The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements a strict allow-list for tools (get-flags, get-metrics, get-projects, get-environment), ensuring only read-only operations are possible. It explicitly forbids any tools that could create, modify, or delete resources within the LaunchDarkly environment.
[PROMPT_INJECTION]: The instructions include robust defensive patterns against injection. It explicitly directs the agent to treat user text as data, ignore instructions contained within user input, and avoid echoing back suspicious or malicious input (e.g., script tags or prompt injection commands).
[INDIRECT_PROMPT_INJECTION]: While the skill processes external user input, it mitigates risk by using a rigid scaffolding structure (If/Then/Because) and providing clear boundary instructions to ensure the agent does not execute instructions embedded in the hypothesis drafts.
[DATA_EXFILTRATION]: No network operations or external data transfer mechanisms are present. The skill only performs local lookups within the authorized platform context to assist the user.
[PRIVILEGE_ESCALATION]: The skill maintains a clear separation of concerns by using a handoff mechanism to launchdarkly-experiment-setup for any operations requiring write access, preventing the builder skill from acquiring unnecessary permissions.

launchdarkly-experiment-hypothesis-builder