The Agent Skills Directory

[PERSISTENCE]: The skill instructions guide the user to install persistent hooks in the agent's configuration file (~/.claude/settings.json).
The PostToolUse hook executes track-tools.sh after every tool invocation to record activity.
The Stop hook executes evaluate.sh at the end of every session to analyze tool usage and potentially trigger suggestions.
[COMMAND_EXECUTION]: The skill's primary logic and monitoring functions are implemented through shell scripts and manual bash commands.
The skill uses Bash to check for existing skills, status toggles, and to verify file creation.
The included track-tools.sh and evaluate.sh scripts use standard Unix utilities (jq, awk, sed, sort, uniq) to process session metadata.
[INDIRECT_PROMPT_INJECTION]: As a tool that generates new skills from session history, it is susceptible to indirect prompt injection if a session contains malicious instructions designed to be captured into a new skill.
Ingestion points: Evaluates the full conversation history and tool outputs from the current session (SKILL.md Step 1).
Boundary markers: The skill relies on a human-in-the-loop confirmation step (Step 3) to review and adjust the proposed skill before it is written to disk.
Capability inventory: The skill has capabilities to write new files and execute shell commands across the project and home directories.
Sanitization: The track-tools.sh script performs basic character normalization on skill names (replacing spaces and colons), but broader sanitization of the generated skill body relies on the LLM's adherence to the provided templates.

auto-skill