The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from screen OCR and audio transcripts which are considered untrusted sources and may contain adversarial instructions.
Ingestion points: Data enters the agent context through the search functionality in screenpipe.mjs which queries the Screenpipe local API.
Boundary markers: The script wraps search results in <private> tags to prevent accidental storage by other tools, and instructions explicitly mandate treating the text as evidence rather than instructions.
Capability inventory: The skill is capable of executing local Node.js scripts, reading the ~/.claude-mem/settings.json file, and making local HTTP requests to 127.0.0.1 or localhost.
Sanitization: Captured text is truncated to a maximum of 2,000 characters per record to limit the amount of external content processed.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute a bundled Node.js helper script to interact with local services.
The execution uses the CLAUDE_SKILL_DIR environment variable to resolve the path to the script securely within the skill's directory.
The helper script strictly validates that all network endpoints are local HTTP origins, preventing the use of the tool for data exfiltration to remote servers.

screenpipe