The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions define a process for injecting 'lessons' derived from untrusted user feedback into the agent's behavioral instructions (system prompt).
[PROMPT_INJECTION]: Mandatory Evidence Chain for Indirect Prompt Injection surface:
Ingestion points: The skill explicitly processes user revisions of output and direct feedback strings to extract preferences (documented in SKILL.md under 'Lesson Sources').
Boundary markers: The skill recommends using a markdown header ('## Lessons Learned') to separate injected content, which is an insufficient boundary for preventing instruction override or bypass.
Capability inventory: The agent is granted file:read and file:write permissions to persist these behavioral modifications in workspace/memory/improvement-log.jsonl.
Sanitization: There is no mechanism described for sanitizing, validating, or filtering the content of these lessons before they are used to define the agent's future behavior.

self-improve