Generated banner for Interface Review
Generated logo for Interface Review

Interface Review

Jakub KrehelCommunityCreated by Jakub Krehel

  • Skill
  • Review and audit
  • Software development
  • MIT
  • No issues found

From the author

Reviews a branch, pull request or uncommitted change for the interface problems it introduced or regressed, across accessibility, layout, writing, typography, color and UI.
— the description by Jakub Krehel

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

Interface-review provides a structured approach to auditing UI changes by focusing on the impact of a specific branch, pull request, or uncommitted set of changes. Rather than auditing an entire codebase, it resolves the scope of the change, identifies affected surfaces, and classifies findings as introduced, regressed, or pre-existing. It emphasizes reviewing the 'blast radius' of changes—such as shared components or design tokens—and ensures that the interface remains consistent with the stated intent of the pull request or commit.

What it does

  • Resolves the scope of changes from branches, PRs, or uncommitted work.
  • Expands the review to affected surfaces and consumers of changed components.
  • Classifies findings based on whether they were introduced, regressed, or pre-existing.
  • Checks for incomplete changes, such as missing states for new components or strings.
  • Routes specific domain issues to specialized skills while maintaining a unified report.
  • Operates in a read-only manner without mutating the user's working tree.

When to use it

  • When you need to verify that a new pull request does not break existing UI patterns.
  • When you want to ensure that a change to a shared component does not negatively impact its consumers.
  • When you need to audit uncommitted changes for accessibility or layout regressions before pushing.
  • When you need to confirm that a new feature implementation includes all required states like error or loading.

Advantages

  • Focuses strictly on the change, preventing scope creep and irrelevant feedback.
  • Systematically identifies regressions by comparing removed lines against known signals.
  • Provides a clear, standardized report format for interface findings.
  • Respects the integrity of the local environment by avoiding destructive git operations.

Usage examples

“interface-review pr 482”
The agent resolves the scope of pull request 482, identifies affected surfaces, and generates a report detailing any introduced or regressed interface issues.
“interface-review”
The agent detects the current branch and uncommitted changes, performs a scope resolution, and provides a summary of interface findings for the current work-in-progress.

Install

  1. Download the skill as a ZIP:

    https://autogineer.com/agent-skills/jakubkrehel/interface-review/download
  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip interface-review.zip -d .agents/skills

Documentation by Jakub Krehel

Interface review

This skill reviews a change rather than a screen. It resolves the scope, expands the changed files to the surfaces they affect, reads both sides of the diff and classifies every finding.

It owns the scope, the classification and the change-scoped report. Domain rules belong to the better-* skills. Severity, consolidation, coverage, the cap and the verdict belong to better-interface, which this skill hands the review to.

Correctness, tests, security and performance belong to the project's general code review. Name such a concern once, point at that review and drop it.

The change, not the codebase

The author is asking "did I make this worse?". Report what the change caused and stay mostly quiet about what it merely touched. Three pre-existing findings is a courtesy; thirty is a different review and one nobody asked for.

Read the whole diff before forming an opinion of it. A skimmed diff produces findings about code the next hunk already fixed.

1. Resolve the change scope first

The whole invocation is the target, so /interface-review pr 482 reviews pull request 482. Scope resolution holds the accepted targets and how each resolves.

With no target supplied, resolve in this order and stop at the first match:

  1. HEAD is ahead of git merge-base origin/<default-branch> HEAD. Review that range plus any uncommitted changes, stating the commit count and the uncommitted file count separately.

  2. The working tree is dirty. Review the uncommitted changes.

  3. Neither. There is no change to review, so follow With no change, ask rather than invent one.

Record the base commit of the resolved range as $BASE and name it in the scope block. Every command in this skill and its references uses it.

Exclude lockfiles, snapshots, generated output, vendored code and binaries per Excluded paths, and name what you excluded. If nothing survives the exclusions, treat it as no change.

2. With no change, ask rather than invent one

A clean tree with nothing ahead of the merge base means there is no change to review. Never fall back to HEAD~1..HEAD on your own. The last commit is whatever happened to land, often a merge or someone else's work.

