The Agent Skills Directory

Indirect Prompt Injection Surface: The skill processes external data from Claude Code session transcripts and repository metadata. To mitigate the risk of embedded instructions, the skill includes explicit directions to treat this content exclusively as data and never as instructions.
Validation Contract (Scout Output Contract): The skill implements a rigorous validation layer for all data returned by its sub-agent. This contract enforces strict regex patterns for counts, file paths, IDs, and links, while also checking for path safety by blocking directory traversal segments like '..' and imposing character length limits.
Controlled Sensitive Data Access: The skill reads local session transcripts from the ~/.claude/projects/ directory. This access is guarded by a consent workflow that requires the user to approve the specific files or folders before they are read.
Command Execution Guardrails: A proactive safety rule is defined to prevent untrusted data from being used in shell commands. The skill explicitly forbids passing any values extracted from transcripts or repository findings into the command line, effectively neutralizing potential command injection vectors.
Exfiltration Controls: The skill is restricted from automatically sending or posting data externally. It is designed to provide text for the user to manually review and paste, with a single, highly specific exception for calendar events that requires explicit approval of the exact content.

climb