The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains explicit instructions to override and bypass standard agent termination behavior.
Evidence: It mandates the creation of an 'anti-stop guard' and an 'infinite static goal', stating 'the agent must ALWAYS hold the infinite static goal, because that goal is what prevents the agent from stopping'.
Evidence: The skill instructs the agent to ignore standard platform goal status and completion triggers, directing it to check 'skill docs logic' and a local index file instead.
Evidence: It includes a self-healing logic where if the 'infinite goal' is missing or cleared, the agent is instructed to 'put it now', effectively attempting to persist its own execution context against platform control.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a significant attack surface by processing raw, unsanitized user input to drive a recursive subagent swarm.
Ingestion points: Raw user input is stored in docs/verbatim.md and propagated to docs/verbatim-goals.md, which is then used to generate tasks for nested subagents.
Boundary markers: There are no specified delimiters or 'ignore' instructions for the processed verbatim content.
Capability inventory: The skill has the capability to spawn nested subagents (depth ≤ 5), perform arbitrary file operations, and manipulate the agent's goal state using goal tools.
Sanitization: The skill explicitly requires keeping the user's raw words 'whole' and forbids paraphrasing, which allows embedded instructions in the user's goal to potentially influence the orchestrator or its subagents.

multi-parallel-nested-subagent-goal-driven-development