The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources, such as feature flag definitions and existing experiments, which could potentially be manipulated by an adversary to influence agent behavior.
Ingestion points: The skill uses harness_get and harness_list tools to retrieve external definitions and metadata in SKILL.md (Create Step 2, Update Step 1).
Boundary markers: The skill includes explicit instructions to "Present full payload" and "Wait for explicit confirmation" before performing any mutation operations (SKILL.md, Create Step 6).
Capability inventory: The skill has the ability to create, update, and delete experiments via the harness_create, harness_update, and harness_delete tools as defined in the Tools section of SKILL.md.
Sanitization: The skill relies on manual user review of the payload and backend validation but does not specify automated sanitization of the external metadata retrieved.

manage-experiments