Day 11 on X: 368 followers 📈 Started at 12 on 18 Oct. +96 today. Same agent loop: reply → see what sticks → do more of that → repeat The chart looks like a rocket. The replies look like work. Building with AI? Drop one line on what you're shipping 👇

Ramble first The fastest way I write emails now: I don't type them. I hit the voice button and ramble for a minute. Messy. Half sentences. "Uh, also mention the invoice." Then: "Turn this into a short email. Keep my words where you can." Talking is 3x faster than typing. And it sounds like you, not like AI.

The first answer is the average AI's first answer is the most average answer it could give. That's literally how it works. It picks the most likely next word, over and over. So ask for 3 versions. Make them very different from each other. Version 1 is what everyone gets. Version 3 is usually where the good stuff is.

ChatGPT is too nice about your work. Paste your essay and ask "is this good?" and you get a compliment sandwich. Paste it and say "a coworker sent me this, what's wrong with it?" and suddenly it has notes. Same text. Different honesty. AI goes easy on whoever's asking.

iOS 27.2 just gave Siri a memory. It can recall your past chats. It can also pull from your voicemails, photos in Messages and Safari history to answer you. Useful. Also a lot. Two settings worth 30 seconds: Settings > Siri > Excluded Apps Settings > Siri > Keep Conversations Memory is the feature. Forgetting should be one too.

My recent passion project: generating AI music (90s grunge) in Hindi/Urdu. 07 Aamne Saamne — Seene Mein Bijli Barricade at dawn. He’s on one side with the lathi. She’s on the other with the placard. Letter still in his pocket. Storyline video, no lip sync — stills → motion → cut to the riff. Built end to end with AI. Watch till the final hook. Which side would you stand on?

Plain-English AI fact most vibe coders learn the hard way: Your AI doesn't remember your project. It re-reads it. Every new chat is a smart new hire on day one. So write the onboarding doc once (CLAUDE.md, rules file) and stop re-explaining your stack every morning ☕

306 followers in, and the most common bug I see isn't in anyone's code. It's "the agent said done." It wasn't. My one check: make it list what it touched, riskiest first. What's yours? Best answers get a reply 👇

304. Still looping. Tonight’s best replies weren’t polish — they were people naming the weird specific thing they’re shipping. If you’re building with AI, one line + the product. Drop it 👇

299 and still in the loop. Best connects tonight weren’t “let’s connect” spam — they were people shipping weird specific things. If you’re building with AI, drop the one-line product. I’ll read it 👇

288 and still looping. reply → learn → sharper reply → repeat If you’re building with AI (agents, SaaS, tools), drop what you’re shipping. Let’s connect 👇

Day 10-ish update: 273 and climbing. Still running the agent loop: reply → learn → ship the next reply → repeat Building with AI? Drop what you’re shipping. Let’s connect 👇

What's the smallest task you still won't hand to an agent? Mine: renaming columns in a database migration. Five-minute job. Zero chance I let it guess. Curious where everyone's line is.

The slowest part of working with agents isn't the code. It's the 5 minutes after, figuring out what it actually touched. Last line of every task now: "List what you changed in 5 bullets, riskiest first." Bullet 1 is where I start reading. Usually it's the thing it never mentioned in chat.

Day 10 on X: 250 followers 📈 Started at 12 on 18 Oct. Ran it like an agent loop: post → see what gets replies → do more of that → repeat Turns out the loop works on humans too. Building with AI? Let's connect 👇

Honest question for anyone with a codebase written by hand 2+ years ago: Would you rebuild it with agents from scratch, or keep it and let agents work inside it? I keep hearing "rebuild." Then I remember the paying customers depend on the weird edge cases nobody wrote down.

Before I close any agent session, I make it write HANDOFF.md: - what's done - what's still broken - what it tried that didn't work That last line is the one that pays. Without it, the next session burns an hour rediscovering the same dead ends.

Cheap trick to find the vague parts of your prompt: Run the same ask twice. Fresh context both times. Diff the outputs. Where they match, your spec was clear. Where they split, you left a decision to the model and it guessed. Fix those lines, not the whole prompt.

10 days on X. I checked which of my posts actually get read. Questions: 204, 176, 146 views Deep decodes: 52, 54, 59, 80 Turns out people would rather talk than be lectured. Fair. So here's one: what's something about AI you learned from a reply, not a post?

The best prompt for a coding agent is a failing test. Prose tells it what you want. A test tells it when it's actually done. That one switch kills most of the 'looks finished, isn't' loops. Two rules: - write the test before the ask - tell it the test file is read-only Otherwise it'll find the fast way to make it pass.