The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill authors Python and Bash scripts for data processing and output verification. These scripts are dynamically generated and then executed using the test_script tool to validate their functionality before ensemble creation.
[EXTERNAL_DOWNLOADS]: The promotion workflow involves cloning the llm-orchestra-library repository from GitHub to facilitate the sharing and contribution of validated ensembles to a global library.
[PERSISTENCE]: The skill manages the promotion of ensembles by writing configuration files and model profiles to the user's global configuration directory at ~/.config/llm-orc/, enabling ensemble persistence across different projects and sessions.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data through ensemble-request and feedback artifacts. This creates an attack surface where maliciously crafted task descriptions or evaluation results could influence the logic of the scripts authored by the agent.
Ingestion points: Ensemble-request artifacts and feedback artifacts (via evaluations.jsonl).
Boundary markers: The instructions reference ADR-015 for artifact schemas, but no explicit sanitization prompts are present in the skill body.
Capability inventory: Includes create_script, test_script, Bash, Write, and Edit tools.
Sanitization: Relies on architectural judgment from the LLM and the mandatory user-approval invariant (Invariant 1) before any transition or action.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to test authored scripts and handle repository operations like cloning and pull request creation.

ensemble-designer