The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to execute watch-skill commands for adding lessons, listing them, and viewing usage statistics. These operations are performed locally using the user's CLI.
[INDIRECT_PROMPT_INJECTION]: The skill implements a feedback loop where user-provided corrections are stored as lessons and automatically injected into future prompts for similar questions. This creates a vulnerability surface where a malicious correction could influence the agent's future behavior.
Ingestion points: User-provided inputs for <question>, <the wrong answer>, and <the correction> are captured via the watch-skill lessons add command in SKILL.md.
Boundary markers: None are specified for the retrieval and injection process of these lessons.
Capability inventory: The skill has access to Bash and Read tools, which could be abused if an injected instruction triggers command execution.
Sanitization: The skill does not mention any sanitization or validation of the correction text before it is stored or re-injected into the model context.

learning-from-mistakes