When you give your agent a complex task, it does everything with the same expensive model. Reformatting 10 files? Opus. Generating boilerplate from a spec? Opus. Running a script and checking output? Still Opus. Meanwhile, most of those subtasks could be handled by Haiku at a fraction of the cost.

Smart Router does two things.

First, it automatically identifies which parts of a task are mechanical or repetitive and delegates them to a cheaper model. Creating a batch of similar files, converting formats, running validation checks, generating boilerplate. The skill writes fully self contained prompts so the subagent has everything it needs without any conversation context. Your main model focuses on the work that actually requires judgment: architecture, debugging, the final user facing deliverable.

Second, it handles context handoffs when you switch models or start a new session. It writes a structured HANDOFF.md snapshot that captures not just what was done, but why. Key decisions and their rationale, rejected approaches and why they didn't work, specific next steps, and any user preferences that came up in conversation. The next model reads that file and picks up exactly where you left off without asking you to repeat yourself.

You can use either feature independently or both together. On a big task, you might delegate boilerplate to Haiku, do the complex work yourself, and then write a handoff when the user wants to continue later with a different model.

Works with Claude Code, Cursor, Codex CLI, VS Code Copilot, and any agent that supports the SKILL.md standard.