The Agent Skills Directory

[SAFE]: The skill consists entirely of instructional markdown and reference files. No malicious scripts, hardcoded credentials, or unauthorized network operations were detected.
[INDIRECT_PROMPT_INJECTION]: The skill features a high surface area for processing untrusted data, as it instructs the agent to mine public forums, social media, and web pages for demand signals.
Ingestion points: Untrusted external data enters the agent context through the use of tools like social-fetch, Firecrawl, and Browserbase described in references/demand-signals.md and references/data-sources.md.
Boundary markers: The instructions do not specify technical boundary markers for the ingested data; however, the framework mandates rigorous cross-verification and human-in-the-loop scoring (Phase 3 and Phase 4 in SKILL.md).
Capability inventory: Capabilities are restricted to generating lead tables, CSV files, and outreach drafts. No instructions for executing subprocesses, modifying system files, or performing network exfiltration of the processed data are present.
Sanitization: The skill relies on qualitative scoring (e.g., the Demand-fit rubric) and manual review rather than automated sanitization to handle untrusted input.
[COMMAND_EXECUTION]: The skill mentions the use of a local CLI tool (node tools/clis/github-prospects.js) for GitHub-based research. This is documented as an internal tool for developer-intent prospecting and does not represent an arbitrary or malicious command execution vulnerability.

prospecting