Automatic LLM model selection for sub-agent tasks. Classifies tasks by complexity and type, then routes to the optimal model (cost vs capability). Use when s...
Security Analysis
high confidenceThe skill's instructions, requirements, and included reference docs are coherent with its stated purpose (automatic model routing); it does not request credentials, installs, or hidden endpoints and appears to be what it claims to be.
Name/description (model routing) matches the SKILL.md and reference content: tier table, classifier prompt, spawn examples and orchestration patterns are all relevant to selecting models for sub-agents. The skill does not ask for unrelated binaries, secrets, or external services beyond the implicit expectation that a model provider exists.
Runtime instructions are primarily prompting patterns and routing logic; they do not instruct reading secrets or unrelated system state. References include operational patterns (tmux, JSON files under ~/.claude, filesystem-based coordination) which are architectural examples — implementers could choose to read/write local paths if they adopt these patterns, so review any integration code you add or follow carefully.
Instruction-only skill with no install spec and no code files. Nothing will be written to disk by the skill itself; risk from install mechanism is minimal.
The skill declares no required environment variables, secrets, or config paths. It references provider models (gemini-flash, haiku, etc.), which implies your runtime will need appropriate API keys configured elsewhere — this is expected and proportional to the skill's function.
always:false and no requests to modify other skills or global agent configuration. The skill describes spawning sub-agents and recommends agent-level rules (AGENTS.md), which are normal for a router but should be applied intentionally.
Guidance
This skill appears coherent and safe as an instruction-only model router. Before installing: (1) Confirm your runtime already has the model-provider credentials and cost controls you want (the skill will encourage calls to various models and can increase token spend). (2) If you adopt any of the reference orchestration patterns, review any code you write that implements them — examples mention writing JSON files under ~/.claude and using tmux; decide whether you want that filesystem access. (3) Test routing behavior in a sandbox to verify tier→model mappings match your available providers and names (model IDs in the docs may not match your environment). (4) If you prefer to approve model choices manually, restrict autonomous invocation or require human approval when spawning expensive tiers. Overall: coherent, but audit integrations and cost controls before broad deployment.
Latest Release
v1.1.0
v1.1.0: Added task orchestration patterns (hierarchical, pipeline, fan-out), chain-of-thought techniques matched to tiers, Claude Code architecture lessons (reverse-engineered), cron/sub-agent routing rules, bootstrap vs auxiliary file guidance (progressive disclosure).
More by @globalcaos
Published by @globalcaos on ClawHub
