The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted security findings from external scanners and interpolates this content into instructions for verifier and ranker subagents.
Ingestion points: The skill accepts JSON, Markdown, or directories of findings via the findings-path argument.
Boundary markers: Verifier prompts use explicit section headers (e.g., "FINDING UNDER REVIEW") to separate the scanner's claims from the analysis logic.
Capability inventory: The skill utilizes Read, Glob, Grep, Write, and restricted Bash commands, along with the ability to spawn subagents via Task.
Sanitization: It implements a "Write-to-file then read" pattern (--from <payload-file>) for the checkpoint helper to prevent target-derived strings from causing shell injection.
[COMMAND_EXECUTION]: The skill uses Bash to execute several local utility commands.
Allowed commands include git log, jq, find, ls, wc, and a Python helper script located at .claude/skills/_lib/checkpoint.py.
The skill documentation explicitly instructs against using shell heredocs or stdin for repository-derived data to mitigate command injection risks.
[EXTERNAL_DOWNLOADS]: The documentation references external security guidelines.
Points to CVSS best practices documentation hosted in a public GitHub repository belonging to the Anthropic organization.

triage