Last week a new ChatGPT app dropped and I figured it was another skin update. It’s not. The creator behind this breakdown, Matt Wolfe, spent two weeks living inside it and burned 6.8 billion tokens on Codex alone, and the walkthrough he put together made me rethink what “AI assistant” even means now. Quick note for honesty’s sake: OpenAI sponsored his video, and he says so upfront, but the demos are his own real workflows, so judge them on the receipts.
Here’s what stuck with me.
What’s actually new
OpenAI took four separate products and mashed them into one app. No more tab juggling.
- ChatGPT Work: the assistant side. Give it tasks, connect your tools, let it go do things.
- Codex: the developer side. Write code, review diffs, hunt bugs, build projects.
- Sites: built-in hosting. Build a site or game, say “send it to Sites,” and it’s live. No CDN, no database setup, no deploy dance.
- A real browser: with an annotate feature. Click a section of a page, circle it, say “add more green here,” and the model knows exactly what you mean without you explaining.
You can also import cookies and passwords from Chrome, so it logs into sites as you. And scheduled tasks let a prompt run on autopilot: daily briefs, weekly reviews, follow-up monitoring.
The twist
The app tour is nice. The twist is what the author did before using it.
He built a project called “personal assistant” and gave it one setup prompt: create a skill for yourself that pulls from everything I have. Gmail, Calendar, Slack, Google Drive, meeting notes, journals, past chats, memories, his Twitter and Instagram through the browser, his YouTube channel. Everything.
Then the part I think most people would skip: he made it read-only. It can see everything, write nothing.
That one constraint is what makes the whole thing safe to use. And the model built itself an evolving source map, so when he adds a new tool later, it gets folded in automatically. No re-setup.
The result is an assistant that knows what he’s overwhelmed with, what shipped, what’s stuck, and who’s waiting on him. I was genuinely a little stunned watching this part.
The mini-workflow
Here’s the setup, stripped to the steps you’d actually follow:
- 🧩 Create a pinned project. Call it whatever. This is your assistant’s home.
- Prompt one: the full-context skill. List every source it can read. Explicitly say read-only. Tell it to fire this skill on every prompt inside the project.
- 🗣️ Prompt two: the voice profile. The author asked it to analyze 12 months of his sent emails and work chats, excluding forwards and pasted text. It built separate email and chat voice profiles.
- Point it at old chats. Right-click any past conversation, copy the session ID, paste it in. Now it reads that thread specifically instead of you re-explaining context.
- ✅ Draft, never send. He lets it write emails and save drafts. He reviews and clicks send himself.
That’s it. Two prompts and a rule.
What it does once it’s dialed in
- What’s on my calendar, Slack, and email that I need to do this week? It returned a real triage list, flagged what had to close before his Thursday flight, and listed work it had already finished for him.
- Book one-pager. He pointed it at three old chats about a book he’s writing. It wrote the pitch sheet with no fresh context needed.
- Slide deck. He handed over an outline and a brief. It pulled screenshots from his own site and built the deck. He says it got him 95% there.
- Cold outreach demo. “Find me 20 leads in San Diego and draft personalized emails.” It researched them, wrote 20 different emails, and dropped them straight into Gmail drafts.
- Folder cleanup. 65 loose files sorted into existing folders in 83 seconds. Then 21 thumbnails renamed by what’s actually in the image, took 79 seconds.
The Codex side
This is where the 5.6 models show off. The author had Ultra grind for 18 hours straight on a single project.
He rebuilt Craigslist with a cleaner spin, and it genuinely doesn’t have that vibe-coded look. He built an original puzzle game called Echo Garden where you solve levels by getting your own clones to repeat your moves. He rebuilt Future Tools with a real recommendation algorithm: 18 hours of build time.
From the community: one person had 5.6 drive Blender through the graphical interface with no MCP, just computer use, and model a cannon. Another built a voxel Manhattan over a full autonomous week. Nate Herk made an entire talking-head video with script, avatar, visuals, and edit from one prompt.
Pro tips and honest caveats
- Slower is the price. Your one prompt spawns workers doing dozens of sub-tasks. Some finish in two minutes, some take an hour.
- The voice profile isn’t you. It’s close. You’ll still strip out the ChatGPT-isms.
- It doesn’t always listen. He said “publish to Sites” and it built on localhost anyway. One nudge fixed it.
- Read-only is the whole game. Full access to your business without a write permission is what makes this sane.
- Build tiny, stupid things. His favorite? A one-button app that toggles his Mac dock. Saves five seconds. He loves it.
That last point is the real takeaway. You don’t need to sell it or scale it. Build the dumb little tool that annoys you and move on.
🎬 The full video has the actual prompts on screen plus the community builds. Worth the watch if any of this clicked.