Agent Skills for building with Gravity UI ecosystem. Compatible with other agents as well via skills.sh.
npx skills add gravity-ui/skillsBuilds, fixes, styles, and composes UI with Gravity UI. Gravity UI is a 30+-package design system. AI agents get three things wrong:
- Pick the wrong package —
chartsvschartkit, uikitTablevs@gravity-ui/table,date-componentsvsdate-utils. - Use outdated APIs — training data lags the installed version; the agent writes props that don't exist (anymore, or yet).
- Invent component APIs / icon names —
Button variant=,Icon name=, guessed@gravity-ui/iconsnames,Text variant="header-3".
This skill fixes all three.
Use when:
- Working with
@gravity-ui/*imports,uikit,ThemeProvider, or--g-CSS tokens. - Building dashboards, tables, charts, forms, navigation shells, or landing pages on Gravity UI.
- Any prompt (EN or RU) mentioning gravity-ui, гравити, дизайн-система, компоненты, дашборд, таблица, график, or форма.
Upgrades @gravity-ui/uikit from v7 to v8 in an app or a library. The migration guides ship inside the v8 package; the skill adds what they leave to the agent:
- Preflight — React 18+, and the third-party dependencies that still pin uikit v7 (
scripts/check-uikit-peers.mjs). - Inventory and decisions — the affected code grouped by guide section, and the choices that belong to the user:
/legacyor a rewrite, thelbreakpoint. - Verification past
tsc— many v8 changes compile cleanly and show only on screen or in tests.
Use when: bumping uikit to 8, estimating that upgrade, or finishing a half-done one.
Skill auto-activation across agents sits around ~50% in public evals — the skill exists, the agent just doesn't always invoke it.
Paste this into your AGENTS.md:
<!-- BEGIN:gravity-ui-agent-rules -->
# Gravity UI
This project uses the Gravity UI design system (`@gravity-ui/*`). Before writing
or modifying Gravity UI component code, you **MUST invoke the `gravity-ui`
skill** — it routes you to the correct package and prevents invented APIs. Run
it on any UI task here.
<!-- END:gravity-ui-agent-rules -->For Claude Code, also add @AGENTS.md to CLAUDE.md (it reads CLAUDE.md
natively, not AGENTS.md).
Does the skill help? Ten models solved 18 Gravity UI tasks four times each without the skill and four times
with it. A run passes when the project builds, deterministic checks pass (the right @gravity-ui/* package,
no hand-rolled replacement) and an LLM judge confirms the intended Gravity UI solution.
| Model | Agent CLI | Without skill | With skill |
|---|---|---|---|
| Claude Sonnet 5.5 | Claude Code | 67% | 99% |
| Claude Fable 5.1 | Claude Code | 69% | 96% |
| Claude Opus 5.5 | Claude Code | 74% | 94% |
| GPT-6 Astra | Codex | 32% | 92% |
| Claude Haiku 5.5 | Claude Code | 68% | 89% |
| DeepSeek v4 Pro | Codex | 53% | 89% |
| GPT-6.1 Sol | Codex | 29% | 86% |
| GPT-6 Luna | Codex | 15% | 86% |
| GLM 5.3 Flash | Codex | 45% | 81% |
| Kimi K3 | Claude Code | 51% | 79% |
All models together: 50% → 89%. Without the skill agents know @gravity-ui/uikit and hand-write the rest
(navigation, tables, charts, dashboards, landing pages, the markdown editor, the date picker); with it they
route to the right package. Method, per-task numbers, the code and transcript of every run and the harness to
re-run it are in gravity-ui/skills-evals.
Distributed under the MIT License. See LICENSE for details.