The Agent Skills Directory

[COMMAND_EXECUTION]: Employs curl for network requests and python3 -c for processing JSON, URL encoding, and performing sample size calculations. These operations are essential for the skill's logic.
[DATA_EXFILTRATION]: Transmits experiment details and authentication tokens to a configurable remote endpoint ($PEPPER_CLOUD_URL). This behavior is consistent with the skill's purpose of persisting experiment state, but involves a non-whitelisted domain.
[PROMPT_INJECTION]: Processes data from bets/bets.md and user-provided inputs by interpolating them into a prompt for another AI agent. This creates a surface for indirect prompt injection where malicious content in the input file could influence the generated specification. This risk is managed through a mandatory user review and approval gate before the data is written to the remote state.
Ingestion points: Reads project data from bets/bets.md and active experiments from experiments/ paths.
Boundary markers: Missing; variables are interpolated directly into the drafting prompt.
Capability inventory: Ability to read/write state via curl and execute shell commands.
Sanitization: Uses json.dumps for API payload integrity, but input variables in the drafting prompt are not sanitized for injection patterns.

experiment-design