← All episodes · Every

How Every's Head of Consulting Uses Codex Every Day

2026-07-05 · original episode

Overview

Dan Shipper sits down with Natalia, Every's head of consulting, for a follow-up conversation about how her AI-driven workflows have evolved. Their last episode introduced Claudie, the internal AI agent that runs consulting operations; this time the conversation moves from Claudie's growth (she now has her own social presence and a self-evaluating 'trust battery') to a bigger theme: when to build your own AI-glued tools versus buying professional software, and how Natalia has become dramatically more capable at building her own tools using Codex.

The heart of the episode is a screen-share tour of Natalia's actual daily Codex workflows: a personal learning skill that turns dense research topics into illustrated cartoon guides, a heavily customized email-triage app built on top of Dan's open-source 'Tend' app, a 'loop' that autonomously rebuilt her company's entire CRM data overnight, and a family care-coordination app she had Codex build to manage her 81-year-old father's medical care across nurses and family members in Spanish and English. She also shows off Claude artifacts used for real-time trip planning in New Orleans.

The conversation threads a recurring theme from Every's philosophy: knowledge work is shifting from 'sculpting' (doing every task by hand) to 'gardening' (building systems and loops that create the conditions for good outcomes) and from software as rigid rule-compilations ('bones') to language models as adaptive reasoning layers ('brain and ligaments'). Natalia closes with practical advice for executives: start by handing AI the same systems and shared context you already use to manage a team, focus on standardizing one task at a time, and resist the urge to reinvent everything at once.

Topics

Claudie's evolution and the limits of AI agents vs. hiring humans

Natalia opens by giving an update on Claudie, the internal AI agent introduced in the prior episode. Claudie has gone from a project that engineer Nitasha was manually 'Wizard of Oz-ing' behind the scenes to something that now runs autonomously every day: managing dashboards, posting to its own LinkedIn and Twitter, and running a self-evaluation loop against feedback, which Natalia jokingly calls its 'trust battery.'

Despite this progress, Every is simultaneously hiring a human operations person, and Dan presses on why. Natalia explains that AI, and Claudie specifically, is excellent at executing against a well-defined standard operating procedure, but still requires constant oversight to maintain quality and taste — a manager still has to direct it toward excellence, which is time-consuming. Just as importantly, she personally wants to interface with people, not just an agent, and believes humans are still needed to interpret the rich data and dashboards Claudie produces, surface what's actually interesting in it, and lead conversations based on those signals. Her expectation is the team will keep growing specifically to build on top of what Claudie surfaces, not to replace it.

Why they bought a CRM (Attio) instead of vibe-coding their own

Natalia describes moving the company off a homemade, Claudie-and-Google-Sheets sales pipeline system onto Attio, a professional CRM, and Asana for project management. Dan frames this against the live debate about whether AI makes all SaaS replaceable by vibe-coding — if you can one-shot a CRM, why pay for one?

Her answer: some companies' entire business is doing one narrow thing extremely well, and maintaining a bespoke tool has real ongoing costs. In the old setup, Claudie had access to her email, meeting notes, and inbound leads, and tracked everything in a Google Sheet that became a de facto database — but keeping that data's quality high required constant human maintenance, similar to managing Claudie itself. She walks through a concrete example: tracking a sales lead through a pipeline with logic rules for how deals should progress. She could mentally track deals over a 2-3 month window, but conversations that stretched longer or multiplied in volume became untrackable by hand. Attio has the same data access Claudie had but with robust built-in logic to track deal movement and flag issues automatically — something she could have trained Claudie to do, but only with significant additional investment. Dan connects this to a broader point: real software is a compilation of thousands of small logical rules discovered over time through use, which AI is good at working within and writing, but not at one-shotting from scratch. Natalia says this experience made her 'a really big fan of PRDs' and better at scoping work and deciding early whether to build vs. buy.

Software as bones, language models as brain and ligaments

Dan offers a metaphor for how software and language models complement each other: software is like bones — without it you'd be a 'flopping jellyfish' with no structure — while a language model is like the brain and ligaments, providing adaptive reasoning and movement. Without the model layer, a system is just 'a pile of sticks.' He notes language models can effectively 'grow bones' (i.e., generate structured software), but that growing a whole coherent 'body plan' is complex. This framing sets up the rest of the conversation about Natalia's shift from being non-technical to actively building her own tools.

The shift to Codex and why it changed how Natalia builds

Natalia confesses that Dan nagged her for weeks to try Codex before she did, and credits her own AI-related growth over the last month or two largely to it. Coming from Claude Code, where she had grown comfortable with file systems and folder structures, she found Codex's combination of terminal, browser, and chat interface, powered by a very capable model, removed much of the mental overhead of understanding how her projects were organized under the hood. She describes literally 'feeling the compute' — the model visibly wants to do hard work.

She frames her personal journey as one of increasingly not needing to become a traditional engineer: she used to spend a year trying to figure out which technical skills were worth learning, but with Codex she doesn't have to think as much about architecture (which she considers a real skill their engineers have) and can instead trust the tool to make good decisions and build working solutions directly for her.

Demo: the 'learning cartoons' skill for turning research into illustrated guides

Natalia shares her screen to demo a favorite Codex skill, originally a prompt built about six months ago (based on work from engineer Nitasha and modeled on Claudie), that codifies how she likes to learn any new topic: its history, the first-principles physics of the topic, how the field arrived at today's state, and the variables at play in its current landscape. Normally this produces a long guide she reads over a weekend, but when she doesn't have 12 hours free, she has Codex's visual model turn the guide into illustrated cartoons/zines she can scroll through on a walk or the subway.

Her worked example is physical education and fitness: after months of sitting at a computer neglecting her physical health, she asked for a guide on the history of workouts and how to make better exercise decisions. The cartoon explained that structured 'workouts' as a concept emerged roughly 200 years ago, tied to the Industrial Revolution as people shifted away from physical labor in daily life. She highlights learning about anatomy and specifically the different time scales at which muscles, ligaments, and bones each get stronger, which shaped how she thinks about progression in physical strength. Dan connects this back to his bones/ligaments/brain metaphor from earlier.

How Natalia organizes projects in Codex (pinning)

Dan asks how she organizes her many Codex projects, specifically whether she uses pinning. Natalia says she only pins one project — her email triage app (built on Dan's 'Tend' app) — and works loosely in everything else, unlike Dan who pins everything to avoid losing work. She shows a project list, including things like sales strategy and a personal 'epistemology' learning project (referencing an offhand dive into Aristotle's syllogistic systems), and explains she organizes Codex projects the same way she organizes her actual work projects, working in whichever one she's prioritizing that day.

She also contrasts the experience with Claude Code, where she felt she had to keep Finder open and actively track file systems and where things were being saved; Codex handles that organization more visually, reducing her mental load.

Demo: the email triage app, built and customized on top of 'Tend'

Natalia walks through her heavily modified version of Dan's open-source email-triage app (formerly 'inbox sweep,' now called Tend). The original version had buttons for archiving or sending emails, custom to Dan's own workflow. Natalia rebuilt hers around her actual needs: delegating tasks, tracking things in Asana, and managing communications across client teams that can include hundreds of employees at once, which creates significant mental load during triage.

Her inbox app is trained on a 'ghostwriter' skill she built about a year ago from roughly 150 of her own sent emails, so it understands how she communicates in different contexts. It's overlaid on her inbox with full context of prospective and existing client work, and drafts replies in her own voice. She can approve and send a draft, ask for a rewrite, archive without replying, or route content into a saved markdown file per client (used even when a reply isn't needed but context should be preserved), or convert an item into an Asana task with one click. She jokes that her email now 'knows what's going on more than I do,' since its drafts are often more accurate than what she'd write from memory. Dan highlights that this ability to spin up a custom button/integration on the go, mid-triage, is what makes the tool feel transformative rather than just a nicer inbox.

The sales-pipeline flowchart, and loops as the new mental model for knowledge work

Dan recalls a moment that convinced him Natalia truly understood this new way of working: finding her in the office on a Sunday building an elaborate flowchart of sales-pipeline logic. Natalia pulls it up — a detailed flowchart of how inbound leads move through stages depending on fit and conversation progress. She explains this is the same logic Attio uses to manage the CRM pipeline and the logic she feeds into Codex for email triage; she PDF'd the flowchart and shared it directly with Codex so both systems operate from identical rules.

Dan uses this to unpack the idea of 'loops,' a term he says is popular but poorly understood. He contrasts old knowledge work, which was like sculpting — every outcome shaped directly by your hands — with the new model, which is like gardening: you create the conditions for growth rather than making each individual output yourself. A loop means building the system that handles a whole category of work (like all your email) rather than doing each instance manually, intervening at key points (the 'human sandwich' at the start and end of a process — deciding what's worth attention, then refining the draft), and compounding learnings back into the system after each cycle so it improves over time. Natalia connects this to Dan's earlier 'model manager' analogy from four years prior: this is an evolution from directing AI as an individual contributor doing one task well, to operating it as a system the way a manager operates a team.

Best loop example: an unattended overnight CRM rebuild

Natalia describes what she calls her best loop yet: while setting up their CRM's logic with a partner team, she gave Codex a single goal — update the CRM to accurately reflect what had actually happened across hundreds of client and prospective-client conversations, based on calls and email context — gave it a detailed prompt and direction, and let it run overnight. About 6 hours later, when she woke up, the CRM was fully populated, representing what she estimates would otherwise have been weeks of manual data entry. She attributes this success to the shared context and logic they'd already built (referencing the flowchart/rules), which let Codex make good judgment calls autonomously, and describes it as a genuine moment of joy and improved quality of life from AI.

Advice for executives: start small, reuse existing systems, don't rebuild everything at once

Asked what her consulting clients (executives at tech companies, hedge funds, and PE firms) should take away from these workflows, Natalia's first piece of advice is to start with the systems you already have: if you're already running a team with KPIs, shared goals, and OKRs, hand that same architecture and shared context to AI (where company policy allows), and think about which tasks should remain exclusively human — for her, that's live client calls, since 'only I can get on calls and have productive conversations with my clients' (with a running joke that her team member has 'cloned' her tone in written communication).

She warns that the single biggest mistake she still makes, and sees clients make, is trying to overhaul everything at once in pursuit of being maximally 'AI-pilled' or AI-first. Instead, the real work is unglamorous: standardizing and clearly writing down how to do one task really well, defining what 'done well' looks like, and only then moving to the next task. Done repeatedly, this compounds into systems capable of sophisticated autonomous work, but it starts with something as basic as a markdown file of simple instructions.

Personal projects: family care coordination app and AI for administrative life

Natalia shifts to personal use cases, describing a growing conviction that AI will absorb much of the administrative burden in people's lives, particularly for women managing complex family logistics. Her flagship example: she asked Codex to build a full application to coordinate care for her 81-year-old father, who works with multiple nurses and has ongoing medical appointments, procedure follow-ups, and family WhatsApp threads to track. Codex took the project from prototype to a complete, live, password-protected app over what she describes as a 13-hour working session.

She demos the app: a central portal aggregating Google Form reports from nurses and casual WhatsApp updates into a single top-line summary. Because she's Colombian and updates often arrive in Spanish, the app can toggle the digest into English for faster reading during the day. Both her family and the nurses (who have their own view, allowing continuity of care between shifts) use the same portal, plus a task tracker showing which family member is responsible for which follow-up, automatically updating based on conversation content. Natalia stresses that the value isn't really about the tool itself — the nurses have noticed the family being more proactive and responsive, making them better care partners, which is the actual outcome that matters. Dan connects this to a broader observation that people underrate how applicable AI is to caregiving and everyday life tasks, not just coding.