Gather the facts in Nothing to review, state them, then offer these routes and wait:

  • The open pull request on the current branch, first when there is one. A branch whose commits already landed resolves to no change, while its pull request is still exactly what the user meant.

  • The last commit, HEAD~1..HEAD, named by short SHA and subject so the user sees what they would get.

  • A target they name, such as pr <n>, a branch, a ref or a range.

  • A whole-repository interface audit, which is not a change review. Hand it to better-interface as a repository-scope review, without this skill's scope block, statuses or pre-existing section.

Where exclusions emptied the scope, name the excluded files in the offer. Never report a review of nothing as Approve.

3. A diff is not a surface

A changed file is evidence, not the review subject. Its blast radius is the set of surfaces it renders in; review those.

Expand one hop by default, to the direct importers and callers. Expand a second hop only for design tokens, theme values and shared primitives, where one line reaches the whole product.

Review at most five consumers in total across both hops, ordered by the rule in Scope resolution. State how many you did not expand, since an unstated cutoff reads as completeness.

4. Read the removed lines

Regressions are invisible in the post-change state. Read the - side of every hunk against Removed signals.

A signal is a lead, not a finding. A removal is a regression only when nothing in the change replaces it, and the domain skill owns that judgement. Route each unmatched removal to its owner, report only what that skill confirms and status it Regression.

5. Classify every finding

Give every finding one status:

  • Introduced: the change created it.

  • Regression: the change weakened something previously correct.

  • Pre-existing: present before the change and not made worse by it.

Status by cause, not by location. A finding a changed line causes is Introduced or Regression wherever it surfaces, including an untouched consumer of a changed token. A finding the change neither created nor worsened is Pre-existing, even three lines from a hunk.

To check whether a line predates the change, blame the reviewed range. A ^ prefix marks a line unchanged since the base:

git blame -L <line>,<line> "$BASE"..<head-ref> -- path/to/file

For uncommitted work, drop the range. A line marked Not Committed Yet or blamed on a commit inside $BASE..HEAD belongs to the change.

6. Hold the change to its stated intent

Read the pull request title and body, the linked issue and the commit messages, then review whether the interface delivers what they claim.

This is how you find the incomplete change, which a surface review misses because it inspects only the states that exist. Look for the absent ones:

  • A new variant, size or theme applied to some states but not all of hover, focus, active, disabled, loading and selected.

  • A new user-facing string with no entry in the translation catalogue the project maintains.

  • A new component with no empty, loading, error, disabled or narrow-width state.

  • A control added to one surface but not to the siblings that already carry its peers.

Do not report scope creep. Whether a change does too much is a process question, not an interface one.

7. Hand the review to better-interface

Hand better-interface the scope block, the affected surfaces and a status on every finding, and report in the format below.

If better-interface is unavailable, report the resolved scope and the file inventory, name it as the missing skill and stop. Do not invent a severity scale, a cap or a verdict.

Never mutate the working tree

A change review is read-only, including the checkout. git fetch writes only to .git and is permitted. gh pr checkout, git checkout, git switch and git stash rewrite the files the author has open, so they are never permitted. Fetch pull request refs and read them in place.

Rendered verification is opt-in here, in place of better-interface's Verify what can be verified. Mark visual and runtime claims Not verified unless the project exposes a cheap preview or the user asks for a rendered review. For working and branch, the checkout already is the head, so render it in place. For any other target, render an isolated worktree such as git worktree add /tmp/review-<n> refs/remotes/pr/<n> and remove it when done.

Before you finish

MistakeFix
The scope block counts only uncommitted files while HEAD is ahead of the merge baseReview the branch range plus the uncommitted changes, with both counts
A report on HEAD~1..HEAD the user never choseState the facts and offer the routes in With no change, ask rather than invent one
A changed shared component with no consumer under Surfaces expandedExpand to its consumers and name the ones you skipped
A deleted aria-label, outline or prefers-reduced-motion with no finding and no noteRoute it to its owner through Removed signals
A Regression whose removed attribute reappears elsewhere in the diffCheck the equivalent replacements first
A finding on an untouched consumer of a changed token statused Pre-existingThe change caused it, so it is Introduced or Regression
A line outside every hunk statused Introduced with no causing change namedBlame it against the reviewed range
gh pr checkout or git switch in the command logFetch the ref and read files with git show
A cited line that does not match git show <head-ref>:pathCite against the head ref named in the scope block

