What is the better-interface skill?
The better-interface skill is the orchestrator of Jakub Krehel's better skills. It routes your interface through accessibility, layout, writing, typography, color and UI polish, collects their evidence and produces one ranked verdict. It owns no domain rules of its own. Each finding cites a file and line and ends in Block or Approve.
Better-interface is a coordinator. It does not carry rules for accessibility, layout or color. It loads the skill that owns each domain, runs those reviews and merges the results into one table ranked by user impact.
The bar for reporting is evidence, and the bar for Approve is that you inspected what you claim. A density, radius or voice you merely disagree with is not a finding.
How a review runs
- Resolve the scope: a screen, a flow or the repository. State it, and if it is too large, narrow it to one complete flow.
- Recon the framework, styling system, tokens and any project docs such as CONTRIBUTING.md or AGENTS.md.
- Review in order: accessibility, layout, writing, typography, colors and UI polish.
- Consolidate one root cause into one row, with every location listed, capped at 15 findings.
- Run safe checks the project offers and mark anything you could not run as Not verified.
Escalation triggers
Once the owning skill confirms them, some symptoms are HIGH on sight, such as a control with no accessible name, a missing focus indicator, a pointer only path, clipping at 320px or 200 percent zoom, and meaning carried by color alone.
Who it is for
Best for
- A whole screen or flow review before launch.
- One report instead of six separate domain reports.
- Teams that want fixes in their own styling idiom, preferring deletion or the platform over adding code.
Not for
- Reviewing a branch, pull request or uncommitted change. Run interface-review, which is user-invoked.
- Setups without the sibling better skills. A missing domain is marked Not reviewed and never recreated from memory.
- A design critique with an opinionated style. Related: critique and polish.
Use it with UiChemy
Review the page your agent built, fix the HIGH items, then move it into WordPress with UiChemy AI to WordPress or Figma to WordPress. Get Started on uichemy.com. Related: fixing-accessibility.
How to install Better Interface Skill: One Review Across Six Domains
git clone https://github.com/jakubkrehel/skills.gitClone the repository
Run
git clone https://github.com/jakubkrehel/skills.git. The skill is in skills/better-interface and includes review-format.md.Install its siblings too
Better-interface needs better-accessibility, better-layout, better-writing, better-typography, better-colors and better-ui. Copy those folders from skills/ alongside it, or a domain is marked Not reviewed.
Plugin route for Claude Code
Run
/plugin marketplace add jakubkrehel/skillsand/plugin install interfaces@interfaces, the commands in the README.Request a review
Name a screen or flow and ask for an interface review. Ask for fixes separately if you want edits.
Check that it loaded
Ask for a review of one screen. The report should open with the resolved scope and a coverage table, then one findings table across domains, then verification and a Block or Approve verdict.
Key features
- Scope firstThe skill infers the screen, flow or repository from your request, states the scope in the output and covers empty, loading, error and narrow width states where they exist.
- Change reviews routed outA request naming a branch, pull request or uncommitted changes is sent to interface-review, because a guessed diff gives the report a scope nobody can check.
- Recon before judgmentIt identifies the framework, styling system, component library and tokens, and reads project docs such as CONTRIBUTING.md and AGENTS.md before judging.
- Six owning skills in orderAccessibility, layout, writing, typography, colors and UI polish run in that order so foundational failures are not hidden by polish.
- Evidence on every findingEach finding cites a path and line and shows the current implementation. Runtime claims need rendered evidence or are marked Not verified.
- Shared severity scaleHIGH blocks a task or hides content, MEDIUM harms comprehension or consistency and LOW is isolated polish, ranked by how many places a fix reaches.
- Thirteen escalation triggersItems like a missing accessible name, ignored reduced motion, a destructive action with no confirmation and an error with no recovery path are HIGH on sight.
- Cheaper fix firstProposed fixes follow an order: delete, use the platform, reuse project tokens, correct the value, and only then add something new.
- Consolidated, capped outputOne root cause becomes one row listing every location. A report holds at most 15 findings, and a short or empty report is valid.
- Read-only by defaultA review request does not edit source. The skill implements findings only when you also ask for that.
Frequently asked questions
What is the better-interface skill?
It is the combining skill in Jakub Krehel's skills repository. It runs a cross-discipline review across accessibility, layout, writing, typography, color and UI polish by loading each owning skill, collecting evidence and consolidating one ranked verdict with file and line citations.
How do I install better-interface?
Clone jakubkrehel/skills and copy skills/better-interface plus the six domain skills it routes to into your agent's skills folder. In Claude Code you can run the plugin marketplace add and plugin install commands from the README. Then ask for a review of a named screen or flow.
Does better-interface contain the actual rules?
No. It owns orchestration only. Accessibility rules belong to better-accessibility, structure to better-layout, copy to better-writing, type to better-typography, color to better-colors and motion and surfaces to better-ui. If a domain skill is unavailable, that domain is marked Not reviewed.
How do I review a pull request with it?
You do not run it directly. Better-interface asks you to use interface-review for a branch, pull request or uncommitted change, because that skill is user-invoked and resolves the change scope. Interface-review then hands the review back for severity, coverage and the verdict.
Which AI coding agents can use better-interface?
The README documents a Claude Code plugin install, and the skill folder includes an agents/openai.yaml file. For other agents, copy the folders to their skills location. The README lists no other agents, so confirm yours loads SKILL.md folders and can read the sibling skills.
How many findings can a review report?
At most 15, and it never pads toward the cap. Escalation triggers rank above everything else, so when more fire than the cap allows they are listed first and the report says how many the cap excluded. A short review or no findings is a valid result.
Does it edit my code?
Not by default. A review request is treated as read only. It edits source only when you also ask it to implement the findings, and then it re-runs the relevant verification afterward. Fixes are written in your project's own idiom instead of a different stack.