AI for UX · For leads · 7 min read
UX agents and skills: a guide for designers
What context files, skills, sub-agents, MCP connectors and plugins are, in plain words; when to use each; how to write a skill that works; five example skills for UX work; and how to review and share them safely.
A prompt helps once. A skill helps every time. Agentic tools let a designer hand whole tasks to AI: "build a brief from this ticket", "review this screen against our design system", "write up this interview". A skill captures how your best designer does that task, so everyone gets the same checks and the same output. It's also software that can read your files and call other systems, so it needs the care of any software you install. This guide is written for leads, and useful to any designer who wants to build one.
Products are named only as examples, as of November 2026. Check each vendor's documentation before you build.
What you'll be able to do after reading:
- Tell apart context files, skills, sub-agents, connectors and plugins
- Decide which one a task needs
- Write a skill with clear steps, output and rules
- Review a skill before you install it
- Share a skill with your team safely
1. The building blocks
Standing rules the agent reads in every session.
Like the brand guidelines on the wallInstructions for one kind of task, loaded when needed.
Like a written process for one jobA separate worker that reads a lot and returns a summary.
Like a researcher who comes back with one pageLets the agent read or act on another system.
Like a key to another buildingA bundle of all of the above, installed in one step.
Like an onboarding kit| Block | What it is | Use it for | Examples (as of Nov 2026) |
|---|---|---|---|
| Context file | A file the agent reads at the start of every session in a project | Stable rules: users, principles, design system rules, "never do" | CLAUDE.md, AGENTS.md, DESIGN.md. See Context files |
| Skill | A folder of instructions (plus optional references and scripts) the agent loads for one kind of task; Agent Skills is an open format for them, published by Anthropic in December 2025 | Interactive work: needs Figma or a browser, asks follow-up questions, applies fixes | Claude Code skills, Anthropic's frontend-design skill |
| Sub-agent | A separate AI worker that runs in its own context and returns a short report | Heavy reading: a 60-minute transcript, many wiki pages; can run in parallel | Claude Code sub-agents |
| Connector (MCP) | A server that lets the agent read from or act on another system (MCP) | Design files, issue trackers, wikis, browsers, analytics | Figma MCP server, Playwright MCP |
| Plugin | A bundle of skills, sub-agents, hooks and connector settings, installed in one step | Sharing a set of tools across a team, with versions | Claude Code plugins |
A useful pattern combines them: the skill talks with you, and hands the heavy reading to a sub-agent, which returns only what you need. (Some tools can also run a whole skill as a sub-agent.)
2. Skill or agent?
- 1Needs you in the loop?
Choices, approvals, fixes in Figma.
→ Skill - 2Reads a lot, returns a summary?
Transcripts, many wiki pages.
→ Sub-agent, called from a skill - 3Runs on a schedule?
Only adds what is new.
→ Scheduled agent - 4More than two or three users?
Others will install it.
→ Package it as a plugin
3. Anatomy of a skill
Most skill systems use a folder with one main instruction file and optional supporting files:
design-brief/
├── SKILL.md # name, description, and the instructions
├── references/
│ ├── brief-template.md
│ └── sources-guide.md # where to look, in which order
├── examples/
│ └── example-brief.md # a good, anonymised output
└── scripts/ # optional: helpers the skill runs
SKILL.md, a generic example
---
name: design-brief
description: Turn a ticket into a design brief for a UX designer. Use when the
user gives a ticket key or URL and asks for a brief, context, or "what do I
need to know" before designing.
---
# Design brief from a ticket
## Goal
Collect what is already known about a demand and point out what is missing,
before the designer opens Figma.
## Steps
1. Read the ticket, then its parent epic and relevant sibling tasks.
2. Read the wiki pages linked from them; search by feature name for older pages.
3. Search for related bugs and support requests; turn them into edge cases.
4. Check for existing designs and screenshots of the current build.
5. Map the demand to known design system patterns (see references/).
## Output
Use references/brief-template.md: goal, users, requirements, data, states and
edge cases, pattern, conflicts and gaps (always), up to 8 questions for the PM.
