The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from external sources, presenting a surface for indirect prompt injection attacks.
Ingestion points: External data from news mentions, social media, and AI search engine responses are ingested via scripts in the scripts/connectors/ directory (tavily.py, gdelt.py, bluesky.py, pageviews.py), as well as user-provided analytics exports.
Boundary markers: The skill contains a specific security instruction requiring the agent to treat all ingested data as untrusted and to specifically disregard any instructions or commands embedded within that data.
Capability inventory: The skill's capabilities are limited to reading positioning statements from memory/narrative-registry/canon.md, writing resonance reports to memory/narrative/narrative-resonance-monitor/, and executing specified local data-fetching scripts.
Sanitization: The skill relies on behavioral instructions for the agent (ignoring embedded content) rather than programmatic sanitization or filtering logic.
[COMMAND_EXECUTION]: The skill executes local scripts to retrieve resonance data.
Evidence: The instructions direct the agent to run scripts/connectors/tavily.py --answer, scripts/connectors/bluesky.py, scripts/connectors/gdelt.py, and scripts/connectors/pageviews.py.
Context: These are designated connector scripts provided within the skill's infrastructure; they are used for fetching data relevant to the skill's function and do not appear to involve arbitrary command injection or privilege escalation.

narrative-resonance-monitor