Transcript: How to Build a Copilot for Product Work — Zeevi Arnovitz (Wix)

Host: Eyal David · Guest: Zeevi Arnovitz · Back to episode

Zeevi Arnovitz from Wix breaks down his method for building a copilot that accompanies product work end to end. Instead of treating AI as a one-off chat, he builds a permanent 'worker' based on three pillars — knowledge, personal instructions, and context — and relates to it like a genius intern who needs a good brief so it doesn't produce generic outputs.

In this episode

A good copilot rests on three pillars: general knowledge (onboarding, like a new hire), personal instructions (custom instructions), and context specific to each project.
Without enough context, even the strongest model returns generic outputs — the real investment is in the context stage, like briefing an intern.
Setting custom instructions solves the people-pleaser problem: you ask the model for pushback instead of flowing along with every idea.
The practical architecture: one project that holds knowledge + custom instructions, and under it a separate chat for each product project.
The copilot accompanies the entire product lifecycle — from discovery, through wireframes and challenges, planning the test, all the way to shipping the product.
The tooling landscape moves fast: at Wix they build in Gemini, Zeevi personally prefers ChatGPT, and every few weeks a new model (like Claude) changes the game.

Eyal David: [00:00] So, practically — how do you build a copilot that helps with product work? Here's a short clip from the full conversation with Zeevi from Wix.

Zeevi Arnowitz: [00:08] Let's dive into it — what did you build it with? Yeah, great. So I've already tried building it in a few tools, and at Wix we build it in Gemini, personally I use ChatGPT, I really really love ChatGPT, but again, every two weeks there's a new model, now Claude's, so it really is a game changer. I'll say right up front that this isn't my idea — I learned it from someone named Tal Raviv, by the way, I really really recommend that everyone listening to this podcast follow him on Twitter and listen to all of his podcasts, he's one of the most influential voices right now, I think, in the world of how to practically use AI as a PM, and broadly, the three important pillars of the copilot, are general knowledge, okay? Which — I look at it as if you just hired someone new to your team, what comes first? All the knowledge you'd give them for onboarding, so they're plugged in. The second thing is personal instructions, which is basically all the problems we hear about — it doesn't challenge me, it just flows along with all my ideas, all of that — so yes, those are problems, but they're problems that are solvable through personal instructions. So we set up a project, okay? and we have the knowledge, which is everything I said we bring in during onboarding, we have custom instructions, which is how I want it to respond to me, don't be a people pleaser, give me pushback, anything like that, and then under that copilot we create chats, and each chat is a project I'm working on, okay? I went into discovery for some new project I want to work on as a PM, I really treat this thing like an intern working with me, and like a real intern — even if this intern is the most genius in the world, if you don't give enough context about what we're working on, and what problem we're setting out to solve, its outputs will be generic, even if this intern is super super genius, and that's why it's really important to invest in the context stage, and then this copilot accompanies me on every project I do, from the discovery level, through the wireframes and the challenges, and how we plan the test, and finally how the product ships, and that's the bulk of the sessions I'm running right now at Wix.

Eyal David: [02:17] That was a Takhles episode. From Product Builder — want to hear the full conversation? Search for it in the feed, and don't forget to follow us.