The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes the entire session history as its primary input to extract instructions and logic.
Ingestion points: The skill reviews the session history to identify successful execution steps and key findings (SKILL.md, Step 1).
Boundary markers: No explicit delimiters or instructions are provided to the agent to ignore potentially malicious embedded commands within the analyzed session data.
Capability inventory: The skill uses the Write tool to save new skills and scripts to the skills/ directory and uses chmod +x to make scripts executable (SKILL.md, Step 5).
Sanitization: The instructions do not specify any sanitization or escaping of the session content before it is interpolated into new skill files.
[COMMAND_EXECUTION]: The skill invokes shell commands to modify file permissions.
Evidence: The instructions explicitly direct the agent to run chmod +x on generated Python scripts to make them executable (SKILL.md, Step 5).
[DYNAMIC_EXECUTION]: The skill dynamically generates executable content at runtime.
Evidence: The skill determines if a script is necessary and generates Python code using a uv script template or JavaScript code for the Bun runtime. These scripts are then written to the disk and prepared for execution based on logic derived from the current session (SKILL.md, Step 3 and Step 5).

skill-extractor