You talk. It builds. You wake up to finished work. That’s not sci-fi — it’s ChatGPT’s new voice stack.
Most beginners hit a wall: prompting is friction. Typing context is slow. You stare at the blank box, unsure what to ask. By the time you’ve crafted the perfect prompt, you could’ve done the task yourself. So you don’t use AI. You stay stuck.
Three features change that equation entirely. They’re rolling out now. Here’s how to use them.
Codex Voice — Your Computer Finally Listens
How to Enable It (Plus/Pro/Team/Enterprise)
Open ChatGPT desktop or mobile. Look for the headphones icon (not the waveform — that’s the old voice mode). The new orb interface. If you don’t see it, update the app. It’s rolling out through August 2026.
Cost: ChatGPT Plus ($20/mo) gets you access. Pro ($200/mo) gets higher limits. Team/Enterprise included.
The Screen-Reading Orb — Context Without Typing
Hit the hotkey (default: Option+Space on Mac, Alt+Space on Windows). The orb appears. It reads your screen via Appshots — sees your browser, your IDE, your terminal, your email.
You talk. It sees. It acts.
“Book me a flight to Austin for next Tuesday. File the receipt to expenses. Check my calendar for conflicts.”
Three parallel threads spin up. One hits travel sites. One forwards the confirmation to your expense system. One queries your calendar. All while you keep talking about something else.
Thread-Forking — Parallel Work Without Context Switching
This is the killer feature. Complex task? Codex forks a thread. Branches from current context. Pulls from previous conversations automatically. You don’t organize — it does.
Example: You’re debugging a failing test. You say “fix this.” Codex forks a thread, reads the test output, explores the codebase, proposes a fix. Meanwhile, you say “also update the README with the new API endpoints.” Another thread. Same context. No copy-paste.
Your First Voice Automation: Book a Flight + File Expenses + Check Calendar
- Enable Codex Voice (headphones icon)
- Say: “I need to go to Seattle next Thursday. Book the cheapest morning flight, add to my calendar, and save the receipt to my expenses folder.”
- Watch the orb spin up three threads
- Review the flight options it presents (you pick)
- Confirm — it books, calendars, files
Total talking time: 15 seconds. Total manual work: one click to confirm.
Heartbeats — AI That Taps You on the Shoulder
What They Are (Proactive, Not Reactive)
Old AI: you ask, it answers. New AI: it watches, it tells you.
Heartbeats run in ChatGPT Work (mobile app). They monitor what you connect — email, calendar, bank alerts, package tracking, GitHub notifications. When something needs your attention, you get a push. No prompt required.
Set Up Your First Heartbeat: Suspicious Charge Alerts
- Open ChatGPT mobile app → Work tab
- Tap “Heartbeats” → “Create Heartbeat”
- Connect Gmail/Outlook
- Describe: “Alert me immediately for any email containing ‘suspicious activity,’ ‘unauthorized charge,’ ‘fraud alert,’ or ‘verify your identity’ from financial institutions”
- Choose notification: push + email
- Save
Done. Next time your bank flags a $0.01 test charge from a stolen card — you know in seconds, not days.
Package Delivery + Email Triage Heartbeats
Package tracking: Connect Amazon/email → “Notify me when any package is out for delivery or delivered today”
Email triage: “Summarize all unread emails from my ‘Important’ label every morning at 8 AM. Flag anything requiring a response today.”
No Code Required — Mobile Setup Only
This is the point. You configure in plain English. The AI builds the monitoring logic. You get the alerts. Zero engineering.
ChatGPT Sites — Deploy Real Apps From Chat
What You Get (SQL, Storage, Auth, Env Vars)
ChatGPT Sites isn’t a prototype. It’s a deployment platform with:
– PostgreSQL database (real SQL, not JSON blob)
– File storage (uploads, assets, user content)
– Environment variables (API keys, secrets, config)
– Email-based access control (invite-only, domain-restricted, public)
– Custom domains (yourbrand.com, not chatgpt.site/random)
– HTTPS, CDN, automatic scaling
All provisioned from chat. No CLI. No dashboard (unless you want one).
