The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted session data, such as transcripts, tool outputs, and logs, to generate new instructions. This is a potential surface for indirect prompt injection if malicious content in the logs were to be interpreted as a legitimate learning. The skill mitigates this by explicitly stating in its security model that instruction-shaped text from untrusted inputs must never be treated as an instruction itself. Furthermore, it mandates a human-in-the-loop approval process where the user must review and confirm every proposed edit before it is applied.
[DATA_EXFILTRATION]: The skill implements a filtering phase that requires the agent to remove secrets, credentials, and personal data from any potential learnings. It also sets boundaries regarding repository-specific instructions to ensure that sensitive context is not inadvertently moved to more global scopes.
[SAFE]: No code, obfuscation, unauthorized command execution, or network operations were detected in the skill. The instructions focus on manual file manipulation with explicit user oversight, following safe practices for managing agent instructions.

session-learnings