The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Node.js scripts (scripts/evaluate.mjs) that invoke the browse CLI tool for site navigation and automation.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests and processes untrusted data from external websites (via execution traces and summaries) to update its automation strategies (strategy.md).
Ingestion points: Browser activity and site content are recorded in summary.md and trace.json files within the ./autobrowse/traces/ directory.
Boundary markers: The skill does not implement clear delimiters or instructions to ignore embedded commands within the processed trace data.
Capability inventory: The agent possesses Write, Edit, and Bash capabilities, which are used to modify project files and execute automation code.
Sanitization: Website content is processed without evident sanitization or validation before being used to inform strategy updates.
[COMMAND_EXECUTION]: The skill includes a 'graduation' feature that programmatically writes new skill files to the user's home directory (~/.claude/skills/), which could allow for the persistence of generated logic across different sessions.

autobrowse