The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted content from arbitrary external URLs.
Ingestion points: scripts/distill-page.ts uses Playwright to navigate to URLs provided as arguments.
Boundary markers: While the HTML is converted to Markdown, there is no explicit instruction to the agent to treat the resulting text as untrusted data.
Capability inventory: The skill fetches and reads content but does not provide the ability for the content itself to trigger secondary shell commands or file writes.
Sanitization: Employs Chromium's Page Content Annotations ML model to distinguish between primary content (e.g., articles) and boilerplate (e.g., ads, navigation).
[OBFUSCATION]: The skill uses Unicode Private Use Area (PUA) characters as internal markers.
Evidence: scripts/semantic_markers.ts defines PRE_OPEN_MARKER using \uE000 and \uE001.
Context: These are used as technical delimiters to reliably track structural boundaries (like code blocks) during the transformation from browser annotations to Markdown, preventing collisions with actual document text.
[COMMAND_EXECUTION]: Includes a maintenance script that executes shell commands for environment setup.
Evidence: scripts/update-proto.ts uses execSync to run npx buf generate.
Context: This script is intended for developer/maintenance use to update Protobuf definitions from the Chromium source and is not invoked during standard agent task execution.
[EXTERNAL_DOWNLOADS]: Fetches configuration files from trusted official repositories.
Evidence: scripts/update-proto.ts fetches common_quality_data.proto from the official Google Chromium source repository.

distill-page