The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (messages and notes from contacts) from the hunter/pipeline.json file. This represents a potential attack surface, but the risk is successfully mitigated by the skill's design.
Ingestion points: Data is read from the contacts and threads sections of the hunter/pipeline.json database in the followup.py script.
Boundary markers: The skill enforces a strict human-in-the-loop (HITL) protocol; SKILL.md explicitly instructs the AI to show every draft to the user and wait for a manual 'yes' response before the user sends the message through separate connectors.
Capability inventory: The script is limited to reading/writing its own JSON data files and generating text; it has no autonomous network access or system-level write capabilities.
Sanitization: The script includes a defensive mechanism that uses regular expressions to detect 'no' or 'stop' requests in incoming messages, which automatically flags contacts for exclusion from future outreach.

hunter-followup