GPT-5.6 + Codex is an Operating System. Run your life and business
Codex is an Operating System. Dan Shipper Runs His Whole Life Inside It.
Dan Shipper has been testing GPT 5.6 for weeks.
He came on the show, shared his screen, and showed the whole setup: how he answers email, how he tracks his company, how he built a SaaS app live, how he trains models over the weekend.
Here is the system.
The mislabel
People hear Codex and think coding tool. They hear ChatGPT and think same thing, different wrapper.
Dan's position: those are two different products, and the second assumption costs you the most.
He calls the Codex CLI trash. The desktop app is the thing. He describes it as what the Claude desktop app would have been if it were rebuilt from scratch, because OpenAI got to watch the agent-on-your-computer paradigm play out first and skipped the messy part.
His words on where he lives: Codex is my operating system for work.
Where 5.6 actually sits
Dan grades it against Fable.
Fable is S plus tier. So good it's illegal. Also slow, also expensive, also built for people orchestrating fleets of agents or doing machine learning. With Fable, he emptied the bug backlog by setting it loose and paying for it.
GPT 5.6 is A tier. Roughly 60 to 70% of Fable, by his count. You still have to think to fix a bug.
His analogy: Fable is a tactical nuke. 5.6 is a Porsche. Great to drive around town, fast, handles well, good for collaborating.
The upgrade from 5.5 is specific. Drafts read as if he wrote them for about 90% of his email. Fewer mistakes. More careful about what it sends. Resourceful enough to browse Facebook Marketplace for his new apartment and render what the guest room would look like.
The model is only as powerful as the context you can give it.
The core shift
This is the line worth stealing.
Dan stopped optimizing how he does his work. He started managing the system that does his work.
He compares it to what happens to a manager, or a founder making a first hire. The individual email stops being the unit.
It's not about any individual email. It's about the system that does my email. That's the job.
Step 01: Feeds, cards, and a next best action
Dan's Codex is built around pinned threads. One per part of his life he cares about.
The inbox thread sweeps every unread email. Each email becomes a card. Each card carries a summary and a draft reply. He clears the queue by talking.
The company thread does the same thing with Slack messages and meeting notes he hasn't seen. Cards. Issues. One decision each: care, or archive.
Then it compounds. After a pass, it reads what he replied to, what he archived, what he ignored, and rewrites its own prompts. There's a feed policy sitting in there describing what Dan cares about and what's on his mind.
Everything becomes a feed. Every card gets a next best action.
The email app he built is called Tend. He plans to release it open source. He also said you can feed the video into Codex and build your own version.
Step 02: Mailroom, routers, heartbeats
Dan gave his Codex an email address.
He uses the plus format, dan+codex, with a random string on the end so nobody can just email it. He calls the setup Mailroom.
A router thread checks that inbox every 5 to 15 minutes and decides where things go. Inbox? Company pulse? Somewhere else?
Then the payoff. There's an agent in his company Slack. When someone needs an ops task from Dan, they ask the Slack agent. The Slack agent emails his Codex. Dan watches it get handled.
A heartbeat wakes a thread on a schedule.
A router decides where an input belongs.
Cards give you the decision.
Three primitives. That's the architecture.
Step 03: Give it context, then ask it what to do
Chronicle takes local screenshots of your computer and turns them into a feed of who you are and what you're working on. It's a research preview. It's optional. Dan leaves it on because he'd rather the responses be better.
Record and replay is newer. You say "watch me do this," it records your screen doing a task, and it turns that into a skill. Dan hasn't used it yet. He called it sick.
The first prompt he recommends after install:
"Can you propose for me things you can do for me based on how I use my computer?"
Download it. Give it access. Ask it what it sees.
Step 04: Build something and stop at 70%
They built one live.
The app is called Turnaround. The premise: anyone can one-shot a SaaS app now, so the signal of a real product is maintenance. A badge for your about page showing commits, issue turnaround time, activity. A status page for the vibe-code era.
Dan ran it through the Compound Engineering plugin built by Kieran Klassen.
