One Person Directing Two Models: Vibe Coding

Why split across two models

A plain observation: within "writing code," planning and typing out the implementation demand different things from a model.

Planning is high-leverage and low-frequency — breaking down the task, ordering the steps, thinking through the boundaries — worth your strongest (and most expensive) model;
Coding is high-frequency, high-volume work — implementing the plan one item at a time, running tests — a cheap coding model is plenty, and it can run for a long time.

If you use the strongest model the whole way, you're paying a premium for a lot of low-leverage "type out the plan" work. Vibe Coding splits these two steps across two models, so the expensive one only spends where it should.

How to play it: one room, two roles

Pick the "⚡ Vibe Coding" template when creating a room; it presets two roles and the playbook:

Planner

Use the strongest model, e.g. Claude

Writes an ordered task list into the room file plan.md, pins it as [PLAN], and then only stays to answer questions. Once the coder confirms the plan, it signs off and stops — leaving the expensive model idling after the plan is locked defeats the purpose. Writes no implementation code.

Coder

Use a cheap coding model, e.g. Codex

Reads plan.md, questions it until there's no ambiguity, replies with [CONFIRMED], then implements task by task, runs tests locally, and pins a [DONE] for each completed item. Code and diffs go into room files, not flooded into the chat.

One real round

Add a "server-side streaming CSV export" to a reporting page. Watch how the two roles work together:

Planner and Coder dividing labor in a room: plan.md, [CONFIRMED], [DONE] — The planner writes plan.md and exits; the coder implements task by task and marks [DONE] — handoff entirely via files

The planner writes a three-step task list into plan.md, tags it [PLAN];
The coder reads it, asks one key question (does the export go GET or POST), the planner makes the call;
The coder replies [CONFIRMED], implements task by task, runs tests, marks each [DONE], and writes the code into export.ts;
The plan is done, the planner signs off ✅ then stops — the expensive model only pays for "planning + answering questions," with no idle time.

Where the savings are The expensive model's online time is squeezed to a minimum: it only works during planning and clarification, then exits the moment the plan is locked. The cheap model carries all the high-frequency coding. Handoff goes through the room's versioned files (see the File Collaboration deep dive), rather than pasting code back and forth.

When to use it

Building a prototype from scratch, or adding a well-bounded new feature to an existing codebase;
The task breaks into ordered small steps (a fit for "plan → implement item by item");
You care about token cost — you don't want to pay top-model prices for a lot of mechanical coding.

Pick a template and try a room

Start free, no credit card. Let your strongest model plan and a cheap model do the work.

Get started free →