Review output format

Open with the scope block:

FieldValue
Targetbranch, working, staged, pr 482 or the range as entered
Base reforigin/main at a1b2c3d
Head refrefs/remotes/pr/482 at e4f5g6h
Commits7 committed, 2 files uncommitted
Files in scope12 after exclusions
Excludedpnpm-lock.yaml, src/__snapshots__/, lockfile and snapshots
Surfaces expandedCheckoutPage, SettingsPanel; 3 further Button consumers not expanded

The coverage table from better-interface follows it. A domain with no evidence in the change scope is Not reviewed: no evidence in the change scope, which is a coverage statement rather than a gap.

Then the findings, with a Status column per Classify every finding:

SeverityDomainStatusLocationBeforeAfterWhy
HIGHAccessibilityRegressionsrc/Dialog.tsx:42aria-label="Close" removed in this changeRestore aria-label="Close" on the icon-only controlThe close control had an accessible name before this change and no longer does

With no Introduced or Regression findings, omit the table and state "No actionable interface findings in this change."

Then Pre-existing findings, at most three, highest severity first, stated plainly as not this change's responsibility. Omit the section when there are none.

SeverityDomainLocationIssue
MEDIUMTypographysrc/Toolbar.tsx:7Numeric badges use proportional figures; predates this change

Verification and the verdict follow in better-interface's format.

Files

scope-resolution.md

Scope resolution

How a review target becomes a file list, and the traps that fail quietly and leave the scope block claiming a count it never delivered.

Default branch

Try refs/remotes/origin/HEAD, then gh repo view --json defaultBranchRef. If the ref is missing, ask the remote with git remote set-head origin --auto rather than guessing. It needs the network and writes only a ref under .git, so it is permitted; note it in Verification. With no remote at all, fall back to a local main or master and state which base you assumed.

Targets

Accepted targets are working, staged, branch, pr <n>, a bare <ref> and an explicit <a>..<b> or <a>...<b> range. Anything else in the invocation is a <ref>.

Diff a branch against the merge base, with three dots. Two dots reports every upstream commit that landed on the base branch as part of the change.

But use the dots the user wrote when they wrote a range. <a>..<b> compares the endpoints; <a>...<b> compares merge-base(<a>, <b>) with <b>. Rewriting release..feature to three dots drops everything between release and the merge base, which is often exactly what was asked for. State the resolved range in the scope block.

git diff HEAD reports tracked changes only. Any target including uncommitted work must pair it with git ls-files --others --exclude-standard. Otherwise a newly added component is silently dropped from a scope the report claims to cover in full.

Pull requests

Fetch the head into a remote-tracking ref, git fetch origin "pull/<n>/head:refs/remotes/pr/<n>", and review it in place. This works for forks, which origin/<branch> does not.

Diff against the pull request's own base, never the default branch by assumption. Read it from gh pr view <n> --json baseRefName,baseRefOid and fetch it if it is not local. A pull request into release/2.0 diffed against main picks up every commit between the two.

Read files at the head ref with git show refs/remotes/pr/<n>:path/to/file. Never open the working-tree copy, which is a different revision.

gh pr diff <n> is a fine shortcut for the patch text. It gives no way to read unchanged context or expand to consumers, so fetch the ref as well.

Citations. Line numbers from a fetched ref need not match the working tree. Cite against the head ref and declare that ref and its SHA in the scope block so the numbers resolve.

Intent. The title and body from gh pr view are the stated intent for Hold the change to its stated intent. Add the commit subjects when the body is empty.

Awkward repository states

Three worth handling. Everything else fails loudly at merge-base: unrelated histories, a repo with no commits, a moved submodule pointer. Say the base is unresolvable and stop. Never review a range you cannot name.

Detached HEAD. Use the merge base against the default branch and name the SHA, not a branch, in the scope block.

Shallow clone, the CI default, where merge-base returns nothing. Fetch --deepen=50, retry, then --deepen=200, then report the scope as unresolvable. Deepening writes to .git and not to the working tree, so it is permitted; note it in Verification.

