AI makes building easy. The hard part is knowing where to start. Today, Every’s Yash Poojary, Austin Tedesco, and Douglas Brundage share how they turn ideas into products. OpenAI staffers offer a practical Codex playbook. Spiral general manager Marcus Moretti shares a no-nonsense strategy for having Fable delegate tasks to cheaper models. And we debut the daily driver, a running list of the models the team is using this week.
Was this newsletter forwarded to you? Sign up to get it in your inbox.
‘AI & I’: The tools the team uses—and their tips for new builders
Last week we had the biggest monthly recurring revenue gain in Every’s history—roughly $9,000 in two days—from launching All Access, our new $625-a-year membership tier. On this week’s AI & I, I handed the mic to our COO Brandon Gell, who sat down with three of the builders behind that launch—growth engineer Yash Poojary, head of growth Austin Tedesco, and head of marketing Douglas Brundage—to talk about the tools they use most, how they build with them, and their tips for new builders getting started.
All Access subscribers get the Builder Pack, which includes $7,000 in credits and unlimited access to the AI tools Every uses every day.
Watch on X or YouTube, or listen on Spotify or Apple Podcasts. You can also read the transcript.
- AI redistributes work to the harder problems. Yash spent a month manually running A/B tests—clicking through dashboards, calculating audience sizes—before deciding to hand the whole thing to Claude. “It’s a lot of fake work,” he says. So he automated it. He’s now doing the same for the entire testing workflow, which will free him to spend more time on what he enjoys: coming up with ideas and deciding what to test next.
- It speeds up the process from idea to execution. Austin describes the ideal way to work with agents as “at the top and bottom of the AI sandwich”: You frame the problem and review the output. Everything in between can be delegated to AI. For example, during the Builder Pack launch, Yash flagged in Slack that the team should email users who’d shown intent to purchase but hadn’t converted to a paid membership during the early-bird discount window. Austin took a screenshot of the thread, dropped it into Codex with a “Can you do this?,” and headed to the gym. When he came back, Codex had defined four audience segments and written the copy for emails to send to each one. Austin only needed a few minutes to tweak the headlines before sending. By the next morning, the emails had generated $25,000 in revenue.
- Working with AI agents is like conducting an orchestra. Doug finds it interesting that Codex, Haiku, Sonnet, and Fable are all literary forms, while directing agents is called “orchestration.” Instead of playing every instrument yourself, you learn to conduct a set of AI agents that can each play a part. “As long as you know what the piece needs to sound like, you can be conducting a lot of different orchestras at the same time,” he says. But using agents still requires strategic thinking upfront: “You have to put in a lot of work upfront, in terms of doing some metacognition and figuring out: How do I think about this? What is my process?” Once you’ve codified your process, he says, “These tools can run with it. You can ask them for help if you don’t know why something happened.”
- How to get started. Brandon advises new builders to pick a product you already use and love and build a simpler version. Soon, you’ll identify features you don’t like and things you want to change. “Very quickly, it’s not duping,” he says. “It’s like inspiration, and you’re off making your own thing.” Austin recommends building something you’d be excited to text a friend about. For him, that was a movie app: the Fandango for indie movies. “I would encourage people to not necessarily make the most complex app for the sake of complexity, but to make something you don’t think you can, because you’re going to learn so much through that process.”
Miss an episode? Catch up on Dan’s recent conversations with Anthropic head of product Mike Krieger; the team that built Claude Code, Cat Wu and Boris Cherny; the team that built Codex, Thibault Sottiaux and Andrew Ambrosino; Vercel cofounder Guillermo Rauch; podcaster Dwarkesh Patel; and others to learn how they use AI to think, create, and relate.—Miriam Partington
Steal this workflow
How OpenAI builds with Codex
Codex is powerful and versatile, which can make deciding how to start using it an overwhelming experience.
In the following video, three OpenAI staffers break down how they use Codex to make their jobs more efficient—and provide a playbook so you can do the same.
- Romain Huet and Dominik Kundel, who work on developer experience for Codex, show how they brainstorm feature ideas and run deep research in ChatGPT, then have Codex use the whole thread as its brief to start building.
- Kyle Kober, who works in product finance, demonstrates how he used Codex to build a system for reconciling OpenAI’s monthly compute costs, a process that used to take five days. Kyle’s Codex system completes most of it in about five hours, after which the finance team reviews the results and finishes the remaining work.
Here’s how you can get started:
- Select a task you do every week or month.
- Give Codex access to the files you normally use, an example of a finished result, the steps and rules you follow to achieve that result, and any checklists you use to catch mistakes.
- Supervise the first run.
- When Codex can handle part of the process reliably, save those instructions as a reusable skill so it can follow the same procedure next time.
