better-interface

Files (7)

• LICENSE
• SKILL.md
• openai.yaml · agents/
• change-review.md
• removed-signals.md
• review-format.md
• scope-resolution.md

SKILL.md (first lines)

# Interface review

This is the single interface-review entry point. It resolves a screen or change scope, routes the requested coverage to the owning `better-*` skills and consolidates one ranked verdict.

Orchestration is all it owns. Accessibility rules belong to `better-accessibility`, structure to `better-layout`, copy to `better-writing`, type to `better-typography`, color to `better-colors`, visual polish and motion to `better-ui`. Never duplicate or override their rules here.

Change-scope handling lives in [change-review.md](change-review.md). Load it only for branch, pull-request, commit-range, or uncommitted-change reviews.

## Evidence, not taste

Press hard on the escalation triggers and leave deliberate project choices alone. Those pull the same way. A trigger is a failure whatever the style guide says; a density, radius, or voice you merely disagree with is not a finding.

So the bar for reporting is evidence, not taste. The bar for `Approve` is that you inspected what you claim to have inspected. A short report from a real inspection beats a long one padded to look thorough.

## Core principles

### 1. Resolve the scope first

Infer the screen, flow, feature, or repository scope from the request and current workspace. State the resolved scope in the output.

Cover the resolved scope across the selected domains, including empty, loading, error and narrow-width states where they exist. Report at most 15 findings. A comprehensive review selects every domain below.

When the scope is too large to inspect credibly, narrow it to one complete flow: the one the request centers on, or failing that the entry path every user must pass through. State the boundary and what it excluded. Never imply uninspected surfaces were reviewed.

### 2. Resolve changes within this entry point

A request naming a branch, pull request, commit range, or uncommitted changes is a change review, not a screen review. A request to review the current work uses change mode even without a named target. Load [change-review.md](change-review.md) before inspecting the artifact. It owns reproducible scope resolution, removed-side inspection and finding classification.

Complete that mode without asking the user to invoke another skill. If it finds no change, follow its target-selection rule rather than inventing a last-commit review or silently widening into a repository audit.

In change mode, severity, ranking, the cap and the verdict cover `Introduced` and `Regression` only. A screen, flow or explicitly requested repository audit uses audit mode and needs no Git change classification.

### 3. Recon before judgment

Identify the framework, styling system, component library, design tokens, supported viewports and any preview or test command. Write every fix in the project's own idiom, so no finding arrives as a request to adopt a different stack. That governs the form of the fix, not whether the code is good enough.

Then read what the project has written about its own interface: `CONTRIBUTING.md`, `CODING_STANDARDS.md`, `AGENTS.md`, `CLAUDE.md`, a design-system doc, Storybook docs, interface ADRs. Name which you found, or that there are none.

Read them to find where a finding belongs, not for permission to drop it. A documented convention is no evidence the convention is good, and "it's in the style guide" does not retire a finding. What they change is **where** you report. When a guideline or shared token is the cause, report it once against that source, with the components as its locations.

### 4. Use domain skills as the sources of truth

For a general cross-discipline or explicitly comprehensive review, select every owner below. For a clearly scoped domain request, select the relevant owners and expand only when concrete evidence crosses another domain boundary. Confirm the selected owners are available, load them and complete their reviews before consolidation. Mark unselected domains `Not reviewed: outside requested coverage`.

Review in this order so foundational failures are not hidden by polish:

1. `better-accessibility`
2. `better-layout`
3. `better-writing`
4. `better-typography`
5. `better-colors`
6. `better-ui`

From a domain skill loaded here, take its principles, its references and its verification checks. Its severity ladder and its format are for standalone use; the consolidated format, shared severity and finding cap in this file replace them.

If an owning skill is unavailable, mark that domain `Not reviewed`, name it and continue with the rest. Do not recreate its rules from memory, substitute a neighbour, or claim holistic coverage.

When two skills appear to cover one issue, assign it to the owner of the underlying rule and note secondary effects in the **Why** cell. Report it once.

### 5. Require evidence

Every finding cites `path/to/file:line` and shows the current implementation. Do not report a code-level finding from visual appearance alone or a visual finding from source code alone when runtime behavior determines the result.

### 6. Rank by user impact

Use one shared severity scale:

- `HIGH`: blocks a task, misleads the user, hides content or controls, causes data-loss risk, or creates a repeated systemic failure.
- `MEDIUM`: meaningfully harms comprehension, efficiency, adaptability, or consistency.
- `LOW`: isolated polish with limited task impact.

Within a severity, rank by how many places the finding reaches and how much one fix buys. A token or shared-component fix outranks the same symptom in one leaf.

**Escalation triggers.** Once the owning skill confirms one of these, it is `HIGH` on sight, never averaged down because the surface is minor:

- An interactive control with no accessible name.
- A keyboard-reachable control with no visible focus indicator.
- A control or path reachable by pointer but not by keyboard.
- Motion or auto-playing content that ignores `prefers-reduced-motion`.
- Content or a control clipped, overlapped, or unreachable at 320px width or 200% zoom.
- Body or control text whose rendered contrast pair fails its required ratio.
- State or meaning carried by color alone.
- A destructive action with no confirmation, undo, or distinct treatment.
- Truncated content with no way to reach the full value.
- Content or a control reachable only past a scroll edge or behind a disclosure that has no visible cue.
- An error that names no way to recover from it.
- A semantic color used against its meaning, such as the danger hue on a non-destructive action.
- A state change carried by motion alone, with no color, icon, or label left behind when the animation does not run.

Triggers rank above every other finding. When more fire than the cap allows, list them first and say how many the cap excluded. A cap may shorten a report; it may never be why a blocker went unreported.

These set severity, not new rules. The owning skill decides whether the symptom is present; this list decides what it costs. In a change review, a confirmed `Regression` against a trigger is `HIGH` even where the same symptom would be `MEDIUM` as pre-existing.

### 7. Prefer the cheaper fix

Severity says how bad a finding is; this says which fix to propose. When more than one would work, take the earliest that does:

1. **Delete.** A separator that space would carry, an animation on a high-frequency interaction, an ARIA attribute a native element makes redundant, a ramp nothing imports.
2. **Use the platform.** The native element, the native control, the browser's own focus ring, in place of a custom rebuild.
3. **Reuse what the project has.** An existing token, spacing step, or motion curve, before any new value.
4. **Correct the value.** The wrong easing, radius, gap, or contrast pair, using the exact value the owning skill gives.
5. **Add.** A new token, a wrapper, a media query, an ARIA attribute the platform cannot supply.

A fix written at step 5 where step 1 was available is its own finding. Report the deletion instead.

### 8. Consolidate systemic findings

One root cause is one finding. List every confirmed location in the same row rather than one row per occurrence. Never pad to reach the cap; a short review or no findings is a valid result.

### 9. Verify what can be verified

Run the safe, relevant checks the project offers. Inspect the rendered interface when runtime behavior or visual judgment matters, and report the exact command or interaction and its result. A check you cannot run is **Not verified**, never a finding.

### 10. Review without mutating by default

Treat a review request as read-only. Do not edit source unless the user also asks you to implement the findings. When they do, keep the consolidated report as the change scope and re-run the relevant verification afterward.

## Before you finish

From cx18121/agent-skills, MIT.

More from cx18121/agent-skills