Mid-rebase or mid-merge, the one that does not fail loudly. git diff succeeds and returns something that is not the change, so the review looks fine and is wrong. Detect it with git rev-parse --git-path against rebase-merge, rebase-apply, MERGE_HEAD and CHERRY_PICK_HEAD. Do not test .git/ paths directly, because they are not directories inside a linked worktree. Stop and say the tree is mid-operation.

Nothing to review

The tree is clean and HEAD is not ahead of the merge base. Gather the facts before asking, so the offer is accurate rather than a guess. They are the current branch, the count ahead of the base, the last commit's SHA and subject and any open pull request from gh pr status.

gh pr status succeeds when no pull request is open. It omits currentBranch, so an empty result is an answer, not an error. It fails outright without gh, without authentication and on a repository with no GitHub remote. Treat any failure as "no pull request found", say so and offer the remaining routes rather than stopping.

Renames

Rename detection is on by default in git diff, and --name-status reports R100 old/path new/path. Lower the similarity threshold with --find-renames=40% when a file was moved and edited in the same change.

Review a rename as a move, not a delete plus an add. Everything that survived the move is unchanged code, and only the genuine edits are in scope.

Excluded paths

Exclude these and name what you excluded in the scope block. They are machine-authored and carry no interface rules.

CategoryPatterns
Lockfilespackage-lock.json, pnpm-lock.yaml, yarn.lock, bun.lock, bun.lockb, Cargo.lock, composer.lock, Gemfile.lock, poetry.lock, uv.lock
Snapshots and fixtures__snapshots__/, *.snap, *.approved.*, test-results/, playwright-report/
Generated outputdist/, build/, out/, .next/, .turbo/, .svelte-kit/, coverage/, storybook-static/, *.min.js, *.min.css, *.map
Generated sources*.gen.ts, *.generated.*, *.d.ts emitted by a build, GraphQL and Prisma client output
Vendored codevendor/, third_party/, node_modules/
Binaries and media*.png, *.jpg, *.webp, *.avif, *.woff2, *.mp4, *.pdf

Two exceptions stay in scope. A font file added or swapped is a better-typography change. An image added to a component is a better-ui and better-accessibility change, through its outline and its alt text. Review the code that references them, not the bytes.

Apply the exclusions as pathspecs so the file count in the scope block is the reviewed count:

git diff --name-only "$BASE" <head-ref> -- . ':(exclude)pnpm-lock.yaml' ':(glob,exclude)**/dist/**'

Two traps under-exclude silently. *.lock catches yarn.lock and Cargo.lock but not package-lock.json or pnpm-lock.yaml, so cover every suffix in the table. And without glob magic, **/dist/** needs a literal / before dist, so it misses a root-level dist/. Run the diff with and without the pathspecs and confirm the count dropped by exactly the files you named.

Expanding to consumers

Use the project's own resolver where one exists, otherwise import paths.

git grep searches the working tree by default. Pass the reviewed ref after the pattern, or on a pull request you search a different revision and miss importers the change itself added. Results come back as <rev>:path/to/file; read them with git show, never the working-tree copy. Pass -e when the pattern starts with a dash, such as a --color-* token, or git parses it as an option.

For a changed token or theme value, search the token name rather than the file, since consumers reference the name and never import it.

Order the consumers by a rule you can evaluate, so the cutoff is reproducible instead of a guess:

  1. Route and layout entry points first, whatever the framework treats as a rendered surface: app/**/page.*, app/**/layout.*, pages/**, routes/**, src/views/**, *.astro pages. Everything else only appears inside one.

  2. Then by importer count, since a component pulled in by twenty files carries more of the change than one pulled in by two.

  3. Break ties by proximity, same package or feature directory first.

Say plainly if the ordering was arbitrary past a point.

Requirements and compatibility

Written for: Claude Code

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Claude Code frontmatter, Needs a command-line tool, Needs Node.js

Security

No issues found by our static scan· No issues found

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by
Jakub Krehel
Publisher
Jakub Krehel (Community)
Source
jakubkrehel/skills · skills/interface-review · commit f41f027
Licence
MIT License (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by Jakub Krehel or Jakub Krehel.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download interface-review.zip