Skip to content

Latest commit

 

History

History

README.md

repo-visuals

repo-visuals — a Claude Code plugin that turns any repo into a bespoke hero visual. Scans, proposes scenarios, renders animated GIF, static PNG, or animated SVG.

A Claude Code plugin for designing bespoke GitHub README hero visualsanimated GIF, static PNG, or animated SVG — for any open-source repository.

The plugin scans the target repo, recommends a format that fits its identity (terminal demo, product UI, brand-first wordmark, banner, diagram-as-hero), generates bespoke HTML through a structured discovery dialog, previews it in your browser, and exports a retina-quality artifact ready to drop into the README.

Heroes shipped with this skill

htmlhint/HTMLHint HTMLHint hero gui-cs/Terminal.Gui Terminal.Gui hero
kunwar-shah/claudex claudex hero rullerzhou-afk/clawd-on-desk clawd-on-desk hero
affaan-m/everything-claude-code everything-claude-code hero emtyty/ast-graph ast-graph hero

Full showcase →

Install

/plugin marketplace add livlign/claude-skills
/plugin install repo-visuals@livlign

After install, restart Claude Code (or start a new session) so the plugin's skills are discovered.

Dependencies (install these yourself)

The plugin install does not install these — you need them on your system before running:

  • Node.js 18+
  • puppeteer — run npm install once inside the plugin directory. This downloads ~170 MB of Chromium the first time.
  • ffmpeg + ffprobe (for GIF export and code-evaluated scorecard rows) — brew install ffmpeg / apt install ffmpeg / choco install ffmpeg.
  • gifsicle (optional) — enables palette-size checks in scripts/evaluate.js.

npm install runs from the plugin's cache directory after install. If you can't find it, use /plugin to locate the install path, then cd into it and run npm install.

How to run

Standard run (no scorecard)

Generate a hero GIF for this repo: https://github.com/owner/name

Claude asks which operating mode to use (see below), runs discovery → build → preview → export → output, and stops. No evaluation is collected.

A run feels like this:

You:    Generate a hero GIF for this repo: https://github.com/owner/name
Claude: [scans repo] Inferred CLI tool — `bin/` entry, README opens with a $ npx line.
        Recommending an animated terminal demo at 1200×675. Which mode — Auto / Semi / Manual?
You:    Semi
Claude: [proposes 3 scenarios] Pick one, or describe a different angle.
You:    #2, but make the final frame land on the success state for longer.
Claude: [builds HTML, opens preview in browser] Looks right? Any copy/timing tweaks?
You:    ship it
Claude: → ./hero.gif (1200×675, 4.2s, 1.8 MB)

Final artifact lands in your current working directory. In dev mode, an evaluation row is appended to ./evaluations/runs/<date>-<slug>.md.

Dev run (with scorecard)

Generate a hero GIF for this repo in dev mode: https://github.com/owner/name

Dev mode enables Phase 6 — a 4-rater scorecard (User / Claude / Code / AI-vision-blind). The scorecard is written to ./evaluations/runs/<YYYY-MM-DD>-<slug>.md in your current working directory (not the plugin cache, which is wiped on /plugin update).

Enable dev mode either by saying "dev mode" in the prompt, or by setting REPO_VISUALS_DEV=1 in the environment.

Retro — improve the skill from logs

After you have a handful of dev-mode runs logged:

Run retro on my repo-visuals evaluations

The repo-visuals-retro skill reads ./evaluations/runs/*.md, spots patterns, and proposes diffs against the main skill. It proposes — it does not auto-apply.

Operating modes

Every run, the skill asks which mode to use. This is about how many decisions Claude asks you to make — it does not relax craft rules.

Mode What you decide What Claude decides silently Phase 6 scorecard Typical back-and-forths
Auto nothing everything (format, scenario, vibe, audience, dimensions, copy, ship) Code + AI + Claude rows only (no User ratings) 0
Semi-auto (default) output format (GIF/PNG/SVG/HTML), one preview-and-iterate review scenario, vibe, audience, dimensions, copy full 4-rater scorecard ~3
Manual every decision point nothing (Claude still suggests and recommends) full 4-rater scorecard 8–12

Any mode can be upgraded mid-run — say "stop, switch to semi" and Claude resumes from the nearest unanswered decision point.

Note: operating mode (Auto/Semi/Manual) and dev mode are independent. Dev mode is about whether Phase 6 runs; operating mode is about conversation density.

Layout