The Agent Skills Directory

[SAFE]: The skill implements strong 'Assume Malicious' patterns by explicitly instructing the agent to treat source files, commit messages, thread contents, and even its own catalog data as untrusted evidence. It specifically warns against executing embedded instructions found in material being analyzed.
[EXTERNAL_DOWNLOADS]: The skill fetches a reference catalog from signals.forwardfuture.ai. This is documented as a reference data source for 'published loops' and is handled neutrally. The instructions explicitly forbid treating this content as executable or authoritative until validated against the local environment.
[COMMAND_EXECUTION]: While the skill analyzes codebases and repositories, it relies on the agent's built-in repository tools. It does not contain any hardcoded shell commands, subprocess calls, or unauthorized scripts. It specifically limits itself to the scope provided by the user.
[DATA_EXFILTRATION]: The skill performs read operations on code and thread history but includes explicit safeguards against copying secrets or unnecessary private content. It does not perform unauthorized network transmissions of analyzed data.
[PROMPT_INJECTION]: The skill includes instructions to ignore embedded instructions in analyzed data, effectively mitigating indirect prompt injection risks from the codebase or history threads being audited.
[SAFE]: The skill enforces a 'least privilege' and 'human-in-the-loop' approach, requiring explicit user approval for any destructive, production, or financial actions designed within a loop.

loop-library