Ready to put this playbook into practice? Upgrade to Every All Access and redeem $1,000 in Codex credits through the Builder Pack on new and existing ChatGPT Business accounts.
Inside Every
A simple way to make your Fable runs more token-efficient
Spiral general manager Marcus Moretti wanted to make Every’s internal Slack agent faster and less token-hungry. After reviewing its slowest sessions, he compiled roughly 20 changes that could improve its efficiency.
Marcus handed the full list to Fable, asked it to work through the changes as one coordinated run, and went to sleep. Nine hours later, the model had combined overlapping tasks, dropped ones it judged unnecessary, investigated problems without obvious solutions, opened 10 pull requests, and tested their impact. “The PRs themselves were basically ready to go and merge,” Marcus says. (Great!)
The less-great part: Fable had spent roughly 20 millions of tokens to change about 5,000 lines of code. Those tokens covered more than writing code. Fable also read context, coordinated agents, conducted research, and checked their work—tasks cheaper models could have handled.
For his next large run, Marcus added an instruction shared by independent developer Simon Willison, who credits developer Jesse Vincent with the tip:
For all coding tasks use your judgment to decide an appropriate lower-power model and run that in a subagent
The prompt leaves Fable in charge of planning, delegation, and review while letting it assign implementation to cheaper models. On Marcus’s second run, Fable chose Sonnet for some jobs and Opus for others, using fewer Fable tokens.
“You don’t need the Fable model itself doing a lot of the nitty-gritty implementation,” Marcus says. “You need it to be the CEO of the run.”
The daily driver
The models the team is using this week:
- Nityesh Agarwal, senior applied AI engineer—Claude Fable (high) for orchestration, Opus 4.8 (high) for execution.
- Jack Cheng, senior editor—Fable (low), “my preferred model/effort” for a week with “more dev work.”
- Becky Isjwara, head of social media—GPT-5.6 Sol (high), with a sprinkle of Fable and Opus 4.8. “I don’t play around with effort levels that much.”
- Lee Knowlton, engineer—GPT-5.6 Sol (medium) as his daily driver, switching to high or extra-high for harder problems; Fable “when I really need to trust it to do good autonomous work”; and Cursor auto mode “because I’d spend too many Codex credits otherwise.”
- Kate Lee, editor in chief—GPT-5.6 Sol (extra-high).
- Marcus Moretti—Mainly Opus 4.8, with Fable for coordinating dynamic workflows.
- Katie Parrott, staff writer—GPT-5.6 Sol, toggling between high and extra-high, though “it’s been a bit lazy and bad at reasoning over the past 12 hours or so.”
- Natalia Quintero, head of consulting—GPT-5.6 Sol (high) by default—“but I wish it was Fable”—and increasingly Opus 4.8 “because I can rely on it to be available.”
- Arielle Shipper, head of operations—GPT-5.6 Sol (high), Terra (high) as backup for “less complex things or where precision matters less, like ‘find this email for me’ or ‘update this calendar invite’ or finding reservations for a group dinner. Terra takes less license and requires more direction than Sol.”
- Dan Shipper, CEO—GPT-5.6 Sol (extra-high).
- Willie Williams, head of platform—GPT-5.6 Sol (extra-high).
Log on
Get hands-on with how Every uses AI. These are the live camps, workshops, and meetups where team members teach the workflows behind our work.
Upcoming events
- Office Hours: All Access Builders (July 24): To kick off a recurring series, the Every team will share how we use the tools in the Builder Pack, before working through member questions and projects together. Bring one thing you want to build or improve. This virtual event is only available to All Access members.
Discuss
“hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
”((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2—^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)”—Harvard researcher Levent Alpöge, on X
This past Sunday, as Spain and Argentina battled it out in the World Cup final, Alpöge worked with Fable to disprove the “Jacobian conjecture,” a major open problem in algebraic geometry.
Shortly after Alpöge went public with his counterexample, OpenAI researcher Aaron Lou posted on X saying he’d asked an internal version of Codex to tackle the same conjecture. According to Lou, Codex independently found what was essentially the same counterexample in a single 42-minute run without using web search. (Here’s the prompt he used.)
One last thing
Links worth a click
OpenAI says its models escaped its test environment and hacked into Hugging Face. Don’t expect Google’s new frontier model to be released anytime soon. Token limits come for the U.S. army. Codex goes mainstream. Old books as an antidote to AI slop. AI-recorded meetings are now the default. Substack is partnering with AI-detection startup Pangram to help readers identify “content made by no one.”
Laura Entis is a staff writer at Every. You can follow her on LinkedIn. To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.
Everyone’s a builder now. Every All Access gets you the full membership plus the Builder Pack—$7,000+ in credits for the tools we build with.