## Rules
- Read-only. Never post or edit anything unless the user explicitly asks.
- Cite a source (ticket, page, design node) for every claim; mark inferences.
- If information is missing, say so. Do not invent requirements.
What makes a skill work well
- A precise description. It's how the agent decides when to load the skill: say what it does and when to use it, in the words people actually type.
- Steps, not wishes. "Read the epic, then the siblings" beats "be thorough".
- A fixed output format in a template, plus one good example.
- Explicit rules about what it must never do (post, edit, publish, invent).
- References for knowledge that changes (design system rules, decisions), kept apart from the instructions so they're easy to update.
4. Five example skills for UX work
Generic patterns: adapt the inputs and sources to your own tools.
| Skill | You give it | You get | Rules that matter |
|---|---|---|---|
| Design brief builder | A ticket | A one-to-two-screen brief with requirements, states, conflicts and gaps, and questions for the product manager | Read-only; every claim cites a source |
| Screen reviewer | A Figma frame | Errors, warnings, layout improvements and what's done well, each with where, which rule and a fix | Edits only after an explicit yes |
| Interview synthesiser (skill + sub-agent) | A recording and transcript | An anonymised draft research page: behaviours, quotes with timestamps, confirms / contradicts / new | Names and recordings stay local; publishes only on "publish" |
| Interviewer coach | A transcript of an interview you ran | A private report on your questions: leading, closed, missed follow-ups, with rewrites | Private by default |
| Handoff spec writer | A finished frame | Components, states, focus order and validation messages for developers | Drafts for a designer to approve |
See AI in UX research and AI-assisted synthesis, step by step for the research rules.
5. Rules for every tool
- Read-only by default. Nothing is posted to the issue tracker, wiki or design files until a designer has read and approved it.
- Every claim cites its source: a ticket, a page title, a design node, a timestamp. Inferences are marked as such.
- Gaps are part of the output. Conflicts and open questions matter as much as the summary.
- Participants stay anonymous in research tools.
- Approved access only. Your organisation's approved AI setup, never a personal account for work data. See AI governance and data protection.
6. Review a skill before you install it
Treat every skill, plugin and connector as code from a stranger until you've checked it. In February 2026, researchers at Koi Security found 341 malicious skills among 2,857 in one public skill marketplace, most of them installing password-stealing malware (eSecurity Planet). See also AI tools landscape for designers.
7. Share a skill with your team
- Package it as a plugin in your organisation's internal repository, not a zip file in chat.
- Write a README: purpose, inputs, outputs, setup, known limits, owner.
- Include an anonymised example output.
- Version it and keep a short changelog.
- Calibrate before rollout: run it on two or three past pieces of human work and compare.
- Name an owner (often an AI champion).
- Retire it when it stops earning its keep. Review the set every quarter.
8. Vocabulary
| Term | Meaning |
|---|---|
| Context file | Standing instructions an agent reads in every session |
| Skill | Packaged instructions for one kind of task |
| Sub-agent | A separate AI worker with its own context, returning a summary |
| MCP | The protocol that connects agents to other systems |
| Plugin | A bundle of skills, agents and settings, installed in one step |
| Sandbox | A safe test space where mistakes don't matter |
9. Try it (no code needed)
- Pick one repetitive task you did this month.
- Write its
SKILL.mdon paper or in a doc: description, steps, output, rules. - Ask a colleague to follow your steps by hand, without you. Where did they get stuck? That's where an AI would too.
- Pick one public skill and review it with the checklist in section 6.
Reflection questions: Which of your team's tasks has a "best way" that lives only in one person's head? What should a skill for it never be allowed to do?
Checklist
Learn more
- Docs: Agent Skills (open format) · Claude Code skills · Claude Code plugins · AGENTS.md · Model Context Protocol
- Examples: Anthropic skills repository
Related
Building an AI agent roadmap for your UX team · Generating design with Claude Code · Agentic design · AI in UX research · AI governance and data protection · Context files · MCP · Module 12: AI and the UX Lead