Skip to main content
skill CX Codex

ux · rainmanjam — Claude Code skill: structure your interface code

Tells the agent how to split a screen into components, pick a state owner and keep markup away from business rules.

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add rainmanjam/poka-yoke --skill ux -a claude-code
Open on GitHub

What ux does

This skill covers forms, flows and screens. Interface code grows faster than anything else in a project, since requirements keep arriving as “when the user does X, also…”. One component ends up responsible for fetching, validation, formatting, submission and error display at the same time, and any edit touches all five.

Inside are four sections: state ownership, component seams, non-success states, and code to avoid. Every value needs a single owner; two components that can write to it will drift apart, and the drift shows up only sometimes. Derived values get computed, not stored a second time. Server data and form input stay separate because they outlive each other differently. A thing that both loads and renders loses the ability to be previewed or fed other data.

Loading (including a refresh over content already on screen), empty, error with something the user can act on, partial data: the skill makes the agent account for each one. A discount rule belongs in a function you can test without rendering anything. Formatting stays at the display edge, so you don't parse your own output later. Colour and spacing are explicitly out of scope for a structural review.

Extracting a component with one caller counts as a cost, not a win. So does threading a value through layers that never read it.

What it helps with

  • split a form component into pieces with typed props
  • pick a single owner for form state
  • compute totals instead of storing a second copy
  • show empty and partial results without pretending they are errors
  • move discount logic out of JSX into a testable function

Example prompts for the agent

  • Break CheckoutForm into components and show me the prop types.
  • Where does state live in Cart, and what should be derived instead of stored?
  • Review OrderScreen and list the non-success states it never handles.

How to install ux

1Skills CLI (Vercel)

npx skills add rainmanjam/poka-yoke --skill ux -a claude-code

2Manually: put the files in place

~/.claude/skills/ux/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/ux/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about ux

How do I install the ux skill in Claude Code?

Run npx skills add rainmanjam/poka-yoke --skill ux -a claude-code. The skill file lands at ~/.claude/skills/ux/SKILL.md.

What does the ux skill do?

It gives the agent a structure for interface code: component boundaries, one owner per piece of state, data loading kept apart from rendering. It also pushes for explicit loading, empty, error and partial states.

Which editors can use this skill?

It installs into Claude, Codex, Cursor, GitHub Copilot, Gemini CLI and OpenCode, each with its own install command and file path.

Source text

SKILL.md · MIT
---
name: ux
description: >-
  Forms, flows and screens. Structure, naming and clarity of user-facing interaction code:
  component boundaries, state ownership, and keeping presentation separable from the rules
  behind it.
---

# UX: Structuring Interaction Code

Interface code accumulates faster than any other kind, because every requirement arrives as
"and also, when the user does X". Without structure a screen becomes one component holding
fetching, validation, formatting, submission and error display, and every change risks all of
them.

The discipline is ordinary: single responsibility, explicit state ownership, and names that
survive being read alone. What is different is that the pressure to skip it is higher, because
interface code is visible and the deadline is usually visual.

## Decide where state lives, once

Most interface bugs are state-ownership bugs wearing other clothes.

**One owner per piece of state.** If two components can both change a value, they will
disagree, and the disagreement will be intermittent. Lift the state to the nearest common
parent and pass it down, or move it out of the tree entirely.

**Derive rather than duplicate.** A `total` stored alongside the `items` it sums is two facts
that must agree. Compute it. If computing it is expensive, memoise the computation rather than
storing a second copy.

**Keep server state and form state distinct.** What the server last said and what the user has
typed are different things with different lifetimes. Conflating them is why "the form reset
itself" bugs are so hard to reproduce.

**Name state for what it is, not what it shows.** `isSubmitting` is a fact. `showSpinner` is a
consequence, and it will eventually need to be true for a second reason.

## Split components along seams that will move

**By responsibility, not by size.** A 400-line component that renders one coherent thing is
easier to work with than four components passing state between them.

**Separate what fetches from what displays.** A component that both loads and renders cannot be
previewed, tested, or reused with different data. Take data as props; let a parent decide where
it came from.

**Keep the rules out of the markup.** Whether a discount applies is a business question with a
right answer that does not depend on React. It should be a function you can test without
rendering.

**Push formatting to the edge.** Currency, dates and pluralisation are presentation. Keep the
underlying values unformatted until the moment they are shown, or you will parse your own
output later.

## Handle the states that are not "success"

Every screen that loads has at least four states, and most code writes one.

- **Loading**, including a second load over existing content, which is a different visual case
- **Empty**, which is not an error and should not look like one
- **Error**, with something the user can act on
- **Partial**, where some of the data arrived

If those live as branches inside one render, the component is doing five things. A small
discriminated union describing the state, matched once, keeps them separable and makes it
obvious when one is missing.

## What good output looks like

- **Show the component boundary and the prop types.** Most of the design is visible from the signature.
- **Say where each piece of state lives and why.** Ownership is the decision worth recording.
- **Cover the non-success states explicitly**, or say which are out of scope.
- **Point at specific components.** "The table" is not reviewable.

## What to avoid

**Prop drilling as a default.** Passing a value through four components that do not use it
couples all four. Either lift it out of the tree or restructure so the intermediate layers do
not exist.

**Premature component extraction.** Splitting a component that has one caller, to "keep files
small", adds a hop with no benefit. Wait for the second use.

**Effects that synchronise state.** An effect that copies one piece of state into another is
almost always a derivation that should have been computed.

**Styling decisions in a structural review.** Spacing and colour are real work and a different
conversation. Mixing them in means the structural comments get skimmed.

© rainmanjam · MIT · rainmanjam/poka-yoke

Similar skills

All skills for Frontend and UI →

ux by other authors comes first.

From the same set7

Project sponsors

Companies that keep this analytics open
Gold sponsor

Impressions, clicks and code copies — open statistics. We keep the numbers public for the same reason NeBlask itself is public. Become a sponsor