The Agent Skills Directory

[PROMPT_INJECTION]: The skill evaluates historical interaction data to suggest improvements to its configuration, which introduces a surface for indirect prompt injection. If past sessions contain malicious instructions, they could potentially influence the suggested updates. * Ingestion points: Historical session data and logs processed in references/retrospective-prompt.md. * Boundary markers: There are no explicit markers used to separate instructions from historical data during analysis. * Capability inventory: The skill has the capability to propose changes to the agent's core instructions (AGENTS.md) and create new skill files. * Sanitization: No content sanitization is performed on the historical data being reviewed. * Mitigation: Significant mitigation is provided by the Human Gate policy, which requires manual user approval for all proposed modifications.

codex-retrospective