Here's the exact setup.
Step 1
Set Your Top Model As The Planner
In Claude Code, run /model and pick a capable planning model available on your plan. Set /effort high for planning. This model writes the plan and reviews the final result, it does not do the grunt work.
Step 2
Create A Cheap Executor Subagent
Make a subagent that runs on a cheap model to do the implementation. Just ask Claude to create one for you, or add a file at .claude/agents/executor.md with this inside:
File — .claude/agents/executor.md
---
name: executor
description: Writes and edits from a given plan. Use for all implementation.
model: haiku # or sonnet for harder work
---
You implement exactly what the plan says. Don't redesign it.Don't skip that last line. This is the part that tripped me up: left alone, the cheap model wants to redesign your plan instead of just building it, and then your expensive model has to clean up the mess, which defeats the whole point. "Don't redesign it" keeps it in its lane.
Step 3
Run The Loop
Now your top model plans, hands the work to the executor, and reviews it. Kick it off like this:
Prompt — Plan, Delegate, Review
Write a detailed implementation plan for [X]. Don't build it yet. Then hand that plan to the executor subagent to implement. Once it's done, review the result yourself against the plan and fix anything it missed.Your expensive model only spends tokens on the two things that matter, the strategy and the quality check, while the cheap model does the heavy lifting.
Level 2 (optional): let it run on its own
Once you're comfortable, you don't have to babysit the plan-execute-review cycle. Run /loop and Claude will keep repeating it for you until the job's done, no switching back and forth by hand.
Should you even do this?
Honestly? If you're only using Claude to chat or write, you don't need this yet. This is for when you're actually building something and watching your limit drain by midweek. If that's not you, save this for when it is.
Made by @chloesinyin. Follow for more Ai content.