The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion and persistence of external data (user corrections, error messages, and tool outputs) into the agent's project-specific instructions and memory files.
Ingestion points: Information is gathered from user feedback in conversation and from the output of shell commands (monitored by the error-detector.sh hook).
Boundary markers: The captured data is stored in .learnings/*.md files and promoted to instruction files like CLAUDE.md or AGENTS.md. The templates used for logging do not include specific delimiters or warnings to ignore embedded instructions within the logged content.
Capability inventory: The skill performs file-write operations to log data and modify project instruction files. It also includes the extract-skill.sh script for scaffolding new skill directories.
Sanitization: There is no explicit sanitization or filtering of the content being logged. Malicious instructions embedded in a failed command's output or a deceptive user correction could be persisted and later followed by the agent when it references these recorded learnings.

self-improvement