UI review skill for Claude: Interface Review verdict, safety and alternatives

Interface Review is Jakub Krehel's user-invoked Claude skill that reviews a branch or pull request for UI regressions. We rate it Conditional, 68/100, with safety grade C. Confidence is Low, because the evidence is thin and nobody has measured its findings. Jakub Krehel is a founding design engineer at Interfere, per his site jakub.kr.

Get the skill: interface-review folder on GitHub · SKILL.md at the reviewed commit · skills.sh listing

Changed since our review: commit da5f6d5 rewrote SKILL.md and both reference files on 2026-10-05. The plugin is now version 1.13.2. Our verdict and safety grade cover only the version we reviewed, commit 267330e, on 2026-09-30.

Which UI reviews does the Interface Review skill for Claude handle?

The Interface Review skill for Claude handles change-scoped UI reviews: a branch, pull request or uncommitted work. It does not audit whole screens or screenshots. Its sibling better-interface reviews whole screens.

An Agent Skill is a SKILL.md package that an agent such as Claude Code loads on demand. Our guide explains how Agent Skills load in Claude. Interface Review is one such package, and it runs only when you call it. It takes one code change, works out exactly which lines changed, and reviews the interface those lines render.

Each finding gets one of three labels. Introduced means the change added the problem. Regression means the change broke something that worked before. Pre-existing means the problem was already in the code. The skill belongs to Jakub Krehel's full interface collection, and it leans on that collection to judge severity.

Interface Review stays out of correctness, tests and security. SKILL.md leaves those to normal code review. Use it next to your usual reviewer, never in place of one.

What does Interface Review check in a change?

Interface Review checks a change across accessibility, layout, writing, typography, color and UI. It expands to up to five affected screens, reads removed lines, and labels each finding Introduced, Regression or Pre-existing.

These rules come from SKILL.md and its two reference files at commit 267330e:

• Resolves the merge-base, the commit where the branch split off, before it reads the working tree.
• Excludes lockfiles, generated files and vendored paths from the diff.
• Expands one hop to consumers, or two hops for tokens and primitives, with five consumers at most.
• Scans removed lines against a list of signals, such as a dropped aria-label.
• Compares the change with its stated intent and flags missing hover, focus, disabled, loading, empty and error states.
• Flags untranslated strings and incomplete variants.
• Caps pre-existing findings at three, so old debt stays short.
• Forbids checkout, switch and stash, so the review stays read-only.
• Hands severity and the final verdict to better-interface.

The output has three parts: a scope block, a coverage table and a findings list with a Status column. The scope block states what was reviewed. The skill refuses to invent a target when none is given.

Is the Interface Review skill worth installing?

Yes, it is worth installing, but only inside the full jakubkrehel/skills collection. The Interface Review skill scores 68/100, Conditional, at Low confidence. Its scope rules are careful and read-only, yet nobody has measured whether its findings are accurate.

Interface Review is a carefully scoped, read-only change reviewer for UI diffs. It depends on sibling skills, and its removed-line scan carried a wrong-range bug at our commit. It fits teams already running the full collection who want branch or PR findings kept apart from existing debt.

It is a poor fit in two cases:

• Installing interface-review alone, without better-interface and the better-* domain skills.
• Running automated CI review of untrusted fork pull requests, where PR bodies reach the agent unsanitised.

The overall score leans toward the worst failure. At commit 267330e, the removed-signal scan read a hard-coded range that misses pull request refs (issue #34). That defect reached one scan, not the whole review. The bug, the unsanitised PR text and the sibling dependency keep it at Conditional. Thin evidence sets the confidence to Low; it does not lower the score.

Our page on how we score skills defines each measure below.

The evidence is thin. The missing piece is accuracy on real pull requests. That covers the precision of Regression statuses, missed removals and behaviour on a fork PR. To see how it ranks against other design skills, read the best UI design skills, ranked.

How many tokens does Interface Review use?

Interface Review loads 2,672 tokens per run, measured from file size, or 5,438 with references. Our estimate for a typical 10-file UI pull request is 40,000 input and 3,000 output tokens, before siblings load.

Beyond its own folder, each run loads better-interface and several better-* domain skills. It reads the diff and up to five consumer files as well. Input tokens grow with pull request size; that is an estimate, not a measurement.

What changed since our review?

