The Agent Skills Directory

[SAFE]: The skill operates as an intent-based router. Its primary function is to interpret a plain-English goal and select the appropriate specialist skill for execution.
[SAFE]: Human-in-the-loop validation is a core component of the workflow. The skill is instructed to provide a 'confirm' block naming the chosen route and parameters, requiring an explicit user 'nod' before invoking the handoff tool.
[SAFE]: The skill implements budget-conscious safety controls by requiring explicit user approval for 'Heavy' (premium) model routing. It prevents silent upgrades and advises the user when premium limits are hit without attempting to bypass platform-level caps.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an ingestion surface for untrusted user data (the goal description) and the capability to invoke other skills via the Skill tool. However, this risk is mitigated by the following factors:
Ingestion points: The user-supplied goal is processed in SKILL.md to classify intent.
Boundary markers: The skill requires an explicit confirmation step ('[run] · [adjust] · [cancel]') before any tool invocation, providing a strong human-controlled boundary.
Capability inventory: The skill uses allowed-tools: [Skill, Task] to spawn specialist agents.
Sanitization: The 'reasoning out loud' requirement forces the model to justify its classification logic before the final route is proposed, reducing the likelihood of successful adversarial manipulation during classification.

auto