The Agent Skills Directory

[METADATA_POISONING]: The skill frontmatter lists the author as 'Claude Code', suggesting it is an official part of the agent platform. However, the documentation links to a third-party repository ('github.com/blader/...') and the skill metadata differs from the provided author context ('ckorhonen').\n- [PERSISTENCE]: The installation instructions in 'DOCS.md' guide the user to add a 'UserPromptSubmit' hook to the global '~/.claude/settings.json' file. This hook executes the 'continuous-learning-activator.sh' bash script on every prompt, ensuring the skill's logic and monitoring persist across all agent interactions.\n- [DYNAMIC_EXECUTION]: The skill is programmed to autonomously generate and save new skills that include executable 'scripts/' subdirectories. This pattern creates a mechanism where the agent writes and subsequently executes its own generated code.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes arbitrary session history and terminal output to generate new system instructions (skills), creating a vulnerability surface for instruction poisoning.\n
Ingestion points: Session history, terminal output, and task results analyzed during the extraction and retrospective processes.\n
Boundary markers: The skill requires a 'dry-run' confirmation from the user before writing, but lacks technical delimiters to isolate ingested data during the generation phase.\n
Capability inventory: The skill utilizes 'Read', 'Write', 'Edit', 'web_search', and 'web_fetch' tools to manage the skill library.\n
Sanitization: Instructions include a policy to avoid sensitive data extraction, but there is no automated sanitization for content processed during the learning phase.\n- [COMMAND_EXECUTION]: The setup process involves manual execution of shell commands ('chmod +x', 'cp') and the registration of a bash script hook in the configuration file.

continuous-learning