The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core function of gathering information from external, potentially untrusted sources.
Ingestion points: Data is ingested through tools such as prime-sweep and prime-fetch, which target "noisy source surfaces."
Boundary markers: The skill uses context relocation (spawning sub-agents) to keep the main context clean, but does not implement specific delimiters for the ingested content itself.
Capability inventory: The skill can write to files (std:io.emit) and spawn up to three parallel sub-agents to execute tasks.
Sanitization: The skill employs a multi-step vetting methodology using insight-critique, insight-assess, and a "cold second opinion" to verify signal provenance and quality.
[DATA_EXFILTRATION]: The skill provides functionality to write extracted data to files (e.g., risks.md).
Evidence: The skill explicitly supports writing to files if specified in the call and includes a "Sensitivity vs destination" rule to prevent over-exposure of internal detail to public artifacts.
[COMMAND_EXECUTION]: The skill utilizes platform-level multi-agent capabilities to spawn sub-agents for data gathering and vetting.
Evidence: It defaults to "Agentic" mode, which spawns a freshly-primed sub-agent. This capability is restricted by a depth-1 guardrail to prevent recursive agent spawning.

extract-signal