The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a feedback loop where data generated in a browser environment (stored in pending.yaml) is later processed by the agent to perform file edits via the Edit tool. This creates a surface where the agent's actions are driven by externally provided data.
Ingestion points: SKILL.md documents the agent reading pending.yaml created by the injected script to apply changes.
Boundary markers: The schema uses JSON marker blocks, but the agent's ingestion of the resulting YAML lacks explicit boundary protection against instructions embedded within data values.
Capability inventory: Read, Write, Edit, Glob, Grep, Bash (mkdir).
Sanitization: The injected script uses standard JSON.stringify for serialization, though the agent's interpretation logic depends on external skill context.
[DYNAMIC_EXECUTION]: The skill functions by injecting a comprehensive JavaScript block into existing HTML files. This script handles UI rendering, state persistence in localStorage, and interactive file writing via the browser's File System Access API.
[COMMAND_EXECUTION]: The skill uses the Bash tool to initialize project directories for session state (e.g., mkdir -p artifacts/tweaks/...).

make-tweakable