Interactive Video: $ARGUMENTS

Generate a self-contained HTML walkthrough from $ARGUMENTS with canvas animations, narrated audio, quizzes, and interactive widgets.

Parse $ARGUMENTS for source (URL, file, topic, or "codebase"), scope hints ("focus on X", "skip Y"), and audience/use-case hints ("onboarding", "knowledge transfer", "walkthrough", "beginner-friendly"). Pre-fill extracted hints into configurator URL params.

Gotchas (things the references and validator don't catch)

• Title-only visuals while narration gives details. Extract 3-5 key phrases per scene and stagger revealAt to match when they're spoken: revealAt = (phraseTime - sceneStart) / (sceneEnd - sceneStart). The canvas should reinforce what the ear hears.
• Scenes too short for narration. Don't create 5-second scenes with a single text element. Each scene needs enough content to fill its duration with progressive reveals.
• Hardcoded colors instead of theme palette. Pull colors from styles.json for the selected theme — hardcoded colors break theme switching.
• `interactions` in content.json is an object keyed by ID, not an array. The engine looks up by ID: CONTENT.interactions[ix.id].
• Visual reveal order mismatches narration order. If the narration describes a stack bottom-to-top, the visual must reveal bottom-to-top (or flip the layout). The most recently revealed element must always be the thing currently being narrated. See design-guide.md § Spatial-temporal alignment.
• Dead scenes. A scene where everything fades in and sits still feels like a slideshow, not a video. Every scene >20s should have at least one source of continuous motion — flowing dashed arrows, a pulsing glow, gentle drift. See design-guide.md § Alive vs dead scenes.

Quality principles

• Research grounds the content. Verify facts before writing. Calibrate depth to the use case.
• Concrete beats generic. "The auth service uses JWT with 15-min expiry" beats "many systems use tokens."
• Narration is conversational. Direct, "you" voice. Like explaining to a new team member on their first day. See design-guide.md § Narration voice for banned patterns and anti-slop rules.
• Lesson title: short noun phrase. Can include audience level if it sets expectations. No "Understanding...", "A Guide to...", "Introduction to...". Examples: "DNS Resolution", "Git Rebase for Beginners", "The Maillard Reaction — Advanced".
• Lesson subtitle: one-line summary of what the lesson covers, followed by the source. Examples: "From query to IP in four hops. Based on cloudflare.com/dns." / "How this Rust CLI aggregates logs. Source: github.com/foo/bar." If multiple sources, list them. Never generic.

Phase 0: SETUP

0.1 One-time setup

Run bash $SKILL_DIR/scripts/setup.sh on first use. Skip if edge-tts is already on PATH.

0.2 Confirm language

Detect from source. Confirm with user — this is the only CLI question.

0.3 Research + extract topics

• URL → WebFetch the page. Optionally WebSearch for supplementary context.
• File path → Read the file. WebSearch for background on frameworks/concepts it references.
• "codebase" → Explore with Glob/Grep/Read. WebSearch for documentation on patterns found.
• Topic string → Research in parallel using Agent tool (WebSearch + WebFetch per agent). Wait for all agents, then extract topics from the research.

Extract the natural topic structure from the source as JSON:

[{ "id": "topic-id", "label": "Topic Name", "desc": "One-line description" }]

0.4 Open configurator

cd $SKILL_DIR && npx tsx scripts/build_audio.ts --generate-previews --lang {lang}
cd $SKILL_DIR && npx tsx scripts/serve_configurator.ts \
  --lang {lang} \
  --source {encoded_source} \
  --topics {encoded_topics_json} \
  --audience-hint {hint_or_empty} \
  --source-type {url|file|codebase|topic}

The server opens the configurator in the browser and blocks until the user clicks Done. Config JSON is returned via stdout. If the command times out (10 min) or the user pastes a Preferences: block manually, fall back to text parsing.

0.5 Set output directory + copy engine

mkdir -p {out}/src/engine && \
cp -n $SKILL_DIR/engine/*.js {out}/src/engine/ && \
cp -n $SKILL_DIR/engine/lesson.css {out}/src/

After this: no more questions. Run Phases 1-4 without stopping.

Phase 1: PARSE PREFERENCES

Parse the config JSON from Phase 0.4 stdout. Extract: audience, length, interaction, style, voice, voiceName, topics, preset.

If the user pasted a Preferences: text block instead (fallback), parse it as before: extract key-value pairs from the text.

Theme keys are defined in styles.json. Length targets and interaction design rules are in design-guide.md.

Phase 2: STRUCTURE

Read $SKILL_DIR/references/design-guide.md. Use it as a reference — adapt to what the content needs.

Phase 3: GENERATE

Read the reference files you need (always in one parallel message):

• Always: page-template.md, styles.json, engine-contracts.md
• For custom draw functions or complex animations: also render-patterns.md

Read(engine-contracts.md)
Read(page-template.md)
Read(styles.json)
# Read(render-patterns.md) — only if this lesson uses custom draw

Write content.json → {out}/src/content/{lessonId}/content.json

• Follow schema in engine-contracts.md exactly
• Leave t and duration as 0 — build_audio fills them

Start audio in background → immediately after writing content.json:

cd $SKILL_DIR && npx tsx scripts/build_audio.ts {lessonId} \
  --content-dir {out}/src/content \
  --audio-dir {out}/audio/lessons

Use run_in_background: true. Audio only needs content.json.

Write HTML while audio generates → {out}/{lessonId}.html

• Follow the template in page-template.md exactly

Phase 4: FINALIZE

After audio completes (you will be notified):

1. Read updated content.json for computed timing
2. Patch HTML: scene s/e values + IX time values

Scene timing rules:

• First scene starts at 0. Last scene ends at meta.duration.
• Each scene's s equals the previous scene's e. No gaps. The audio gap (1.2s silence between segments) does NOT create scene gaps — scenes are wall-to-wall.
• Use interaction time values and narration t values as guides for where to place boundaries, but always ensure contiguity.

Validate + serve in one command:

cd $SKILL_DIR && npx tsx scripts/validate.ts {lessonId} \
  --content-dir {out}/src/content \
  --audio-dir {out}/audio/lessons \
  --html {out}/{lessonId}.html && \
cd {out} && npx --yes http-server -p {port} -c-1 --silent & \
sleep 1 && open "http://localhost:{port}/{lessonId}.html"

Show the user: title, duration, scene count, interaction count, scene breakdown, files created.

Optional: visual verification — skip if it fails (Playwright may not be installed):

cd $SKILL_DIR && npx tsx scripts/visual_verify.ts {lessonId} \
  --html {out}/{lessonId}.html --out {out}/verify-{lessonId} || true

If it succeeds, screenshots are saved to verify-{lessonId}/. If it fails, continue — the lesson is already validated and served.

Log the lesson — append to ${CLAUDE_PLUGIN_DATA}/lessons.json (create if missing):

{ "id": "{lessonId}", "title": "...", "theme": "...", "duration": 0, "scenes": 0, "interactions": 0, "date": "YYYY-MM-DD", "output": "{out}" }

This lets future invocations see what's been built before.