A solo founder's daily workflow with an AI agent that commits code, writes content, and ships products.
Jeff Weisbein · @jeffweisbein · OpenClaw
I run 4 products. Solo. No employees. No co-founder.
My AI — Fubz — has commit access to every repo, database access, social media accounts, and deploys to production multiple times a day.
This is what that actually looks like.
Live audio rooms for friend groups. iOS app. The heart of the operation.
AI second brain that actually remembers everything. Web + Telegram.
AI journalist discovery for PR teams. Find who actually covers your beat.
AI-powered RFP response for PR agencies. Save 32+ hours per proposal.
Part 1
What Fubz actually has access to — and why that matters.
Git commit + push to all 4 repos. Deploys to Vercel automatically. Never asks me to run commands.
Direct Supabase access. Runs SQL, creates triggers, fixes user data, manages subscriptions.
Twitter/X (3 accounts), Reddit (2 accounts). Searches, posts, replies. Runs outreach campaigns.
30+ cron jobs. Heartbeat every 30 min. Proactive checks on email, calendar, opportunities.
Fubz wakes up fresh every session. No built-in memory between conversations.
Everything it "knows" comes from markdown files it reads at the start of each session — and writes to throughout the day.
If it's not written down, it didn't happen.
Part 2
What a typical day actually looks like.
Everything happens over iMessage. I send a screenshot, a bug report, or a vague direction — and Fubz figures it out.
No tickets. No PRs for small stuff. No standups. Just continuous conversation.
4:30 AM. I report getting logged out of Cackles overnight. Fubz queries the database, finds the smoking gun, and codes the fix — all in one conversation.
🔍 Root cause found
8 token rotations in 30 seconds. Race condition — multiple refresh calls hitting Supabase simultaneously, rotating the token out from under each other.
A user fell asleep with a room open for 7 hours. I ask Fubz to add a "still there?" system. 15 minutes later it's pushed to main.
📦 3 files changed
New RoomIdleMonitor service, integrated into RoomCoordinator + RoomView. In-app alert with 20-second countdown, local push if backgrounded.
February 21st. One session. Ten bugs squashed, each one shipped to main. This is what "run with things" looks like.
5-6 rounds in 20 minutes. Multiple times a day.
iOS Swift, Next.js, Supabase SQL. Reads code, finds root cause, pushes fix.
Blog posts, landing page copy, SEO. 12 posts across 4 sites in one session.
Tweets, replies, opportunity scanning. 17 rounds of promotion in a day.
User management, subscription comps, ghost cleanup, trigger creation.
Built a 20-sec ad from screen recordings. 20 iterations with ffmpeg.
Spawns background agents for big tasks. Full copy rewrites in 7 minutes.
Part 3
How I make it smarter over time.
Early on it confused "portals" with "rooms." Used "we" instead of "i" in tweets. Modified pages I liked.
Each correction goes into MEMORY.md. Future sessions don't repeat old mistakes.
Week 1: constant corrections.
Week 4: it predicts what I want.
Part 4
4 weeks. From "can it build?" to "it runs half my company."
Set up Twitter CLI, Reddit accounts, 5 cron jobs. The "can this thing actually build?" test.
Broke Stripe, Google OAuth, Clerk, everything. Lesson: use printf not echo. Permanently logged.
3,000 lines of code. 30 files. StoreKit 2, Supabase schema, pricing, UI. One afternoon.
First time the AI was literally inside the product it helps build. Agora Conversational AI integration.
12 blog posts, copy rewrites, SEO improvements, blog post capitalizing on Discord's age verification backlash. Plus bug fixes. Plus building this deck.
Commit access, database access, social accounts. Day 1. It's not useful if it can't act.
iMessage. No tickets, no PRs, no standups. 5-6 rounds in 20 minutes.
Everything in files. Corrections, decisions, gotchas. The files ARE the continuity.
"Run with things, I'll course-correct." The default is action, not asking permission.
Every correction compounds. Week 1: constant fixes. Week 4: it predicts what I want.
Best work happens late at night in flow. ClawdGuard, Infinity, The Bridge. Trust gets built here.
"Jeff steers, Fubz rows.
When Jeff says 'make those look good and function' — that's enough to go on."
Multiple agents working on the same products need shared context. Without it, your content agent contradicts your dev agent's decisions.
Two layers:
Product context, voice guide, decisions log, user signals. Every agent reads these at startup.
Competitors, trends, market opportunities. When you share an article, it gets filed and cross-referenced. Future conversations pull from it automatically.
Mid-conversation, I realized I was the bottleneck. We'd be debugging code, I'd want to pivot to a tweet draft, and the first task would just… vanish.
So I told Fubz: "I want to do less remembering."
Two minutes later, we had a system:
Every task gets tracked in tasks/active.md automatically. Nothing slips through the cracks.
Say the words and Fubz snapshots exactly where we are — context, progress, what's left — then we switch. Say "what's parked?" to see everything in flight.
The best agent features come from real friction. I didn't plan this — I complained, and we built it together in 2 minutes. That's the loop.
Everything you need to build your own AI agent workspace — open source.
github.com/jeffweisbein/openclaw-starter-kit
Fork it, fill in your details, point OpenClaw at the folder. You're live in 5 minutes.
Jeff Weisbein · @jeffweisbein
Fubz 🤪 · The AI that ships
cackles.club · sidebra.in · whocoversit.com · wizardrfp.com