Build a Micro-Tool in 15 Minutes: Personal Expense Tracker
Prompt:
“Build a personal expense tracker. Needs: add expense (amount, category, date, notes), list expenses with filters, monthly totals by category, CSV export. PostgreSQL backend. Email auth — only my email can access. Deploy to ChatGPT Sites.”
What happens:
1. Codex writes the full stack (Next.js + Prisma + NextAuth)
2. Creates the database schema
3. Sets up email auth for your address only
4. Deploys to a live URL
5. Gives you the link
You get: A working app at expense-tracker-yournames.chatgpt.site (or your custom domain). Real database. Real auth. Zero server management.
Use Cases: Lead Magnets, Client Portals, Internal Tools
- Lead magnet: “Build a ROI calculator for [service]. Capture email before showing results. Deploy to Sites.”
- Client portal: “Build a dashboard where clients can see project status, download deliverables, approve milestones. Email auth per client.”
- Internal tool: “Build a script runner — I paste a SQL query, it runs against my DB, shows results as a table. Only my email access.”
These aren’t toys. They’re production apps. Built in minutes. Hosted for free (within limits).
The Video Workflow (Adapted for Beginners)
Nick’s Pro Pipeline: 50 Clips → Finished Verticals Overnight
Nick Baumann (OpenAI DevEx) records 50-60 raw clips. Uploads to a custom plugin. Plugin transcribes everything. Scans footage. Chooses strongest takes. Assembles finished vertical videos overnight.
What ate a content team’s day → handled while he sleeps.
Your Version: Codex Voice + Descript/CapCut
The custom plugin isn’t public yet. Here’s the beginner workaround:
- Record clips on your phone (talking head, screen capture, B-roll)
- Upload to Descript (auto-transcribes, removes filler words, fixes eye contact)
- Open Codex Voice → “Here’s my Descript project. Create a 60-second vertical cut: hook in first 3 seconds, three main points, CTA at end. Output editing instructions I can follow in CapCut.”
- Follow instructions in CapCut (free, mobile/desktop) — or pay Descript to auto-edit
Time: 30 minutes active work vs. 4+ hours manual editing.
Step-by-Step: Record → Upload → Describe → Polish
| Step | Tool | Time |
|---|---|---|
| Record raw clips | Phone | 10 min |
| Upload + transcribe | Descript | 5 min (auto) |
| Describe edit | Codex Voice | 2 min |
| Polish + export | CapCut | 10 min |
| Total | ~27 min active |
Putting It All Together — Your Daily Voice-First Routine
Morning: Heartbeats Briefing
Wake up. Check notifications. Heartbeats already summarized:
– “3 packages delivering today”
– “1 suspicious charge flagged on Chase ending in 4829”
– “5 emails need responses (2 urgent)”
– “Calendar: 3 meetings, 1 conflict detected”
You haven’t opened email. You know what matters.
Work: Codex Voice for Coding + Ops
Coding: “Add rate limiting to the API middleware. Use Redis. 100 req/min per IP. Write tests.”
Ops: “Reschedule my 2 PM to 3:30. Email the client. Book a Lyft for 3:15 to the airport.”
Research: “Find 5 competitors to [company]. Compare pricing, features, reviews. Put in a table.”
All voice. All parallel. All while you focus on the hard thinking.
Evening: Sites for Deploying What You Built
That expense tracker you prototyped? Deploy it. The client portal? Deploy it. The lead magnet calculator? Deploy it.
“Deploy this to Sites. Custom domain: tools.mysite.com. Email auth for my team.”
Live in two minutes.
Takeaway — Voice Is the Interface AI Was Waiting For
Typing prompts is the training wheels. Voice is the bike.
The friction of “what do I type?” disappears. You just… talk. The context you’d never bother typing flows naturally. The parallel work you’d never orchestrate manually happens automatically. The proactive alerts you’d never think to set up just… arrive.
Start with one feature. Codex Voice is the easiest entry — it’s in the app you already have. Enable it. Try one voice automation today.
Talk to your computer. See what happens.
You won’t go back to typing.
