The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local scripts provided with the skill package, including scripts/setup-labels.sh (a bash script for GitHub label configuration) and scripts/validate_spec.py (a Python script for markdown schema validation). It also leverages the GitHub CLI (gh) to manage issues and retrieve context.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from various external sources which could contain malicious instructions.
Ingestion points: External data is ingested via the WebFetch tool for arbitrary URLs and gh issue view for GitHub issues in Phase 1 (Hunt).
Boundary markers: The skill explicitly attempts to mitigate injection by instructing the agent to wrap external context in XML tags (e.g., <brief_and_evidence>) when passing data to subagents, as seen in references/subagent-prompts.md.
Capability inventory: The skill has the capability to write files (mkdir and save markdown artifacts), execute local shell/python scripts, and perform network requests through available tools.
Sanitization: The workflow includes reframing the problem and manual verification steps, though it relies on LLM-based parsing of the untrusted data rather than strict programmatic escaping.

forge