I've been using Codex's new browser for a couple of weeks and it's taken over about 90% of what I used to do by hand. It's a normal browser inside the ChatGPT app, signed in as me — I'm just not the one using it. Which means every job that never had an API, I don't do manually anymore. Skool has no public API and I now run my whole community through it. - The three jobs I actually run. A Meta ads campaign built from my last one plus the GoHighLevel funnel it points at — opt-in and thank-you page off a template, form wired to the right pipeline, everything left as a draft. A weekly Skool owner's report — every zero-comment post and unanswered question from the last 7 days with draft replies in my voice, new-member welcomes, pending requests, the leaderboard and the calendar, posted to Slack. And my X feed read through Grok for what's actually moving in AI, turned into a draft post, now running on a schedule. - The rule for when to use it. Good API → use the API (faster, cheaper, same every time — that's still where most of my automation lives, and that's not changing). Already has a plugin → even easier. No API, or one that can't do the job → browser. All three demos are that: Skool has no public API, GoHighLevel's has one but it won't build funnels, X has one but it charges you to read a feed you already read for free. - Cost + settings. It runs on your normal ChatGPT plan — no separate bill, no API key. But the model matters: the heavy model burns credits fast and you don't need it to click around a website, so drop to a lighter one for browser jobs and keep effort at medium. My take: the thing to actually be careful about is prompt injection. This browser is signed into everything I'm signed into, and a page can carry hidden instructions written for the AI. OpenAI says they can't fully solve it — which is why every prompt I ran had limits written in: read-only on Skool, draft-only on X, campaigns stay paused. That's not caution for the camera, that's how I run it.

I ask it how the ads did and it answers out loud, in the car. It has my email, my calendar, ClickUp, the CRM my leads sit in, and the channel my team talks in — and I rebuilt the whole thing on Fable 5.1 this week. Underneath there's no app, no service and no server: it's Claude Code running headless on my Mac inside a folder of text files. - Why it can do so much. Because it's Claude Code, it inherits everything I'm already signed into — email, calendar, ClickUp, Slack, Drive, call recordings. And the part that actually changed how I use it: it drives my real browser, the one with all my logins. Anything without a proper connection (my ad account, GoHighLevel) it just opens the page and reads it the way I would. Two other doors: hand a job to Codex when it's code, or let it drive the Mac itself after it asks. - The voice loop, and what leaves the machine. Everything runs through Telegram, so the interface is a text or a voice note. Voice arrives → transcribed locally with Whisper (no key, nothing uploaded) → Claude Code answers → ElevenLabs speaks it → back as a voice note. No webhook, no server, no public address pointing at my house — my Mac just asks Telegram every few seconds if anything's new. Two things leave: the message, and the sentence to speak. My recordings never do. - Where Fable actually earns it. Not everywhere, on purpose. "What's my next meeting" is one calendar lookup — running that on the biggest brain is taking a helicopter to the corner store. So it picks a model per message: plain lookups to Sonnet, ordinary work to Opus, the genuinely hard ones to Fable. Hard means holding five or six tool results in its head at once and coming out with an opinion. And Fable built the thing too — when the browser timed out, it found the cause and swapped the library; when a tool got quietly blocked, it audited the log and found four more I hadn't noticed. My take: it's a butler, not a specialist. Ask it to write a LinkedIn post and it hands you the brief rather than writing it badly — something mediocre at everything is worse than something that knows what it isn't. It's also not fast (30 seconds to a minute and a half), and it lives on my Mac, so Mac asleep, Jarvis asleep. On the $200 Max plan it costs nothing extra; the dollar figure it prints is an API estimate, not a bill.

If you've been getting AI slop in your outputs, pairing Opus 5.5 with Jev eliminates it. Opus 5.5 is Anthropic's newest model and it writes a lot better. Jev can't write a single word — it only checks. One writes, the other checks everything before it goes out. I ran three weeks of my own AI news through it: 1,300+ stories in, about 200 out. - Two kinds of slop, two doors. Slop with nothing in it (nothing specific, nothing new, or it just wants a DM/click/sale) can't be fixed — you throw it out. Slop that sounds like a machine can be completely true and still read like filler — that one you rewrite. The first kind comes in with what your AI reads; the second goes out with what your AI writes. Jev stands at both doors. - Why Jev and not just a big model. Jev never writes — you give it text and a question, and the only answers allowed are the ones you wrote down, each with a confidence number. Under a second per item, ~4¢ per million tokens in, answers free. At that price you stop sampling and check every item. A frontier model costs 100x+ more per check — at 1,000 items a day that's $400–700/mo versus about $2.50. Small judge on everything, big model only on what it's unsure about. (Anthropic's own docs say the same: lightweight model in front as a screen.) - The rule that makes it work. The writer never grades its own work. A prompt with ten writing rules just hopes all ten land; checking each rule on its own tells you which one missed. If a check fails, the draft goes back to Opus with only the failed checks as feedback — two rewrites max, then it posts the link with no write-up. A bad write-up never posts. My take: this isn't about news. AI made writing almost free but it didn't make reading free — a Stanford study found 4 in 10 office workers got AI work in one month that looked finished and said nothing, ~2 hours each to deal with. Same gate works on customer replies, sales emails, or anything before you publish. The only thing that changes per box is what happens when the checker is down: a news feed can fail open, a customer reply fails closed.

Am exited to connect with people and knowing more about you.🤝

A developer open sourced a live spy-satellite globe you can run in your browser for free. Real flights, real military traffic, real cargo ships, real public cameras, real fires burning right now — one Earth you can spin and search. Here's why it matters 👇 Every Monday you rebuild the same report. Four exports, a pivot table, a chart pasted into a doc, and one number you're not quite sure about. That globe is structurally the same object: a dozen live feeds pulled onto one page so somebody can make a decision. The only difference is nobody built yours. Now a coding agent will — but only if you make it interview you about the report BEFORE it writes any code. A lazy prompt gets you a beautiful dashboard of the wrong numbers. The full step-by-step guide is attached to this post as an HTML file. Inside: the free keyless install (two paths, no terminal needed), the exact interview and build prompts, and a table of what to point the agent at first vs what will burn you. 👉 Grab the HTML here https://claude.ai/artifact/Ch5KUL1FjkqUuktvK7xSAH Then tell me what your Monday report actually is 👇