Skip to content
workingBenPublic

About

Claude Code skills that show their work: design reviews, visual QA, copy edits and decision pages to communicate visually with Claude.

Topics

Resources

Stars

1 star

Watchers

1 watching

Forks

Repository files navigation

Claude skills that show their work. Skills for Claude Code.

Use Superpowers to plan and build,
and Proofs to see and decide.

Claude Code plugin 12 skills MIT license tests Trigger evals: 108 of 108 on clear requests, 42 of 42 on ambiguous ones

Stop reading walls of terminal text. Every skill here includes a page you can look at, click through and decide on, then hand back to Claude in one paste. Design reviews, visual QA, copy edits and decision pages for Claude Code: Claude builds the review, you choose from real options, and Claude builds what you chose.

In Claude Code:

/plugin marketplace add workingBen/proofs
/plugin install proofs@proofs

Install the whole plugin, not single skill folders. The skills share a page builder and reference files at the plugin root; a skill copied on its own into .claude/skills/ can't find them, and says so.

Screenshots shrink automatically when ImageMagick or libwebp (cwebp) is installed: they're converted to WebP as each page is built. Without either, pages still work, just heavier.

Browse the site for all 12 skills and 230+ features, each with a short write-up.

The 30-second Proofs tour: a design review shows the live page beside two coded options, each with a Pick this button; a before and after wipe; a page checked on a phone with its fixes; Tap. Paste. Done.; ending on the two install commands.

The 30-second tour. Every picture is a real page, before and after. Full-size MP4 (1920x1080, silent).

Watch the one-minute tour on YouTube

See it work

Real pages, before and after one review. See all of them, as sliders.

design-review: A pale boxed card becomes a full-bleed sunset hero. (Retreat page hero, before and after) design-review: A bare date picker gains the room photo and a teal stay. (Bare date picker to a room-and-dates step, before and after) article-copy-edit: A plain sentence becomes a tracked edit with a claims check. (Plain sentence to an edit with a claims check, before and after) design-review: Three plain links become fanned stacks of real report pages. (Plain report links to page stacks, before and after)

More examples (4)

design-review: A text-only card back becomes a wildflower-photo card. (Text-only marketing flier to a photo card, before and after) design-review: A text section gains a ranked green-to-red bar chart. (Supplement test text to a ranked bar chart, before and after) design-review: A dense readings table becomes seven colored range gauges. (Biofield measurement table to range gauges, before and after) design-review: A long Gantt list becomes a two-lane timeline. (Gantt list to a two-lane study plan, before and after)

After installing

Ask for a skill by name, like design review for /pricing, or run /proofs:visual-qa /checkout. To get new versions later: /plugin marketplace update proofs.

No claude.ai artifacts? Pages open in your browser instead, or on your phone over Wi-Fi. See Where the Page Goes.

What runs automatically

One hook, and it is off until you turn it on. At the start of each session (a SessionStart hook in hooks/hooks.json), it tells Claude how many pull requests are open in the current repo and lists them oldest first with their age, so an unreviewed one cannot sit forgotten.

  • Off by default. It does nothing unless PROOFS_PR_REMINDER=1 is set for the project.
  • Never blocks a session. No token, no network, GitHub down: it exits quietly and the session starts as usual. Its one GitHub call gives up after 5 seconds.
  • Turn it on from your project's root folder (the script edits the .claude/settings.json of the folder you run it in): python3 <the proofs plugin folder>/skills/pr-qa/scripts/install_session_hook.py. Or add "PROOFS_PR_REMINDER": "1" to the "env" block of that file yourself. It reads GitHub access from GITHUB_TOKEN, then gh auth token.

How every skill works

Four steps. 1: Claude shows you first: mockups, measurements and drafts, before anything changes. 2: Publishes a page. 3: You decide visually, with clear buttons on every item and room for a comment. 4: Paste it back, with one Copy button, and Claude does exactly what you chose.

Every skill below plugs into this loop.

Why paste it back?

Because your agents keep working while you decide, and you can decide from anywhere, with nothing running on your computer. Tools that send your answers straight back to the agent are smoother, and they pay for it in other ways:

Proofs Lavish make-pages-interactive
Keeps working while you decide Yes No, it waits Yes, it watches an inbox
Decide on your phone Yes, any device With Tailscale Not stated
Server on your computer None Yes Yes

Both are good tools, and neither replaces our visual communication skills.

How it's built

Proofs is a small, tested codebase, not just a folder of prompts.

One page builder for every skill. Design reviews, copy edits, photo picks, style audits, questions, QA checklists and the working log all come from one builder (shared/review_page/). A skill writes the content as a JSON spec, and the builder makes a single self-contained page with every mechanic built in: decision buttons, variant galleries, a sticky tally, saved progress, light and dark themes, and a Copy button that falls back to a selectable text box when the clipboard is blocked. Every page is checked before it's published, and can be clicked through in a real browser (--smoke).

Questions come first. Anything waiting on you is at the top of the page, above every recap and log entry. The builder enforces it: a page that puts anything above the open questions fails its check.

Tests and evals.