Claude artifacts for real-time trip planning

Natalia shares a lighter example of AI for 'rich information transfer' beyond text documents: during a trip to New Orleans with her mother for French Quarter Fest (initially misspoken as Jazz Fest), she used Claude artifacts on the go to generate guides in Spanish about things they encountered, such as New Orleans's storm pump system, which she found impressive engineering but didn't have time to deep-research herself. She also had it read her Spotify playlists to infer her music taste, cross-reference it against the festival lineup, and recommend which bands (timba and salsa acts stood out) they were most likely to enjoy, letting them use their limited time efficiently while exploring the city.

Visuals

1. Email arrivesAI ghostwriter (trained on 150 of Natalia's past sent emails) drafts a reply in her voice, using full context of client history.
2. Approve & sendOne click sends the drafted reply as-is.
3. RewriteAsk the AI to redraft the response before sending.
4. ArchiveNo reply needed; email is archived.
5. Save to markdownContext preserved into a per-client markdown file even without a reply.
6. Create Asana taskOne click converts the email into a tracked task.
7. SpamDiscard irrelevant messages.

	Homemade (Claudie + Google Sheets)	Professional (Attio + Asana)
Data tracking window	Reliable for ~2-3 months of manual oversight	Tracks deal movement indefinitely with built-in logic
Maintenance burden	Constant human oversight needed to keep data quality high	Vendor maintains the underlying logic and reliability
Flagging & alerts	Requires training Claudie extensively to replicate	Built-in pipeline logic flags issues automatically
Team burden	High — team must supervise and correct the system	Low — less burden on the team to maintain

Takeaways

• AI agents excel at executing standing procedures but still need human oversight for taste, judgment, and the parts of work that are inherently interpersonal — build the agent, but keep hiring humans to interpret its output and lead conversations.
• Buying professional software (like a CRM) can beat vibe-coding your own once data quality and hidden business logic accumulate past what's tractable to maintain by hand — it made Natalia 'a really big fan of PRDs.'
• A 'loop' is not doing one task well with AI — it's building the system (with a human sandwich at the start and end) that handles a whole category of recurring work, then compounding learnings back into it over time.
• Feed AI tools the same shared context and operating logic you already use to run a team (KPIs, OKRs, flowcharts) rather than building bespoke logic from scratch — a single flowchart, PDF'd, drove both the CRM and the email-triage app.
• The biggest mistake executives make adopting AI is trying to reinvent everything at once; the real leverage comes from unglamorous, one-task-at-a-time standardization written down in something as simple as a markdown file.
• AI's administrative leverage extends well beyond coding — an overnight Codex loop rebuilt a company's entire CRM from call and email context, and a 13-hour Codex project produced a full family care-coordination app for an aging parent.

Notable quotes

What Codex helped me do was basically create kind of like an operating system. My email knows what's going on more than I do.— Natalia

I'm just so bullish on all of the administrative tasks that will suddenly kind of like be taken care of because now we have this sort of like super alien tool that can support on those things.— Natalia

Knowledge work now is turning into something like gardening, where when you're gardening, you're creating the conditions for the growth to happen, but you're not like making the plant with your hands.— Dan Shipper

Software is a little bit like your bones in your body. And a language model is a little bit like your brain and your ligaments... if you didn't have any bones, there would be no structure and you'd be like just sort of a flopping jellyfish on the floor.— Dan Shipper

I think the single biggest mistake that I often still ambitiously make and also see our clients make is you want to just remake the whole thing. You want to be AI-pilled, be AI-forward, just be an AI-first organization. And so often that just means you need to standardize and write down how you do a single thing really well.— Natalia

I gave it a more of a robust kind of prompt and direction... and I think like 6 hours later, I went to sleep and 6 hours later it was complete. I woke up to effectively a CRM that was fully set up and had done I think what would have been like weeks of work.— Natalia