The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from arbitrary external URLs to capture branding information (via brand.mjs) and screenshots (via capture.mjs). This data is then written to local markdown and JSON files (BRAND.md, brand.json) and presented to the agent, potentially allowing malicious content on a target website to influence downstream agent actions.
Ingestion points: brand.mjs and capture.mjs scripts which navigate to user-supplied URLs.
Boundary markers: Absent in the generated analysis files provided to the agent.
Capability inventory: File system access, network fetches, and shell command execution (ffmpeg, ffprobe, uvx).
Sanitization: Uses standard DOM extraction via headless Chrome; lacks specific filtering for prompt injection sequences in the extracted text.
[EXTERNAL_DOWNLOADS]: The skill instructions and scripts download and execute external packages and machine learning models at runtime without strict version pinning.
Evidence: uvx rembg i -m birefnet-general for background removal, uvx --from lama-cleaner for inpainting, and npx playwright install chromium for browser automation.
[DYNAMIC_EXECUTION]: The skill uses dynamic JavaScript execution within a headless browser to extract branding information from web pages.
Evidence: scripts/_cdp.mjs utilizes the Chrome DevTools Protocol Runtime.evaluate command to run scripts from brand-page.mjs on the target website's context.
[COMMAND_EXECUTION]: The skill invokes several system-level CLI tools for media analysis, transformation, and browser automation.
Evidence: Subprocess calls to ffprobe, ffmpeg, google-chrome/chromium, uv run, and uvx are pervasive across SKILL.md and the accompanying scripts.

asset-prep