Model Router Beginner: Right Model for Every Task

The problem: most AI agents use one model at one effort for everything. Simple lookups burn your strongest model, hard architecture calls get a quick pass, and nobody checks the work before it ships.

What you get: a routing method your agent follows inside the tool you already use. It sorts every task into planning, working or cheap-model work, picks the effort to match, and builds from written checks so "done" means proven.

What your agent learns

• Three model tiers: a planning model for ambiguous decisions and architecture, a working model for scoped builds and review, a cheap model for bulk reading and formatting.
• Task classification: a decision tree covering bulk work, reading many files, live data, review, verification, planning and builds.
• Build from checks: quote the request, freeze acceptance checks, probe what tools exist, and collect one shared context file before work starts.
• Task briefs: a template that hands any worker the scope, files, non-goals, checks and what it has and lacks.
• One honest review: an independent review by a different model family when you have one, and a recorded limitation when you do not.
• 10 protocols: building, context files, acceptance checks, decision logs, renames, gap analysis, research, calculation, notes and documentation checks.

How setup works

On first run your agent asks which AI tools and models you have, copies the rules into your project, and fills in its own routing table from what it can verify. No installer, no account, no config to edit by hand.

Works inside one agent: Claude Code, Codex, Cursor, Gemini CLI or any agent that reads SKILL.md. When you add a second AI tool, the Intermediate pack adds delegation across them.