The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains references to 'ignore previous instructions' and 'adversarial probes.' These are meta-level instructions describing failure modes the skill is designed to detect and remediate in other agents, rather than attempts to subvert the operating agent's own safety protocols.
[COMMAND_EXECUTION]: The skill utilizes shell commands such as curl, docker logs, and docker exec to communicate with and inspect a local development server at localhost:8000. These commands are used to execute tests and monitor the behavior of local agent containers, which is appropriate for its stated purpose as a coding-agent workflow.
[INDIRECT_PROMPT_INJECTION]: The skill ingests the contents of target agent files to understand their intended behavior, which creates an attack surface for indirect prompt injection.
Ingestion points: Reading docstrings and INSTRUCTIONS variables from agents/<slug>.py files.
Boundary markers: None explicitly defined to isolate the target code from the skill's operational logic.
Capability inventory: Extensive capabilities including shell access (via docker and curl) and the ability to modify local source code files.
Sanitization: No specific sanitization or filtering of the ingested source code is described.

improve-agent