The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates and executes multiple bash scripts in the ~/.claude/hooks/ directory to automate developer tasks such as formatting, type checking, and reflection reminders.
[REMOTE_CODE_EXECUTION]: The skill employs dynamic script generation and execution by creating local .sh files and configuring the agent's ~/.claude/settings.json to run these scripts automatically via SessionStart, PostToolUse, and SessionEnd hooks.
[PROMPT_INJECTION]: An indirect prompt injection surface is created by the load-mistakes.sh hook, which reads content from .claude/mistakes.md and injects it directly into the agent's context as a systemMessage at the start of every session without sanitization.
Ingestion points: .claude/mistakes.md (read by ~/.claude/hooks/load-mistakes.sh).
Boundary markers: Absent; the extracted lesson content is interpolated directly into a JSON systemMessage block.
Capability inventory: The agent has full access to the shell (Bash), file system (Read, Edit), and version control (git).
Sanitization: Content is escaped for JSON structure using jq, but no validation or filtering is performed on the actual text to prevent embedded instructions from influencing the agent.

workflow-optimization