  • 97 unit and browser tests cover the builder, the exports, the site and the README tiles (python3 -m unittest discover -s tests).
  • 50 trigger evals check that the right skill fires for a real request, that its nearest sibling doesn't, and that a plain coding request fires none (claude plugin eval . --ablation none). The latest run passed 108 of 108 on everyday requests and 42 of 42 on deliberately ambiguous ones.

Repo checks.

  • bin/check-skills holds every skill to the same rules: a description that says when to use it, a version and a changelog.
  • bin/bump-version bumps the plugin version whenever a commit changes what ships (skills, shared code, references, hooks), so claude plugin update sees every release. bin/install-hooks runs it, with the other checks, on every commit and merge.
  • bin/privacy-scan blocks private names and details from ever reaching a public commit, since the examples come from real projects.

Works with Superpowers

Superpowers is the best-known workflow skill set for Claude Code (brainstorming, plans, writing tests first, subagents), and it is excellent. It works as text in the terminal. Proofs sits next to it as the visual decision layer, not a rival: use Superpowers to plan and build, and Proofs to see and decide. This project has no affiliation with or endorsement from Superpowers.

Four skills up close

design-review: see your live page beside real mockups, then approve what ships

A live retreat page wiping into the mockup that shipped.

Claude puts a real screenshot of your page next to one to three working mockups per section. You pick one, and Claude builds it.

  • A real screenshot as the Before
  • Mockups drawn in your site's own fonts, colors and spacing
  • One decision per item, then one paste back

Read the design-review page

visual-qa: measure your pages the way people see them, then fix what fails

Before and after: lines on screen 1 to 8, tap targets 32 to 48 px.

Opens every page at desktop, laptop and phone sizes and measures it: text too small, tap targets too tight, text cut off, too few items on the first screen. Every problem gets a mockup fix, and you get a before and after.

  • Real numbers, not opinions
  • Before and after screenshots
  • Hands its findings to design-review

Read the visual-qa page

working-log: never scroll endless terminal output again

A session page: a top bar with a Copy my replies button, Waiting on you asking whether to ship the refund flow with answer buttons, a background log entry under it, and Checks for you and Decided at the side.

Claude keeps one page per session up to date while it works. Anything waiting on you sits at the top with answer buttons, then checks only you can run, then the newest news, each with a read tick. Tap your answers, press Copy, paste. Check it from the couch on your phone, and nothing you need is buried 400 lines up.

  • Questions you answer with a tap
  • Read it on your phone, anywhere
  • Decided items fold out of the way

Read the working-log page

article-copy-edit: review copy edits word by word, with a claims check first

A tracked edit with a claims check flagging a medical promise.

Shows the edits on a page styled like your real article. A claims check comes first, and nothing changes in your files until you have decided.

The claims check is set up for your field: medical and wellness promises, financial guarantees, legal outcomes, product-safety or environmental claims, and more. It asks once what's risky for you, remembers the answer, and also catches what every site gets wrong: guarantees, "results vary" left off, and missing affiliate disclosures.

  • Keep, use, comment or revise, per edit
  • A reason on every suggestion
  • Risky claims flagged before anything else

Read the article-copy-edit page

All 12 skills

Most useful first. Each name links to that skill's page.

working-log: One page per session: what's waiting on you first, then the log, with one Copy button. (Top pick: Most used) design-review: Your live page next to one to three mockups per section. You pick, Claude builds it. (Top pick: Start here) pr-qa: Before and after screenshots of each open pull request, with one decision per PR. (Top pick: Critical for UI changes) questions: Claude's questions on a page, with real buttons, its pick marked and one Copy. visual-qa: Measures pages at three screen sizes, finds layout problems, and designs the fixes. article-copy-edit: Line edits shown on your real article, with a claims check first. review-cycle: Report bugs and ideas as they come; each is fixed, reviewed or saved for later. qa-checklist: The checks only you can run, each marked Pass, Fail or Couldn't test. todos: Your to-do list as a tickable page. Done items are checked before they are archived. photo-picker: Licensed photo candidates for every spot that needs one, in a gallery you choose from. style-guide: Audits the whole site for one consistent look and saves the rules you accept. changelog: A

FAQ

Do I need anything besides Claude Code?

Review pages are published as Claude artifacts when you have them, and open in your browser otherwise. Beyond that, a few skills need small extras:

  • Python 3 for todos and pr-qa (their scripts use only the standard library).
  • gh (the GitHub CLI) or a GitHub token for pr-qa.
  • Node and Playwright for visual-qa and for screenshots.
  • Superpowers is optional. review-cycle uses it when it is installed and works without it.
Will these work in my stack?

Any stack. The pages are plain HTML, and each skill asks your project how it starts, tests and deploys instead of assuming. They were built and tested on Rails sites, so reports from other stacks are welcome.

Can I change a skill for my own team?

Yes. Fork it, or put your house rules in your own CLAUDE.md and the skills will follow them.

License

MIT. Made with Claude Code. Improvements welcome. See LICENSE.

About

Claude Code skills that show their work: design reviews, visual QA, copy edits and decision pages to communicate visually with Claude.

Topics

Resources

Stars

1 star

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages