The Agent Skills Directory

[PROMPT_INJECTION]: The skill introduces a surface for indirect prompt injection by having the agent store and later retrieve task outcomes and error logs.
Ingestion points: The agent reads stored data from the .claude/learnings/ directory using scripts like pre-task.sh and commands during the 'Weekly Synthesis Loop'.
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands are present when the agent retrieves past logs.
Capability inventory: The skill uses local shell operations (grep, cat, mkdir, echo) to manage the knowledge base.
Sanitization: There is no sanitization of logged content. Malicious instructions contained within external error messages or task data could be retrieved and followed by the agent in future sessions.
[COMMAND_EXECUTION]: The shell scripts provided for task hooks (post-task.sh, pre-task.sh) perform file and search operations without input validation.
Evidence: The scripts directly interpolate arguments into shell commands (e.g., grep -r "$1" and echo "... $1" >> ...). If the input strings contain shell-special characters or flags, it could result in argument injection or unintended shell behavior.
[EXTERNAL_DOWNLOADS]: The references/learning-resources.md file contains a curated list of links to official documentation (e.g., react.dev, docs.python.org) and well-known learning platforms.

continuous-learning-agent