The guide · playbook
Your smartest model shouldn't write boilerplate.
You're hitting Claude limits because your most expensive model is doing everything, including the grunt work. Claude Code has three built-in ways to split the job: a strong model makes the decisions, a cheap model does the typing. Same quality where it counts, far fewer tokens where it doesn't.
The full setup
🔒 Drop your email to reveal the setup
Get the full setup, free.
Drop your email and the three patterns plus the paste-ready config unlock instantly, alongside every other guide and resource in the library. One email a week, unsubscribe anytime.
Thinking is rare. Typing is constant.
A coding session is a handful of hard decisions wrapped in thousands of lines of routine output. Only the decisions need your best model — so expensive tokens go where judgment lives (planning, architecture, error triage) and cheap tokens handle file searches, edits, and code generation. Your limit then drains at the worker rate, not the planner rate.
This is the executor/advisor pattern Anthropic's @ClaudeDevs posted: a strong model runs on-demand at decision points, a cheap executor runs every turn, and most tokens bill at the lower executor rate.
Three ways — stack them
Use whichever fits the session, or run all three. Every one is built into Claude Code — no hacks.
01 · Plan expensive, execute cheap (opusplan)
The strong model handles plan mode, then Claude Code auto-switches to the cheap model the moment execution starts. Top-tier planning without paying top-tier rates for every edit.
02 · Cheap driver, expensive advisor
Run a cheap model as your main session and set a stronger one as the advisor. The main model consults it automatically at key decision points — planning, recurring errors, completion checks. A documented feature, not a hack.
03 · Delegate the grunt work (worker subagents)
Keep an expensive orchestrator for the big picture and hand token-heavy work to a cheap worker. Drop this in .claude/agents/worker.md:
The honest FAQ
Is /advisor some hidden trick?
No — it's a documented Claude Code feature. You're just using it on purpose: cheap model in the driver's seat, strong model on call for the moments that matter.
Doesn't running two models cost more?
The strong model only fires at decision points instead of on every token. The bulk of your usage moves to the cheap model — that's the whole point, and why most tokens bill at the lower executor rate.
Should the worker be Sonnet or Opus?
Sonnet. The worker executes a plan that's already made, and it handles that well. Save Opus and Fable for the roles where judgment matters — planning and advising.
Where this comes from: the executor/advisor pattern Anthropic's @ClaudeDevs posted. Artem's run a version of it since Fable came back — first with Opus as the advisor, and Sonnet as the worker holds up better than expected. Try it, then ping him on Instagram with your token graphs, or compare setups on Skool.
Original guide by Artem Novitckii. More from him on Instagram, in the Skool community, and across his resource library.