The Agent Skills Directory

[COMMAND_EXECUTION]: The skill references the use of the caveman cloud CLI tool to retrieve experiment data as a fallback when the primary MCP interface is unavailable. These commands are limited to read-only retrieval of lists, experiment details, and results.
[INDIRECT_PROMPT_INJECTION]: The skill processes experiment data and results which could potentially contain untrusted content. However, the skill implements robust internal logic to mitigate this risk by requiring specific evidence verification and strictly forbidding the agent from performing autonomous mutations.
Ingestion points: Experiment metadata and results fetched via caveman_experiment_get and caveman cloud experiments CLI commands (SKILL.md, Step 1).
Boundary markers: Instructions mandate reporting specific fields (Step 2) and fail closed if required data is absent.
Capability inventory: Reading experiment data via MCP/CLI and generating recommendations for experiment lifecycle transitions.
Sanitization: The instructions explicitly block the emission or execution of lifecycle commands (Step 4) and require re-reading the server state after any external operator action to prevent inferring success from intent alone (Step 5).

caveman-manage