The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates and executes Node.js and Bash scripts at runtime.
Evidence: SKILL.md describes Phase 2 as generating a script at .agent/recordings/<slug>/record.mjs and Phase 3 as running it.
Mitigation: The skill uses JSON.stringify to escape user-controlled inputs (URL, selectors, interaction parameters) before injecting them into the template, reducing the risk of code injection.
[COMMAND_EXECUTION]: The skill executes various shell commands for tool verification and video processing.
Evidence: rules/preflight.md and templates/crop.sh.template show calls to npx, ffmpeg, ffprobe, jq, and curl.
[CREDENTIALS_UNSAFE]: The skill is designed to read and use sensitive browser authentication states.
Evidence: rules/preflight.md and templates/record.mjs.template reference reading .browser/auth-state.json to populate the Playwright storageState for authenticated sessions.
Mitigation: The skill includes an explicit user consent gate in rules/preflight.md when auth state is detected for non-localhost URLs.
[INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface where malicious content on a recorded web page could influence downstream AI agents.
Ingestion points: The skill navigates to user-provided url values in SKILL.md Phase 1.
Boundary markers: None specified for the recording process.
Capability inventory: The skill captures video that is subsequently processed by the video-analyser skill, which uses vision models to extract findings.
Sanitization: No sanitization of the target page content is performed before recording.
[EXTERNAL_DOWNLOADS]: The skill provides instructions and automation paths for downloading external software.
Evidence: rules/preflight.md suggests running npx playwright install chromium and provides instructions for installing ffmpeg via brew or apt-get.

screen-recorder