UX · CX · BX
Working repo, not a sales page
Use Claude properly
A public working repo of the commands, tools, workflows, and sources worth actually using with Claude. Curated by someone who uses it every day and checked against the official docs. No newsletter trap, no guru funnel, no recycled prompt advice.
Designing with Claude
Anthropic now ships an official Design plugin, and it changes what design work looks like in Claude Code. The skills below cover the design process end to end, from research through critique to developer handoff, then connect Claude straight to your Figma files. Pair it all with a DESIGN.md so the output builds to your design language rather than a generic one.
The plugin installs from Anthropic's official marketplace: see the Design plugin page Official. Each skill below is a focused job you invoke by name. The one red dot is a skill named in community coverage that the official page does not list explicitly, so confirm it in your own build.
Two bundled commands connect Claude Code to your actual design system. /design-login authorises access, then /design-sync converts your components so Claude builds with them rather than around them. A first sync verifies every component and can take a few hours on a large repo. Both are in the commands reference Official.
A DESIGN.md is a design system written for an AI agent to read. styles.refero.design is a searchable library of over 2,000 of them, with colours, typography, spacing and rules extracted from leading product sites, and a Refero MCP so your agent can query the whole collection. Community Or write your own, so first pass output lands on brand.
shadcn/ui is the open code component system agents build with best, and one of the most starred UI projects on GitHub. The components live in your repo rather than behind a package, so Claude can read and adapt them directly. Its official MCP server lets Claude browse, search and install components from any registry by prompt Official. A community server adds Svelte, Vue and React Native variants. Community
The official Frontend Design plugin generates production grade frontends that avoid generic AI aesthetics Official. Claude Design turns a description into a landing page, prototype or deck on a canvas Official. For a designer's walkthrough of the plugin, read the UX Planet piece Community. And for the research grounded UX principles these skills lean on, Nielsen Norman Group remains the reference.
The quickest route is the official Figma plugin for Claude Code, which bundles the MCP server settings and skills for the common workflows. Install it, share a Figma URL, and ask Claude to implement the design. It extracts layout, typography, colour and tokens, maps components to your codebase through Code Connect, and works with both the cloud server and the desktop app.
Prefer no MCP at all? The figma-cli route in Enhance Claude drives Figma Desktop through DevTools with no API key.
How Claude performs
An honest ranking of where Claude leads and where it falls behind, across the disciplines designers actually use. Many teams are standardising on Claude alone because consistency beats cherry picking: one model, one voice, one set of habits, and output that lands on brand every time. Others run Claude only for the disciplines it wins, coding, design to code, writing and security, and reach elsewhere for image and video generation, which Claude does not do. Both are sensible strategies, and the chart below shows why.
Positions are our reading of the public evidence, kept honest: none means the capability does not exist in the model, not that it scored badly.
Resources
New to Claude? Start here. These cards start with what Claude can actually do, then cover how to extend it with skills and connectors, how to spend fewer tokens, and how to give it enough context that the output sounds like you rather than generic AI. A short, strictly vetted reading list sits at the foot of the page. Tap any card to open it.
These are real capabilities, each linked to its official documentation. The opposite of a hyped feature list: nothing here needs a thread or a paid course to unlock, and several are the setup behind The Stack and the command reference further down.
A skill is a folder containing a SKILL.md file that Claude loads only when a task calls for it. It teaches Claude a repeatable job (building slides, trimming tokens, following a house design style) without taking up space in every conversation. You add one by pointing Claude at the folder. See the official skills documentation Official.
Find them in a public skills directory Community or on GitHub Community. Read the SKILL.md first and check the source before installing, because a skill runs with your permissions. A few worth looking up to start: Agent BrowserSuperpowersFrontend Design, officialDesign TasteCavemanRemotionGSAPMarketing SkillsSocial Media Skills
MCP servers and connectors link Claude to Drive, Gmail, Calendar, GitHub and Slack. See the official MCP setup guide Official.
Drive Figma Desktop through its built in DevTools with a community CLI, no API key. Read: Into Design Systems. Tool: silships/figma-cli. It is third party and work in progress, so vet it before running.
A library of DESIGN.md specs you can hand Claude: styles.refero.design. Or write your own so the output builds to your system, not a generic one. More on this in Designing with Claude at the top of the page.
A CLAUDE.md in your repo gives Claude your patterns and rules, so first pass output is on brand instead of generic.
Terminal only. The slash commands on this card, such as /compact, /cost and /usage, are Claude Code commands you run in the terminal. They do not work in the Claude web or desktop chat.
