The Agent Skills Directory

[COMMAND_EXECUTION]: The skill integrates with a local CLI tool named caveman to perform token analysis, simulation, and memory operations. Commands such as caveman learn report, caveman mem remember, and caveman learn apply are central to the skill's functionality.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from local repository files (e.g., CLAUDE.md, AGENTS.md) and JSON reports generated by the caveman CLI. This creates a surface where malicious instructions embedded in the analyzed content could influence the agent's behavior during the optimization process. The skill mitigates this by requiring explicit user consent for every change and verifying content hashes before processing.
[DYNAMIC_EXECUTION]: The skill can generate and write new instruction files (skills) to the filesystem based on observed repetitive tool usage patterns. This automated code generation is consent-gated and monitored by a holdout experiment harness to verify efficacy.

caveman-learn