The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands for data processing and documentation gardening. Evidence includes the use of make learning-corpus, uv run scripts/learning_corpus.py, and make doc-garden. These commands interact with the local filesystem and execute scripts defined in the user's environment.
[PROMPT_INJECTION]: The skill design creates an attack surface for indirect prompt injection (Category 8) by ingesting unstructured data to drive agent logic. Ingestion points: The skill processes LEARNING.md files from multiple repositories under the ~/Development path and governance documentation files such as CLAUDE.md.example and README. Boundary markers: The instructions lack explicit delimiters or instructions for the agent to ignore potentially malicious embedded content within the processed files. Capability inventory: The agent is empowered to propose significant 'harness changes' including pre-commit hooks, system rules, and skill updates based on its analysis of the ingested data. Sanitization: There is no mention of sanitizing or validating the content extracted from external repositories before it is used to influence the agent's proposals.

learning-loop