The Agent Skills Directory

[SAFE]: The skill consists of Markdown files providing logical guidelines for the agent. No malicious patterns such as command execution, data exfiltration, or obfuscation were detected.
[INDIRECT_PROMPT_INJECTION]: The skill describes a process for handling untrusted data from sub-agents and external tool outputs. Evidence Chain: 1. Ingestion points: Sub-agent handover files and command output tails (mentioned in acceptance.md and what-counts.md). 2. Boundary markers: Not technically specified, but the skill mandates manual verification of facts. 3. Capability inventory: None; the skill contains no scripts or subprocess calls. 4. Sanitization: None. The skill's primary purpose is to defend against deception by requiring deterministic evidence (exit codes, raw output) instead of natural language summaries, which reduces the risk of the agent being misled by poisoned inputs.

think-evidence