The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates a feedback loop where untrusted user input directly informs future agent instructions, creating a risk of indirect prompt injection.
Ingestion points: User feedback strings are accepted through the --feedback argument in scripts/self-improvement.mjs and scripts/agent-evolver.mjs.
Boundary markers: There are no boundary markers or delimiters provided to the agent to distinguish between evaluative feedback and malicious instructions embedded in the data.
Capability inventory: The skill's primary purpose is to generate "Suggested Actions" that are intended to auto-update the agent's system prompt or behavior instructions. This gives processed data a path to influence privileged agent configuration.
Sanitization: The scripts perform only basic string truncation and case normalization, offering no protection against prompt injection payloads hidden within user feedback.

self-improvement