LLenny's Podcast
← All frameworks
StrategyZevi Arnovitz (Meta)

Model Casting by Personality

Route each task to the model whose 'personality' matches, and split plans to fit

Difficulty
Moderate
Time to result
~days to results
Steps
4
Confidence
83%

Treat each AI model as a distinct coworker with strengths and weaknesses, then assign work to play to strengths and mitigate weaknesses — including splitting a plan into backend and frontend so each half goes to the best-suited model. Use it once you're working across multiple models (Cursor's Composer, Gemini, Claude, Codex) and want the best result per task rather than one model for everything.

Origin

Zevy finds the easiest way to reason about AI is to imagine each model as a person: Claude as the communicative opinionated CTO, Codex as the antisocial genius who solves the worst bugs, Gemini as the terrifying-but-brilliant designer. He routes work accordingly.

Core principles

  • 01Every model has distinct characteristics; match the task to the model
  • 02Use one model's strength to cover another's weakness
  • 03A markdown plan split by concern lets you route halves to different models
  • 04Judge by output, not by how scary the process looks

How to run it

  1. 1

    Write a plan you can split

    Produce the plan as a markdown file structured so backend and frontend work are separable.

  2. 2

    Route frontend to the design-strong model

    Have Gemini read the plan and do the frontend/UI, which it excels at.

    Watch out Gemini's visible thought process can be alarming (e.g. proposing to delete the dashboard or edit the database) — judge the final result, not the process.

  3. 3

    Route fast simple work to Composer

    Use Cursor's Composer for non-complex execution because it's blazing fast and keeps you in flow.

  4. 4

    Reserve the specialist for hard bugs

    Bring in Codex for the worst bugs — the 'closes the door for two hours and comes out having fixed it' specialist.

    Pro tip Keep Claude as the communicative lead/CTO for collaboration, learning, and adjudication.

In the wild

Backend/frontend split

Zevy splits a plan into backend and frontend, then has Gemini 3 read the plan and build the frontend because 'Gemini 3 that just came out is unbelievable at UI.'

Each half of the feature is built by the model best suited to it.

Common mistakes

Judging a model by its process

Gemini's chaotic visible reasoning looks fireable, but the finished design is beautiful — reacting to the process instead of the output leads to using the wrong model.

Is it for you?

Best for

Multi-model builders optimizing quality per task across Cursor, Claude, Codex, and Gemini

Not ideal for

Beginners still on one tool who'd add complexity by juggling models too early

From the transcript

a lot of times I'll split the plan into backend and front end and then I'll have Gemini just read the plan and do the…

30:30

Gemini 3 that just came out is unbelievable at UI

30:00

using all these models um and basically playing to their strengths and mitigating their weaknesses by using other models is is a game changer for…

42:30

From the episode

The non-technical PM’s guide to building with Cursor

Zevi Arnovitz (Meta)