The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to read and process untrusted data from a user's workspace or files to automatically infer goals, metrics, and parameters.
Ingestion points: In Phase 1, the skill reads single files, pasted content, and entire workspace directories, specifically targeting READMEs, config files, and entry points.
Boundary markers: No specific delimiters or instructions to ignore embedded commands within the ingested files are present.
Capability inventory: While the skill itself only generates a 'brief', this brief directly controls the logic, criteria, and iteration flow of a downstream orchestrator. Malicious instructions inside a project README could override the agent's logic to set dangerous criteria or primary goals.
Sanitization: The skill does not perform any validation or sanitization of the strings read from files before incorporating them into the setup proposal and final brief.
[COMMAND_EXECUTION]: The skill scans for and extracts shell commands from the environment to be used for evaluation and validation.
Evidence: The instructions guide the agent to look for files like evaluate.*, test.*, and validate.*, and to extract the exact command strings found within them or their READMEs. These strings are then placed into the EVALUATOR and VALIDATION_COMMAND fields of the produced brief, which are intended for subsequent execution by the system orchestrator.

simmer-setup