The Interface Review files changed after our 2026-09-30 review. Commit da5f6d5 rewrote the removed-line scan behind issue #34 and the pathspec note, and the plugin moved to 1.13.2. The verdict stays Conditional, 68.

We checked the repository on the GitHub API on 2026-10-21. These are repository-confirmed file changes; we have not re-evaluated them.

• 2026-10-05: commit da5f6d5 edited SKILL.md (+46 −49), removed-signals.md (+17 −12) and scope-resolution.md (+21 −17).
• 2026-10-05: the removed-signal command now diffs "$BASE" <head-ref>, and it drops the head ref for uncommitted work.
• 2026-10-05: the pathspec note in scope-resolution.md was rewritten.
• 2026-10-05: the skill description was rewritten to name a branch, pull request or uncommitted change.
• 2026-10-07: the maintainer closed issue #34 as "not planned".
• 2026-10-21: PR #21 is still open and unmerged.
• 2026-10-21: the plugin manifest reads 1.13.2, and HEAD sits 28 commits after 267330e.
• 2026-10-21: the collection added build-design, state-machine and technical-drawing.

Two things did not change. disable-model-invocation: true still keeps the skill user-invoked. No guidance on untrusted PR text was added, so the Snyk W011 concern stands. Issue #34 is closed as not planned, yet the file now uses a resolved range.

Is the Interface Review skill safe to install?

Yes, the Interface Review skill is safe to install with care: safety grade C. It has no scripts, but it runs git and gh, fetches pull request refs and reads untrusted PR text. Snyk flags that risk.

The grade comes from reading the files at commit 267330e. Our guide on how to vet a skill before installing walks through the same checks. We found nine facts in the files:

1. Four files ship: SKILL.md (10.7 KB), scope-resolution.md, removed-signals.md and agents/openai.yaml.
2. No scripts, hooks or MCP config are bundled; the folder is roughly 5.4K tokens with references.
3. The agent runs git merge-base, diff, blame, fetch and worktree, plus GitHub CLI (gh) pr view, status and diff.
4. Network use covers git fetch of PR refs and gh calls, which need a GitHub remote and gh auth.
5. Writes stay under .git by design; checkout, switch and stash are forbidden.
6. Worktrees are opt-in, and the skill removes them afterwards.
7. PR titles, bodies and commit messages count as stated intent, with no instruction to treat them as untrusted.
8. disable-model-invocation: true and allow_implicit_invocation: false mean it runs only when you call it.
9. It needs better-interface to assign severity and verdict, and it stops when that skill is missing.

Three audits on skills.sh, a public directory of Agent Skills, were recorded on 2026-08-30. Snyk, a security scanner, returned Warn, with finding W011 at medium risk. Socket returned Pass. Gen Agent Trust Hub is listed as Pass. Its audit page rates it Safe overall, with medium notes on shell use and missing injection guidance. We show both labels because both sources say so.

The practical risk is prompt injection through a pull request description. A hostile PR body reaches the agent as stated intent. Keep a human on every review, and avoid running it unattended on fork pull requests. Snyk's own ToxicSkills research reports prompt injection in 36% of the skills it tested and 1,467 malicious payloads. Those are Snyk's figures, not ours.

What does the community report about Interface Review?

The community reports on Interface Review are thin: nine items, four of them strong. Those four cover a wrong-range bug, an unmerged errata PR, a Snyk warning and an Agent Trust Hub note.

Our evidence comes from GitHub issues, pull requests and audits dated before the 2026-09-30 review.

What are the Interface Review skill's known failures?

The Interface Review skill had two known failures in its own files at commit 267330e. Its removed-signal scan read a hard-coded diff range (issue #34), and a pathspec note inverted git's glob rule. Both files were rewritten on 2026-10-05.

The worst plausible failure is a review of the wrong diff. The skill then reports Approve or false regressions. Its scope block still claims full coverage. That is why the first row weighs most in our score.

The rewrite's effect on issue #34 is repository-confirmed in the file but not independently reproduced. Treat the output as a structured checklist for human-reviewed pull requests, never as an unattended gate.

Which agents does the Interface Review skill work with?

The Interface Review skill works in Claude Code, the only verified agent. It ships Codex metadata, but no Codex run is documented. It needs git, and gh with authentication for pull requests.

Interface Review needs git for every target. Pull request targets add the GitHub CLI, gh, with a logged-in account and a GitHub remote. It needs no MCP servers. Our list of Claude Code skills ranked by verdict shows other verified options.