LFG loops the cycle: plan, execute, review, compound the learnings, keep going.
goal gives the model a long-term target and verifiable done criteria, then lets it run.
His current goal has been training a copy-editing model for four hours. Over the weekend he had runs going 20 hours each.
5.6 asked good questions before building. Should the badge promise alive, responsive, or dependable? It had a recommended option. Dan's read: earlier models had no opinion.
The output came back at 70 to 75%. 26 commits, last 30 days, rendered in a flat aesthetic with a warm paper background it wanted for no reason he could name.
His grade on 5.6 design: a B, where previous models were a C or D.
His actual move: go to Claude for the design, port it back into Codex. He thinks Claude is still ahead there.
Pirates and architects
Dan splits builders into two types.
The pirate gets it to 70%. Explores the field of possibility. Builds enough to answer one question: is this valuable?
The architect takes the 70% thing and makes it a well-oiled machine.
Dan knows he's a pirate. So he pairs himself with an architect who gets energy from taking his half-formed excitement and polishing it.
The old advice was find a technical cofounder. Everyone can code now. The real pairing is speed with polish.
The SaaS take
Dan thinks the SaaS apocalypse is dumb, and his reasoning is a business model, not a vibe.
He vibe-coded a CRM. It became a mess quickly. They moved to a CRM vendor with a CLI, because that vendor already thought through the corner cases and will keep thinking through them.
99% of the world wants software somebody else maintains.
You can one-shot the first version. Maintenance is what people pay for.
Then the opportunity, stated plainly: build Codex-native SaaS. Not a CLI you delegate to. Software designed for you and your agent to work on the same surface, inside the in-app browser, with the agent already holding all your context. Shorter feedback loop. Programmers already know this, which is why the browser exists.
The margin argument is the sharpest part.
SaaS companies installing agents into their products pay token costs. Dan's example: $80M in revenue against $60M in tokens. In the Codex-native model you host software, the user brings their own agent, and your margins are a software business again.
He expects app stores. Within a year, from OpenAI and Anthropic, for apps built to be used inside these services. He was clear he has no inside information. He just thinks it's too good a business to skip, and everyone who lived through the App Store knows what that ecosystem did. [The host built an app in 2013, got featured, and pulled 25,000 downloads a day from the feature alone.]
Establish yourself as the Codex-native app in a category now. Rank for it. Build the landing pages. Be the place people go when it's mainstream.
Where the models still lose
Copy editing.
Dan's team has made tens of thousands of copy edits, possibly hundreds of thousands, over five or six years. No frontier model does it like a copy editor does. Not even to 70%.
So he's fine-tuning. That means open source models, because the major labs stepped away from fine-tuning.
His claim about the frontier: training your own model is the step after making your own skill. When the skill almost works, fine-tune. Until recently that was out of reach for anyone outside machine learning. 5.6 can now stand up the whole pipeline, from grabbing data to generating synthetic data to running experiments.
Any organization sitting on years of a proprietary judgment call has the same opening.
Start here
Download the Codex desktop app.
Give it access to your computer.
Ask it what it could do for you based on how you work.
Pick one problem you actually have. Solve that one.
Add a second thing after the first one earns its place.
Dan's warning about his own setup: copying it wholesale means spending weeks building something that doesn't fit you. He started with one problem. He solved it. He built from there.
He named the failure mode too. It's the Notion and Roam pattern, where the system becomes the work and the work never happens.
And FOMO is the wrong engine. Curiosity is the right one. Come to it because you want something from it.
Final thought
The shift is smaller than it looks and bigger than it sounds.
Every repetitive task moves inside the agent. Every input becomes a feed. Every feed produces cards with one decision attached.
You stop being the person who answers.
You become the person who designs what answers.
Same work. Different job. That's the era.
Checkout the full episode:
Apple: https://podcasts.apple.com/us/podcast/the-startup-ideas-podcast/id1593424985?i=1000776150885
Spotify: https://open.spotify.com/episode/02gwxXQfEgyEecaL8TTyOq?si=5decee4d94034bd3