The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a pattern where agents read project-local configuration files (.claude/*.local.md) to adapt behavior. This creates a surface for indirect instructions if the files are compromised.
Ingestion points: .claude/plugin-name.local.md files located within the project root.
Boundary markers: The skill utilizes standard YAML frontmatter delimiters (---) and provides logic to separate configuration from the markdown body.
Capability inventory: The skill uses Bash, Read, and Write tools. Included examples demonstrate the ability to notify other terminal sessions via tmux.
Sanitization: Implementation examples explicitly include path traversal checks (blocking ..), sensitive file protection (blocking .env and secret keywords), and recommendations for safe JSON construction using jq.
[COMMAND_EXECUTION]: The skill provides bash scripts for parsing and validating local files. These scripts use standard unix utilities (sed, grep, awk) and include defensive programming patterns such as set -euo pipefail and atomic file updates using temporary files to prevent corruption.

plugin-settings