Use /compact to compress a long session, /context to see what is filling the window, and /clear to reset. Full list in the Claude Code commands reference. Official
The Caveman skill has Claude reply in a terse, compressed style that cuts token use by about 75 percent while keeping technical accuracy. It is community made (MIT licensed, actively maintained), so read the source before installing. Community
Haiku for simple and fast, Sonnet for the everyday default, Opus for hard reasoning, and Fable for the ceiling. The best alias picks the newest Fable available to you. Using the smallest model that works is the biggest saving. Anthropic's guide to choosing a model and effort level breaks down the trade offs. Official
Each revision optimises from the model's own previous output and converges on the generic, so you burn tokens getting further from distinctive, not closer. Start fresh for original work. See the read on AI design convergence. Community
/usage checks usage against plan limits and shows session cost; /cost and /stats are aliases of it. See the cost tracking docs. Official
/batch runs changes across worktrees and subagents (/agents) split work, which avoids re running the same context.
Each of these is a copy and paste prompt built on a well known way of working through a problem. Used properly, Claude asks questions and challenges your reasoning instead of agreeing. Each is a small markdown file you can download and reuse.
Checked Reviewed September 2026. Every tool below was confirmed live; pricing tags reflect what each site states today and can drift, so check before you commit.
Each tool connects to Claude in one of five ways: a direct API call, an MCP connector, a plugin, a skill, or an automation platform. You do not need any of these to begin. They are grouped below by how they connect. Each name links out to the tool, and the tag shows how it is priced.
Recipes
Things worth doing with Claude, written the way LinkedIn never writes them: what it actually does, why it is useful, whether it is real, and where it comes from. Each recipe teaches one Claude idea that usually gets reduced to a buzzword. Not workflows; recipes. Ingredients, steps, outcome. Tap Try it on any card to copy a prompt that runs the recipe in your own Claude.
Command reference
Claude Code is the version of Claude that runs in your terminal and can read and edit your files. The commands below are its slash commands, the typed shortcuts like /init or /compact that control a session.
They are grouped by the job they do. Expand any group to see the commands inside it.
Each group also links out to the official documentation for those commands, under Learn it. Where the page points to anything beyond Anthropic, a tag rates the source.
Notes
Where this command list comes from, the corrections made to it, and what the confirmed and unverified marks mean.
Every command, flag and shortcut on this page was re-checked against the official Claude Code docs in September 2026. Nearly everything previously marked unverified is now documented, so the red dots you see are genuinely rare. Anything removed from the product moved to the Deprecated section at the foot of the page, with its replacement named.
- ultracode is no longer an effort alias. It is a setting and prompt keyword that has Claude plan a workflow for the task.
- /ultraplan was removed from the product. Plan mode replaces it.
- /ultrareview lives on as an alias of
/code-review ultra. - Model references use the current line: Opus 5, Sonnet 5, Haiku 4.5, and Fable 5.1.
Cross checked against Anthropic's official Claude Code docs at code.claude.com. A red dot means the item is plausible but was not independently confirmed, so availability varies by version, plan and platform. Type / in your own build to confirm.
Trusted sources
A short, deliberately strict reading list. Everything here is free, kept current, and comes either from the people who build Claude or from independent experts worth the time. It is the opposite of the recycled threads, engagement bait, and low trust prompt advice that fills most AI feeds.
Deprecated
Commands and claims that were accurate when first published and are not any more. Nothing is deleted, because knowing what changed is part of the record. Each entry names what replaced it. These entries stay searchable and appear beneath live results.
Bleeding edgeNew release
The newest things happening around Claude, gathered before anyone can sensibly evaluate them. Nothing here carries a confirmed dot, because that is the point: this is the front of the wave. As each topic settles and proves itself it graduates into the main sections above, the way the /design plugin will once it has been out longer than a week. Expect churn.
Claude Code is the reference harness, and the numbers behind that are startling: it now drives a meaningful share of public GitHub commits. The insight spreading through 2026 is that harness quality matters as much as model quality, which is why the same model behaves so differently in different tools. Anthropic's own explanation of the loop is the best grounding: how Claude Code works Official.
Then came the challenger. In August 2026 DeepSeek released DeepSeek Harness, an MIT licensed, model agnostic harness where everything is a plugin: models, tools, skills, sandboxes, even the loop itself. It passed 141,000 GitHub stars in four days and runs Claude, GPT, or a local model through Ollama. Community Whether it matures is exactly the question this section exists to watch.
Two releases changed the economics this year. Qwen3.8-27B is Apache licensed, downloads at 18GB, runs on a single consumer GPU through Ollama, sees images, and scores well enough on agentic coding benchmarks to drive a harness by itself. DeepSeek V4 ships open weights under MIT and holds its own on reasoning. Community
The emerging pattern is a mixed bench: Claude as the orchestrator and reviewer, where judgement and consistency matter, with local models taking the bulk work that would burn tokens, first drafts, batch transforms, test generation. The wiring is either a gateway that routes some requests to local endpoints, see the official LLM gateway docs Official, or a second harness such as DeepSeek Harness running the local model, with Claude checking its output the way The Stack already uses Codex CLI as a second eye. Cost of the local half: your electricity.