The Agent Skills Directory

[COMMAND_EXECUTION]: The skill defines a shell variable ENGINE pointing to a local script at ~/.claude/scripts/learnings-engine.mjs and instructs the agent to execute various subcommands (capture, score, lint, promote, etc.) through the Node.js runtime.
[PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by processing historical session data and user corrections to influence future agent context.
Ingestion points: Untrusted or processed data enters the context via core/learnings/episodes.jsonl, which records errors and corrections from the user or session experience.
Boundary markers: The skill does not specify any delimiters or safety warnings for the top-scored learnings injected during the SessionStart hook.
Capability inventory: The agent has the authority to autonomously modify SKILL.md, TOOLS.md, and LEARNINGS.md (Green Zone) and can execute shell commands via the learnings-engine.mjs script.
Sanitization: No explicit sanitization or validation of the correction text is mentioned; the system relies on a scoring threshold (Recency, Frequency, Impact) to promote content.
[SAFE]: The skill implements a privilege model for its own configuration, designating sensitive files like rules.md and CLAUDE.md as 'Red Zone' (Owner only), preventing the agent from autonomously modifying core safety or system instructions.

learnings