Sergey Chudarin

A design agent for Figma. A plugin for Claude Code. The agent works in your file, from your components and variables. Plans flows, builds screens, finds design system debt, fixes bindings, makes presentations, and writes text that doesn't read as generated.

Agents can write into Figma now. Left to themselves they design badly and break the file quietly: a variable that never bound, a component set that deleted itself when its last variant moved out, a screen assembled from hardcoded values that looks right until someone switches the theme.

These skills fixed exactly those breakages on real product files. A skill for each stage of design work, plus a pack of recorded Plugin API failures with the way around each one. Every rule here comes from a real breakage, not a reading of the docs.

Overview and install on one page: skills.chudarin.com · Русская версия

👋 Who it's for

Designers. You install it once, then work in a chat: paste a Figma link, say what you need in one sentence. No code, no code editor, no plugin to build.

What that looks like in practice: you paste a link to a frame and write "design a settings screen for this app". In a project it hasn't seen, the agent first looks for the answers it needs, a design system document, existing screens, the variables already in the file. What it didn't find, it asks you, at most two questions, and writes the answers down so it never asks again. Then it works from your components and variables, reads back every binding it set instead of assuming it applied, and shows you the result. You ask to change one element, and it changes that element, not the screen around it.

🧩 The skills

flowchart LR
  plan["🗺️ Plan<br/>plan-user-flows"] --> design["🎨 Design<br/>design-screens"]
  audit["🔍 Audit<br/>audit-design-system"] --> design
  design --> fix["🔗 Repair<br/>figma-fix-variable-bindings"]
  rules["🧰 figma-plugin-api-rules"] -. before every write .-> design
  rules -.-> fix
  text["✍️ writing-texts"] -. interface text .-> design
  text -.-> slides["🖼️ Slides<br/>design-slides"]

Skill	You type something like	What happens
plan-user-flows	"What screens do we need for checkout?"	You get every screen and state, entry and exit points, and a page structure for the Figma file, before anything is drawn
audit-design-system	"Go through this file and tell me what the design system holds"	The agent reads the file and reports components, tokens, patterns and design debt, and writes documentation pages if you ask. Read-only otherwise
design-screens	"Design a settings screen" · "Change the header on this frame"	The design process itself: where the style comes from, a brief per screen, checks afterwards, edits to approved work one element at a time
design-slides	"Make a pitch deck from this doc" · "Fix this deck, it looks generated"	Story first, one claim per slide, style from your template or brand, numbers checked against sources, export checked before you get it
figma-plugin-api-rules	nothing, it loads on its own	Real Plugin API failures and the way around each one. Loads before the agent writes anything into your file
figma-fix-variable-bindings	"Variables on this page are detached, fix them"	Finds detached, orphaned and hardcoded values, shows you a report, and rebinds only after you approve it
writing-texts	"Rewrite this so it sounds like a person" · "Check the button labels"	Plain words instead of formulaic ones, in Russian and English: interface text, docs, articles, slides. Learns how you write from your corrections, with your consent, in ~/.agents/voice.md

Every skill has a SKILL.md and topic files. Some of those make up the Plugin API pack: components and variants, instances, layout and geometry, text and fonts, variables and modes, connectors, annotations, publishing hygiene, FigJam, prototype links, the tool layer, building a screen from code. The rest carry the audit checklists, the per-screen briefs, the checks that run afterwards, and the UX rules routed by what the screen contains. All plain Markdown. Nothing runs on your machine, nothing phones home.

💡 Why it's not just prompting

Most of what goes wrong between an agent and Figma is silent. The call returns success, the file looks plausible, the binding was never applied. setBoundVariable('cornerRadius', v) binds four per-corner fields and the aggregate key you checked stays empty. A licensed font missing from the sandbox makes text writes fail in ways that read as your mistake. An instance's children omits hidden descendants, so the inventory you built from it is wrong.

Asking nicely fixes none of that. What helps is knowing where exactly it breaks and checking the result, and that is what these files carry.

📦 Install

Five steps: a Figma seat, Claude Code, three plugins, one restart at the end.

Or hand it to the agent. If Claude Code is already installed, paste this into the chat:

Install https://github.com/schudarin/chudarin-design-skills: run the commands from install steps 3, 4 and 5 of its README

The agent runs those commands itself. Three things stay with you: the seat from step 1, the restart, and the Figma login in step 6.

