The Agent Skills Directory

[SAFE]: No malicious behavior or vulnerabilities detected.
[DATA_EXPOSURE]: The skill reads project-level Markdown files (CLAUDE.md, SKILL.md) and uses shell commands to measure their size. These operations are local and confined to the project directory. No network operations or access to sensitive global paths (like .ssh or .aws) were found.
[COMMAND_EXECUTION]: Uses common shell utilities (wc, awk, fd, sort, head) for file analysis. These are used in a limited, non-dangerous way for measurement and file discovery within the project root.
[PROMPT_INJECTION]: The instructions are task-oriented and include "Hard Rules" to prevent misuse, such as refusing to edit files without user confirmation or modifying external skill frontmatter (SKILL.md).
[INDIRECT_PROMPT_INJECTION]: As an optimization tool, the skill processes content from project files (CLAUDE.md). While this represents a theoretical ingestion surface for untrusted data, the skill's logic is focused on structural analysis (measuring lengths, identifying patterns) rather than executing content as instructions. Safety is reinforced by mandatory user diff reviews before any changes are applied.
Ingestion points: Reads CLAUDE.md, .claude/rules/*.md, and skills/*/SKILL.md (metadata only).
Boundary markers: None explicitly defined for file content, but processing is limited to structural analysis.
Capability inventory: Uses Bash for measurement, and the standard Edit and Read tools for file manipulation.
Sanitization: Uses lexical overlap checks and size measurement; requires manual user approval for all edits.

optimize-claude-md