How a non-technical PM builds real products using Cursor and AI

Executive overview

Most non-technical people assume shipping code requires engineering skills. It doesn't — it requires the right workflow. Zevi Arnovitz, a PM at Meta with zero technical background, ships production apps solo using Cursor, Claude Code, and a set of reusable slash commands he designed himself.

The framework moves through six phases: capture the idea, explore it, plan it, execute it, review it with multiple models, then update documentation. The biggest unlock is treating AI not as an autocomplete tool but as a CTO that challenges your thinking and owns technical decisions.

Getting started without a technical background

• Start with a ChatGPT or Claude project — familiar chat UI, no code in sight
• Use the project to create a virtual CTO: prompt it to challenge you, push back, and own technical decisions
• Graduate from Bolt/Lovable to Cursor once you outgrow the opinionated scaffolding
• Treat code exposure as therapy — move gradually from chat to light mode to terminal
• The models underneath Bolt, Lovable, and Cursor are often the same; the difference is how much control you have over decisions

The six-phase workflow

Zevi's full development loop, triggered via slash commands in Cursor:

1. /create-issue — captures a bug or feature idea fast, creates a Linear ticket via MCP, minimal interruption to current work
2. /exploration-phase — pulls the Linear ticket, reads the codebase, asks clarifying questions before touching any code
3. /create-plan — produces a markdown plan with status trackers, critical decisions, and broken-down tasks
4. /execute-plan — runs the plan; use Cursor Composer for speed, split front-end to Gemini if UI-heavy
5. /review — Claude reviews its own code; run the same review with Codex and Cursor Composer in parallel
6. /peer-review — Claude acts as dev lead, receives the other models' reviews, and either explains why the findings are wrong or fixes them
7. /update-docs — updates markdown documentation so future agents have better context

Planning before coding

• Bolt and Lovable jump straight to writing code — fun early on, dangerous for complex features
• Anything touching payments or database changes needs a plan first; skipping this causes gnarly bugs
• The exploration phase is the most important — Claude reads file structure, understands current state, asks questions a real engineering manager would ask
• The markdown plan becomes a shared artefact: other models can read it, work from it, and reference it later

Multi-model code review

• Claude (Claude Code): communicative, opinionated, collaborative — the ideal dev lead
• Codex: non-communicative but solves the hardest bugs; treat it like the engineer in a hoodie in a dark room
• Gemini: exceptional at UI/design; chaotic in its approach but the output is often beautiful
• Running /peer-review forces Claude to either justify its decisions or fix issues raised by the other models — models will argue back if they've been challenged three times already
• Use /learning-opportunity during review to get 80/20 explanations of anything you don't understand

Keeping the workflow sharp over time

• When Claude makes a mistake, ask it what in its system prompt or tooling caused the error
• Update the CLAUDE.md file or slash commands based on the answer — the same mistake shouldn't recur
• Iterating on prompts after failures is one of the clearest divides between average and skilled AI users
• Good documentation inside the codebase (plain-text markdown files explaining how things work) helps agents navigate and write better code on the next feature

Using AI to learn, not just to ship

• /learning-opportunity primes Claude as a technical teacher — explains what's being built using the 80/20 rule
• Building side projects is a way to get reps on decisions (strategy, marketing, architecture) that junior PMs rarely touch in large companies
• The concern that AI weakens PM skills misunderstands the role — PMs should harness everything available to get to the right solution faster
• AI output quality scales directly with the context and constraints you give it; vague prompts produce vague results

On the current state of vibe coding tools

• Bolt, Lovable, Replit, Base44, v0: excellent for getting started, handle auth and database setup automatically, but limit your control
• Cursor + Claude Code: full control, all the hard decisions surface explicitly, requires more judgment
• Neither is wrong — the right tool depends on how much ownership you want over technical decisions
• Making a codebase AI-native (well-structured markdown docs, clear file organisation) is a prerequisite before PMs at larger companies can contribute meaningfully

Using AI for job interviews

• Created a Claude project as an interview coach, loaded it with the best PM interview frameworks available
• Built a quiz app in Base44 to practice segmentation questions during commutes
• Used AI for mock interviews with feedback, then used real humans for the final preparation — human mocks remain irreplaceable for competitive roles
• Had Claude play the candidate on questions where time ran short, learning from model answers