1. A Figma seat that allows agent access

Figma limits an agent by seat, not by skill:

Seat	Starter	Professional	Organization	Enterprise
View, Collab	20 reads a month	6 reads a month	6 reads a month	6 reads a month
Dev, Full	20 reads a month	200 a day, 10 a minute	200 a day, 15 a minute	600 a day, 20 a minute

A View or Collab seat is enough to look at the thing once, not to work. You need Dev or Full.

The limits count reads. Writing into a file is exempt and free while Figma's beta lasts. Source and current numbers: Rate limits & access in Figma's developer docs, as of September 2026.

2. Claude Code

Claude Code is Anthropic's agent. It runs in Terminal, the black window, not a code editor.

Open Terminal (⌘ + Space, type Terminal, Enter), paste this and press Enter:

curl -fsSL https://claude.ai/install.sh | bash

When it finishes, type claude and press Enter. That starts the chat. /exit closes it. Full instructions: docs.claude.com/en/docs/claude-code/setup.

3. The connection to Figma

Figma publishes its own plugin for Claude Code. It carries the connection (the MCP server) and Figma's own agent instructions, which these skills build on top of:

claude plugin install figma@claude-plugins-official

4. These skills

claude plugin marketplace add schudarin/chudarin-design-skills
claude plugin install chudarin@chudarin

Both commands also work inside the chat as /plugin marketplace add … and /plugin install ….

5. superpowers

These skills decide how to work in Figma once the task is clear. They don't decide what the task is. Anthropic's superpowers does: on "add a button" its brainstorming fires first, asks which project and which Figma file, and only then hands over. Without it the agent starts on the wrong page.

claude plugin install superpowers@claude-plugins-official

6. Restart and log in to Figma

Quit Claude Code and start it again. Plugins load on start, so one restart covers all three. Then type /mcp in the chat, pick figma, and follow the login prompt: Figma opens in your browser and asks you to authorize the connection once.

Check the skills landed:

claude plugin details chudarin@chudarin

Every skill is in the list. Inside a chat they're chudarin:design-screens, chudarin:figma-plugin-api-rules, and so on. The agent picks them up by itself when the task matches, and you can name one directly.

📁 Where to run it

Make one folder per product and always start Claude Code from that folder:

mkdir -p ~/Design/my-product
cd ~/Design/my-product
claude

The folder matters. design-screens writes what it learned about your project (design system, style source, the Figma file) into .claude/design.md inside it, and every later session reads that instead of asking you again. Start from a different folder and you start from zero.

🚀 Your first task

1. In Figma, right-click the frame or page you want → Copy link to selection.
2. In Terminal: cd ~/Design/my-product, then claude.
3. Paste the link, add one sentence.

https://figma.com/design/…  — audit this page and tell me what the design system holds

https://figma.com/design/…  — plan the screens for password recovery

https://figma.com/design/…  — the padding on these cards is hardcoded, put it back on variables

Always paste a link. Without one the agent guesses which file and page you mean, and it guesses wrong.

🧭 Which skill when

• "What screens do we need for this feature?" → plan-user-flows
• "What's in this design system, where is the debt, document it" → audit-design-system
• "Design / change / rebuild this screen" → design-screens, which pulls in figma-plugin-api-rules by itself
• Anything that writes to Figma → figma-plugin-api-rules
• "Variables are detached / hardcoded / Figma says «Variable was deleted»" → figma-fix-variable-bindings
• "Rewrite this so it doesn't read as generated" → writing-texts
• "Make / restyle / fix a presentation" → design-slides

🔄 Update and remove

claude plugin update chudarin@chudarin      # newest version, restart to apply
claude plugin uninstall chudarin@chudarin   # remove
claude plugin list                          # what is installed right now

🩺 When it doesn't work

What you see	What it means	What to do
The agent says it can't reach Figma	The connection isn't authorized	Type /mcp, pick figma, log in through the browser
It stops after a handful of reads	Seat limit: 6 reads a month on a View or Collab seat, 20 on Starter	Ask for a Dev or Full seat on a paid plan
Reads start failing later in the day	The daily read limit: 200 on Professional, 600 on Organization and Enterprise, across all tools	Continue tomorrow, or split the work across days
The agent works in the wrong file or page	It had no link and guessed	Paste the link every time. And install superpowers from step 5
The skills don't show up	The plugin is installed but the session is the old one	Quit Claude Code and start it again
The agent changed something you didn't want	Figma keeps the file's history	File → Show version history, restore the earlier version