How do you install the Interface Review skill?

Install the Interface Review skill with the whole collection: npx skills add jakubkrehel/skills. In Claude Code, add the interfaces plugin instead. Then run /interfaces:interface-review on a branch or pull request.

These commands come from the repository README, checked on 2026-10-21:

1. Install the whole collection with the skills CLI:

npx skills add jakubkrehel/skills

1. Add the marketplace and the plugin in Claude Code instead, as two separate commands:

/plugin marketplace add jakubkrehel/skills
/plugin install interfaces@interfaces

1. Authenticate gh with gh auth login before you review pull requests.
2. Invoke /interfaces:interface-review pr <n> after a plugin install, or /interface-review after a skills CLI install.

Interface Review stops without better-interface, so a single-skill install gives you only a scope block. Our complete Claude Code skills setup guide covers folders, scopes and updates.

Can the Interface Review skill run a UX audit or design review?

No, the Interface Review skill does not run a full UX audit. It reviews the interface changes in one diff. For a design review of a whole screen, use better-interface or a heuristic UX audit skill.

UX audit skills are a different kind of tool. Those skills audit screenshots or live pages against heuristics such as Nielsen's 10 usability heuristics and WCAG. The designcareer.guide skill, for example, takes screenshots and writes an HTML report. We have not evaluated those UX audit skills.

Our evidence supports three tasks:

1. Review an open pull request for interface regressions before merge (primary, per PR).
2. Check the current branch plus uncommitted work before pushing (primary, several times a week).
3. Route change-scoped requests from a wrapper skill into Interface Review (secondary, occasional).

One more use is claimed only, in SKILL.md section 6. It catches incomplete variants, missing states and untranslated strings against a PR's stated intent. Nobody has shown that in a dated public report yet.

Test it yourself:

• Open a pull request that removes an aria-label.
• Check out a different branch locally.
• Run /interface-review pr <n> on that pull request.
• Check whether the removed signal is reported against the right range.

What are the alternatives to the Interface Review skill?

The closest alternatives to the Interface Review skill are four evaluated skills. Web Design Guidelines (66) checks UI files against Vercel rules, and Anthropic Design (66) critiques designs. UI/UX Pro Max (68) and UI Skills (72) guide builds.

None of these alternatives reviews a single diff the way Interface Review does. They check files, screens or designs as they stand. Our hub shows every design skill we rank by score.

How does Interface Review fit into Jakub Krehel's skills?

Interface Review is one of eleven skills in Jakub Krehel's collection at review. It hands every finding to better-interface for severity and verdict. The collection scores 70/100, Recommended with caveats.

The six better-* domain skills hold the rules for accessibility, colors, layout, typography, UI and writing. The better-interface skill combines them for one screen. Interface Review adds the change scope on top. The collection had eleven skills at review and fourteen on 2026-10-21. Interface Review had 12.9K installs at review, against 210.2K for the whole collection. Read our verdict on the whole collection for the pack score.

Where is the Interface Review skill on GitHub?

The Interface Review skill lives on GitHub in jakubkrehel/skills at skills/interface-review, under MIT. We reviewed commit 267330e, plugin version 1.6.3. The repository has no tags and 8,074 stars on 2026-10-21.

• Repository: github.com/jakubkrehel/skills.
• Skill folder: skills/interface-review/.
• Reviewed commit: 267330e1adfc66a718fb65fa6918c1f06d0a689e, committed 2026-08-29.
• Plugin version: 1.6.3 at review, 1.13.2 on 2026-10-21.
• Stars and forks: 7,385 and 274 at review; 8,074 and 310 on 2026-10-21.
• Code-search copies: 189 SKILL.md files named interface-review, forks included.
• File hash: SKILL.md sha256 65783ac2….

Common questions

Is the Interface Review skill free?

Yes. It is MIT-licensed.

Does Interface Review work without better-interface?

No. It reports the scope and file inventory, names the missing skill and stops.

Does Interface Review work in Codex?

Not verified. Codex support is claimed through agents/openai.yaml, and no Codex run is documented.

Can Interface Review gate CI on fork pull requests?

No, we advise against it. Pull request text reaches the agent without injection guidance (Snyk W011).

Can Claude do an SEO audit with this skill?

No. Interface Review covers interface quality only; see our Claude SEO skill verdict for SEO audits.