redesigning UI.

Typical contents:

• When to activate (landing pages, redesigns, marketing sections)
• Hard bans (gradients, default Inter, icon-card grids, glow spam)
• Process (read DESIGN.md, one section per request, prefer existing primitives)
• Pass/fail checks before the agent claims done

Browse compare demos — same prompt, skill off vs on — in the Skills explorer.

Why every model needs one

Model	Without taste skill	With taste skill + DESIGN.md
Claude	Eloquent rationale, generic layout	Constraints survive long plans
ChatGPT / Codex	Competent React, SaaS-default look	Structure follows your bans
Cursor (Composer)	Fast multi-file slop	Scoped diffs with craft rules
Kimi	Speedy median UI	Same floor as closed models when skill loads
GLM	Capable code, weak prior on brand	Tokens and bans do the taste work
Gemini	Strong multimodal assist	Still needs written craft rules

Taste is not a mysterious Claude-only property. It is instructions plus examples.

Taste skill vs DESIGN.md vs templates vs prompts

Artifact	Owns	Weak alone because…
Taste skill	Decision process + bans	No brand tokens
DESIGN.md	Tokens, type, layout map	Easy to ignore if never triggered
Gallery template	Proven composition	Needs an agent workflow to apply
One-off prompt	This task only	Forgotten next session

The winning combo for Claude or ChatGPT or Kimi:

1. Taste skill (process)
2. DESIGN.md (tokens)
3. Template / section brief (composition)
4. Short prompt ("hero only")

See agent-ready DESIGN.md and anti-slop checklist.

What to put in a taste SKILL.md

Description (retrieval)

description: Use when designing, redesigning, or generating UI, landing pages, marketing sections, or dashboards. Enforces anti-slop layout, type, and color rules. Read DESIGN.md first when present.

Body (minimum viable craft)

# Tasteful UI

## Always
1. Read DESIGN.md before writing UI. Do not invent fonts or colors.
2. One section per request. Stop and wait for review.
3. Prefer existing components over new one-offs.

## Never
- Purple / indigo gradient heroes on dark mesh
- Inter (or system UI) unless DESIGN.md specifies it
- Three identical feature cards with icon-in-circle
- "Unlock the power of…" headline templates
- Decorative motion without meaning
- Drop shadows on every card

## Before done
- [ ] First viewport has one clear hierarchy
- [ ] Accent color only on primary actions
- [ ] Mobile stack checked
- [ ] No new font imports unless specified

That file improves Claude Code, Cursor, ChatGPT, Kimi, and GLM equally.

How to evaluate taste skills (before/after)

Run this protocol on any candidate skill:

1. Fix a prompt: "Build a SaaS landing hero for [product] in Next.js + Tailwind"
2. Run without the skill on your daily model (Claude or ChatGPT)
3. Run with the skill
4. Compare first viewport: type, layout uniqueness, color restraint
5. Repeat once on a second model (Kimi, GLM, or Gemini)

Keep the skill only if the with-skill pass wins on craft — not on word count.

Agent's Design embeds this compare pattern on /skills for eligible design skills.

Recommended workflow (2026)

In Cursor

1. Install taste skill under .cursor/skills/ (see install guide)
2. @DESIGN.md + invoke skill
3. "Implement hero only per skill + DESIGN.md"
4. Browser check → next section

In Claude Code

1. Project skill in .claude/skills/
2. "Use tasteful-ui. Read DESIGN.md. Hero only."
3. Let Claude plan against the checklist, then implement
4. Optional: open the diff in Cursor for local polish

In ChatGPT / Codex

1. Ensure .agents/skills/ (or Codex path) contains the same skill
2. Attach or point at DESIGN.md
3. Same section-scoped prompt as Claude
4. Port code into the real repo if the session is sandboxed

In Kimi or GLM

1. Confirm Agent Skills discovery path
2. Install the same SKILL.md — do not rewrite for the model
3. Use for draft volume; promote winners to Claude/Cursor for merge

Mistakes that keep slop alive

• Five taste skills at once — descriptions collide; agents thrash
• Skill says "be beautiful" — no bans, no effect on ChatGPT or Claude
• No DESIGN.md — skill bans Inter, model picks Roboto, still generic
• Whole-page generation — even Claude loses hierarchy across 12 sections
• Model hopping without the skill — testing Kimi "raw" then blaming the model
• Ignoring demos — install popularity ≠ craft

Pair with gallery templates

Taste skills raise the floor. Templates set the destination.

1. Pick a layout in the gallery
2. Copy DESIGN.md + agent prompt
3. Enable one taste skill
4. Build section by section in Claude, Cursor, or ChatGPT
5. Spot-check with Kimi or GLM using the same files

Related: Ship tasteful UI with AI agents, Stop AI UI looking templated.

FAQ

Do I need a different taste skill for Claude vs ChatGPT?

No. Keep one portable SKILL.md. Adjust invocation paths, not craft rules.

Will a taste skill fix v0 / Bolt / Lovable output?

Those generators are less skill-native. Export to a repo, then run Claude or Cursor with the skill to normalize. See Lovable and Bolt design resources.

Are taste skills the same as Cursor rules?

Rules are always-on preferences. Skills are triggered procedures with richer bodies. Many teams use both: short anti-slop rules plus a full taste skill for UI tasks.

Can Kimi or GLM replace Claude for final UI?

They can draft. Final taste still depends on skill + DESIGN.md + human review — same as Claude or ChatGPT.

Next step

Open /skills, pick a taste skill with a compare demo, install it beside a gallery DESIGN.md, and rerun your last landing prompt in Claude, ChatGPT, or Cursor. The gap between skill-off and skill-on is the entire product thesis: agents need craft instructions, not vibes.