The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection by interpolating untrusted external data into agent-generated content.
Ingestion points: The skill fetches lead signals such as {comment_snippet} and {job_posting_detail} from Supabase tables or user-provided CSV files, as detailed in the Phase 1 Lead Selection and Phase 3 Message Generation sections of SKILL.md.
Boundary markers: The sequence templates in templates/sequence-templates/ do not utilize explicit delimiters (e.g., XML tags or triple backticks) or specific 'ignore' instructions to isolate the lead-sourced data from the generation prompt.
Capability inventory: The skill has permissions to write CSV files to the local output/ directory, perform network operations via the Supabase client to log outreach attempts, and optionally interact with Google Sheets for reporting.
Sanitization: To mitigate risks, the skill includes a 'Calibration Loop' in Phase 3 for initial review of sample messages and a mandatory 'Hard Approval Gate' in Phase 5 that requires explicit human confirmation before finalizing the campaign.
[COMMAND_EXECUTION]: The documentation references the execution of local scripts for environment setup.
Evidence: SKILL.md provides instructions for running python3 tools/supabase/setup_database.py to prepare the database schema for the skill's operations.

linkedin-outreach