➕ Recommended alongside

Nothing in this section is required. With the plugins from steps 3 and 5 in place the skills work on their own. Everything below is someone else's skills, not mine.

design-screens also uses one visual skill per task: a mobile-UI, web, motion or charts skill, whichever matches the screen. It requires none of them. It reads the list of skills your session actually has, loads the one matching the task type, and if none is installed, says so and works from your design system, your existing screens and its own rules for text and controls instead. The names it looks for: mobile-app-ui-design, frontend-design, impeccable, design-motion-principles, emil-design-eng, dataviz, and design-critique for the review pass.

Installing them

frontend-design is Anthropic's own and comes as a plugin:

claude plugin install frontend-design@claude-plugins-official

dataviz ships inside Claude Code. Nothing to install.

The other five live on GitHub and install with the Skills CLI from skills.sh. It needs Node.js: type node -v in Terminal, and if it says command not found, install the LTS build from nodejs.org first. Then one command per skill, in Terminal, not in the chat:

npx skills add pbakaus/impeccable@impeccable -a claude-code -g -y
npx skills add kylezantos/design-motion-principles@design-motion-principles -a claude-code -g -y
npx skills add emilkowalski/skills@emil-design-eng -a claude-code -g -y
npx skills add ceorkm/mobile-app-ui-design@mobile-app-ui-design -a claude-code -g -y
npx skills add anthropics/knowledge-work-plugins@design-critique -a claude-code -g -y

The flags: -a claude-code picks the agent, -g installs for every project rather than the current folder, -y skips the confirmation. Drop -g to install into one project only. For Codex, replace -a claude-code with -a codex.

Quit Claude Code, start it again, then check what landed:

npx skills list

Later, npx skills check shows which of them have a newer version and npx skills update pulls them all.

Any other skill on skills.sh installs the same way, npx skills add owner/repo@skill-name, and npx skills find <word> searches the catalogue from Terminal. Read the licence before installing someone else's work. If a name above has since moved or gone, nothing breaks: design-screens checks which skills you have installed on every run rather than trusting this list, so an outdated name gets in the way of nothing.

Button labels and interface text go through writing-texts. If you already have a writing skill of your own, keep using it instead.

🤖 Other agents, and installing by hand

Every skill is a folder: SKILL.md + references/ + agents/openai.yaml for Codex. The Skills CLI from skills.sh installs them anywhere it knows; it needs Node.js.

npx skills add schudarin/chudarin-design-skills -a codex -g -y          # Codex
npx skills add schudarin/chudarin-design-skills -a claude-code -g -y    # Claude Code, no plugin system

-g installs for every project; drop it for one project. npx skills update pulls newer versions later. In Codex, call a skill explicitly with $skill-name; implicit invocation is enabled on all of them.

Without Node.js, clone and link the folders where your agent looks for skills, ~/.agents/skills/ for Codex and ~/.claude/skills/ for Claude Code:

git clone https://github.com/schudarin/chudarin-design-skills ~/chudarin-design-skills
mkdir -p ~/.agents/skills
for s in ~/chudarin-design-skills/skills/*; do ln -sfn "$s" ~/.agents/skills/; done

Re-run the loop after git pull to pick up updates.

On any other agent this is ordinary Markdown. Hand SKILL.md to the agent as a context file before Figma work.

Either way you still need the Figma MCP server, and Figma's own figma-use instructions loaded before the first use_figma call: the Figma plugin provides them in Claude Code, elsewhere the MCP resource skill://figma/figma-use/SKILL.md does.

🤝 Contributing

Each skill documents how to extend it; for figma-plugin-api-rules the protocol is at the end of its SKILL.md. Facts about one specific Figma file (node IDs, local conventions) don't belong here. Keep those in <project>/.claude/design.md or your own notes.

🎛️ Also by me

chudarin-context-mod: a side panel for Claude Code that shows what fills your context and what each action costs. A separate plugin in the same chudarin catalog.

📄 License

MIT, see LICENSE.