SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillswalkthrough-record

walkthrough-record

Testing & quality

View repository

Updated 10-16 07:48

About this skill

walkthrough-record

Runs the bundled Playwright recorder: a scenario YAML in, a cinematic webm master out, plus mp4 (H.264, faststart) and gif (two-pass palette) exports. The recorder drives a real Chromium, injects an overlay engine (eased bezier cursor, click ripples, spotlight ring, privacy masks, explanation cards, camera zoom) and captures with Playwright's page.screencast at crisp 2× resolution. An optional framing pass then insets the capture into a wallpaper canvas — rounded corners, soft shadow, optional macOS window header.

It fixes the failure of hand-driven capture: shaky cursor, missed beats, unreproducible takes, soft text, and 20 MB GIFs.

When NOT to use

  • Writing or refining the scenario YAML itself (beats, card copy, zoom placement, themes) → walkthrough-storyboard.
  • OS-level screen capture of arbitrary desktop apps or live presentations.
  • Playwright test authoring, screenshots, or generic ffmpeg conversions.

Workflow

  1. Setup once per machine. Run bash "${CLAUDE_SKILL_DIR}/scripts/setup.sh". It verifies Node ≥ 20 and ffmpeg, installs pinned deps (playwright 1.61.1, yaml 2.8.0) next to the scripts, and installs Chromium. It exits non-zero with the exact missing prerequisite — surface that message, don't improvise installs.
  2. Get a scenario. Use the one the user provides, or storyboard one first (that's the sibling skill's job). No scenario and they just want to see it work? Record the bundled demo: "${CLAUDE_SKILL_DIR}/assets/demo/orbit-tour.yaml".
  3. Record. 
    node "${CLAUDE_SKILL_DIR}/scripts/record.mjs" scenario.yaml --out ./walkthrough-out
    Flags: --validate-only (schema check, guaranteed no browser), --webm-only (skip exports while iterating), --headed (watch live), --out DIR. The recorder validates the scenario BEFORE opening a browser and exits 1 listing every schema violation — relay those verbatim; fix the YAML, never the validator. Authenticated targets: a storageState scenario performs real clicks with a real session — confirm the origin is a staging/test account with the user before recording anything production.
  4. Verify before declaring success. Check the logged mp4 check: line (width,height,frames | seconds), then extract 3–4 spot frames and look at them: 
    ffmpeg -y -i out/<name>.mp4 -vf "select='not(mod(n\,400))'" -vsync vfr /tmp/frame%d.png
    Confirm: cursor visible and smooth, cards readable and inside the frame, zooms centered on their targets, no masked content leaking.
  5. Frame it (optional, off unless the scenario asks). output.frame insets the capture into a wallpaper canvas. It runs at export time — the webm master stays full-bleed, so a frame can be added, restyled or dropped by re-exporting. 
    output:
  frame:
    background: auto      # auto | studio | midnight | spotlight | paper | any CSS colour/gradient
    chrome: false         # macOS window header (12 pt lights, 28 pt bar) — default off
    title: "Orbit"        # header title, chrome only
    pad: 0.062            # canvas padding, fraction of canvas width
    radius: 13            # corner radius in CSS px (10 = real macOS, the default when chrome is on)
    output.frame: true takes every default. background: auto samples the recorded app's brand colour and light/dark-ness and derives a matching wallpaper — the right choice when you don't know the app in advance. Pick a named background when the video has to match a brand you already know.
  6. Tune exports when asked for platform targets (README, YouTube, socials): budgets and recipes live in references/export-profiles.md. GIF too big? Trim to one beat (gif: {start, duration}), then narrow width, then lower fps — re-export from the existing webm, don't re-record (framed runs keep their *.frame-plate.png / *.frame-mask.png next to the outputs precisely so that re-export still works; the recipe is in references/export-profiles.md).
  7. Failures (selector not found, wrong zoom area, blank video, sticky-header artifacts): diagnose with references/troubleshooting.md. Never edit scripts/ to work around one app; fix the scenario or document the limitation.

Output spec

  • <out>/<name>.webm (master, always full-bleed — the frame is never baked into it), <name>.mp4 (H.264 CRF 18, yuv420p, faststart, CFR at output.fps), <name>.gif (palettegen/paletteuse two-pass). Each file is written atomically (never truncated); a failure mid-run can still leave the earlier completed files of that run — same-named outputs are overwritten.
  • Framed runs additionally keep <name>.frame-plate.png (canvas-sized: wallpaper, shadow, rim, header) and <name>.frame-mask.png (the rounded-corner alpha) — the re-export inputs, not junk to clean up.
  • A verification note: probe line + which frames you inspected.

Gotchas

  • Camera zoom is a CSS transform on <body> — by design (GPU-smooth easing). Apps whose fixed/sticky chrome or JS reads viewport rects during a zoom can misbehave; the fix is scenario-side (zoom-free beat or different target), see troubleshooting. Compositor-side zoom is roadmap, not shipped — don't promise it.
  • 2× crispness uses html {zoom} with a doubled viewport, so media queries see base×2 width. Desktop-first apps render identically; apps with max-width breakpoints in between need output: {scale: 1}.
  • Masks are opaque and fail-closed: solid covers (not blurs — blur leaks shape/length), re-resolved every frame; a mask selector matching nothing aborts the run; navigations paint a curtain until masks re-apply. Don't "fix" an abort by deleting the mask — find the right selector. Masks can't reach inside cross-origin iframes.
  • wait: {until: networkidle} hangs on apps with persistent websockets — prefer wait: {for: selector}.
  • Exports overwrite same-named outputs (iterative renders are the norm); use distinct --out dirs to keep takes.
  • Recording is realtime: keep the machine unloaded during capture; verify pacing with the ffprobe commands in troubleshooting if output looks stuttery.
  • A frame costs ~20 % of the app's linear resolution (2560 → 2054 px at the default padding). The master is 2× raster, so the app still lands at 1.6× and text stays sharp — but don't stack a frame on output: {scale: 1}.
  • background: auto samples once, at the first overlay injection. A tour that toggles the app to dark mode halfway keeps the wallpaper it started with; the rim and shadow still separate the window. Pin a named background if that bothers you.
  • Framed GIFs change palette defaults to 256 colours + sierra2_4a: gradient wallpapers band visibly under the unframed 128-colour bayer default. Diffusion at 128 colours is both uglier and bigger — don't "optimise" it back down.

Files

  • scripts/record.mjs — recorder CLI (validates, records, exports; exit ≠ 0 on any failure)
  • scripts/overlay.js — injected overlay engine (cursor/cards/ring/masks/camera)
  • scripts/frame.mjs — export-time framing (wallpaper/shadow/corners/macOS header); node frame.mjs --self-test
  • scripts/setup.sh — one-time dependency setup, scripts/themes/ — theme presets
  • assets/demo/ — bundled demo app (orbit.html) + demo scenario (orbit-tour.yaml)
  • references/export-profiles.md — mp4/gif budgets per destination (README, YouTube, Vimeo, in-app)
  • references/troubleshooting.md — failure→fix table, zoom compat notes, pacing audit commands