screencast-recorder

Record WebM/MP4 screencasts of browser interactions using shot-scraper video. Accepts YAML storyboards natively or RDF Turtle via the ttl-to-yaml.py converter.

Prerequisite Check

Before recording, verify shot-scraper is available:

command -v shot-scraper >/dev/null 2>&1 || uv tool install shot-scraper

Install Playwright browsers if missing:

shot-scraper install 2>/dev/null || python3 -m playwright install chromium

Output Convention

Recordings go to {LLM_ROOT}/{model}/screencasts/ where {LLM_ROOT} is the canonical LLM root from preferences artifact-routing rules ({LLM_ROOT}), and {model} is the model-specific output directory (e.g., Big Pickle, DeepSeek, Claude Generated). Use descriptive filenames: {YYYY-MM-DD}-{topic}.mp4 (with companion .webm).

When recording, derive the path at runtime by looking up the model's output path per agent-rdf-memory/howto/artifact-routing.ttl and appending screencasts/.

Workflow

Step 1: Determine Mode (disambiguate automatically)

Mode A — curl → screencast: User explored resources via mtls-curl commands. Parse the conversation for curl URLs, present the extracted URL list to the user.

Mode B — direct browser: URLs were discussed directly (not via curl). Present them as-is.

If neither is clear, ask: "What URLs or resources should the screencast show?"

Step 2: Elicit Scene List

Present a curated list to the user with numbered scenes:

"I found these URLs from the session:

1. {url-1}
2. {url-2}
3. {url-3}

Want all of them? Adjust order? Add scene names? Set pauses between scenes?"

Capture user preferences:

• Scene ordering — reorder, drop, or add scenes
• Scene names — labels for each scene (shown in progress output)
• Pauses — how long to pause on each scene (default 2s for readability)
• Cursor — visible cursor with click rings enabled by default
• Viewport — default 1440x900 (wider for SPARQL result pages)

Step 3: Build Storyboard