The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill injects and executes JavaScript into the user's active Chrome browser session to interact with Claude.ai's internal APIs. This is performed via AppleScript (scripts/runjs.applescript) or Chrome DevTools Protocol (scripts/cdp_channel.py). The skill also uses uv run to execute local Python scripts at runtime.
[COMMAND_EXECUTION]: The scripts/cdp_channel.py and scripts/selftest_fidelity.py scripts use Python's subprocess module to execute system commands such as ps for process enumeration and sys.executable for running auxiliary scripts.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted conversation data from claude.ai, including content from shared links.
Ingestion points: scripts/render_transcript.py processes exported JSON transcripts.
Boundary markers: The transcript is rendered using Markdown headers (## human, ## assistant) and code fences to distinguish between participants and tool outputs.
Capability inventory: The skill has the ability to write to the local filesystem and execute shell commands via Python scripts.
Sanitization: Content is fenced in markdown to minimize accidental instruction obedience; however, the agent remains susceptible to instructions embedded within the processed chat history.
[OBFUSCATION]: scripts/download_files.js uses Base64 encoding to facilitate the transfer of binary files (such as images and Excel documents) from the browser context to the local file system. This is a functional requirement for bridging binary data through string-based communication channels.

read-claude-web-conversation