The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires a validation command as input and executes it within the environment to verify candidates during the optimization loop.
Evidence: Defined in SKILL.md under Required inputs and Core workflow step 10.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted datasets and optimization outputs, creating a surface for indirect prompt injection. It specifically instructs the agent to answer prompts returned by the optimization stage, which could contain adversarial instructions designed to bypass safety filters or execute malicious code via the validation command.
Ingestion points: The skill reads from datasets/train.jsonl, datasets/val.jsonl, and optimization artifacts like manual-followup-report.json.
Boundary markers: There are no explicit instructions to treat the ingested data as untrusted or to use boundary delimiters to prevent the agent from following embedded instructions.
Capability inventory: The skill has filesystem write access (to write candidates back to source) and shell execution capabilities (to run the validation command).
Sanitization: No sanitization or validation of the content provided by the datasets or the optimization stage is performed before it is processed by the agent.
Evidence: Documented in SKILL.md under the Manual follow-up recovery section and the Core workflow loop steps.

trainer-train