An Every Consulting Lead's Field Notes on Codex: How a Non-Technical Operator Builds and Maintains Her Own Private Work Systems
- Natalia Quintero, head of Every's consulting business, comes from a non-technical background. Over the past month or two she's shifted from Claude Code to OpenAI Codex to build and maintain several private work systems of her own.
- Her favorite move so far: give Codex a goal — clean up hundreds of customer records in her CRM based on what actually happened in calls and emails — then go to sleep. Six hours later, done.
- She built 3 custom apps in Codex: one that auto-converts her study material into comic strips she can flip through on her commute, an email triage tool with 7 action buttons that drafts replies in her own voice, and a care-coordination app for her 81-year-old father's multi-caregiver team, built in 13 hours.
- Despite being capable of building her own CRM, she deliberately walked away from her own AI-assembled system and bought Attio instead: software is the skeleton, the AI model is the brain and ligaments — and maintaining data quality is literally what a CRM company does for a living.
- A year into running Claudia, Every's internal AI employee, the team found it still needs constant human oversight and direction — so they hired an additional human operator.
What she gained by moving from Claude Code to Codex
Natalia Quintero, head of Every's consulting business, recently gave a live demo on the AI & I podcast of the private work systems she's built and maintained entirely in OpenAI Codex over the past month or two.
She used to be a heavy Claude Code user, comfortable inside a folder structure she'd built herself. Dan Shipper spent weeks nagging her to install Codex. Her verdict after switching: change the tool, and your life changes with it.
For someone who doesn't write code, two changes stood out immediately. First, Codex folds the terminal and browserthe terminal is the black command-line window that tells the machine what to do, the browser is the familiar web window — both are now embedded in the same chat interface directly into the chat interface, and paired with a model (5.5) willing to grind hard, you can feel the compute slamming into the problem. Second, the burden is lighter: with Claude Code she had to manually plan the folder structure and keep Finder open to track where things lived before the AI could work inside it; with Codex, the structure grows as it works, so she no longer has to lay the foundation in her head first.
Running through the whole conversation is a mental-model shift she and Dan kept coming back to: knowledge work is moving from "sculpting" to "gardening."
Sculpting means every outcome is carved by your own hand. Gardening means you set up the soil, light, and water, and let the crop grow itself — you're not doing every plant by hand. Applied to work: her job is no longer writing every email or updating every record herself, but building a system that produces those outcomes automatically, stepping in only at a few key points.
Give it a goal, sleep on it, come back to a finished task
Her current favorite move happened while setting up Attio, a CRM (customer relationship management software). Her team needed to fill in and correct every customer record based on what had actually happened in calls and emails.
She didn't do it record by record. She gave Codex a goal: make the CRM accurately reflect what had happened across her conversations and inbox, covering hundreds of customers and prospects, backed by a solid prompt and set of directions. Goal delivered, she went to sleep. Six hours later, work that would normally take weeks of manual effort was done.
This is the "Loop" everyone talks about but rarely explains clearly. Its structure is a human sandwich: a person judges upfront whether the task is worth doing, the middle hands an entire category of work to the AI to build a system and run it, and the person closes the loop by reviewing and refining the output, then feeding what was learned back into the system so the next round is more accurate.
A person gatekeeps both ends of the loop; the middle is entirely handed to AI to execute: 1. Start (human): judge whether the task is worth investing in — is this a goal worth pursuing 2. Middle (AI): don't process tasks one by one — build a system that can handle the entire category of task in bulk 3. End (human): review and refine what the AI produced 4. Feedback: feed what was learned this round back into the system, so the next round is more accurate