Tim Toulas

easily missed

@philkellr Maybe „How OpenClaw works“ at 2:44:35. Also 31:57 Self-modifying AI agent. The first 30 min are ok to skip 😉

@etoulas thanks for chiming in Tim 👋 Yes I was considering to watch this too - but maybe a bit long - any segments you could recommend?

I finally got OpenClaw - after 20h 🦀 Here's how you can learn it faster I so wish I knew this before! 1. Start with a real use case 🎯 There's no way around this. If you just install and think "I will think of a use case later" you'll get frustrated. Take a simple problem you want to solve for yourself. e.g. Do web research (using brave search api) and then write a blog post for it (connect to github). Once you built your first agent you will get better ideas. But you MUST start with something real. 2. The minimal setup 🧱 Ask ChatGPT/Claude how to… - install OpenClaw on your laptop (forget about macmini for now - you can move the project later) - configure only the telegram connector (it's easy, takes 5 minutes) - choose a beefy model (I chose anthropic/claude-sonnet-4-5 - don't worry about the costs. I was at $12 after playing a full day) 3. Build OpenClaw from within OpenClaw 🧠 That was the hardest for me to understand: Once you have your telegram connection to your main agent you build agents with the prompts to the main agent. Don't watch Youtube tutorials (I watched some, they all didn't help), this is a learning by doing exercise. You'll also learn the concepts bit by bit. Don't prompt ChatGPT/Claude to build because they are not up to date with the latest OpenClaw versions (there are new releases every day!) You just prompt in Telegram and ask the agent to build other agents. Start with the big vision and then ask it to build step by step - the normal vibe coding way. 4. Keep using Cursor / Claude Code 🧑‍💻 You probably need custom code for your agents: you'll build scripts your agent will call (for me it's python scripts my agents exec with python myscript​.py). These scripts you'll still build with Cursor or Claude Code. Ask your main agent to give you prompts for it. The division of labour: Your main agent will build the wiring (AGENTS​.md etc.) and Cursor builds the software to it. 5. Lock down OpenClaw 🔒 OpenClaw runs as your user so it can execute anything you can. There are two ways to go: isolated docker environment or an allowlist of commands. Then make it only to respond to your telegram user - or else someone else might find it and can prompt your main agent 😱 But again - ask OpenClaw to do it for you. One extra tip - backup openclaw.json I trashed my config 2 times already. On my OSX I've set up this cronjob: 0 * * * * cp ~/.openclaw/openclaw.json ~/.openclaw/backups/openclaw_$(date +\%Y\%m\%d_\%H\%M).json

@philkellr I immediately went with docker. The disadvantage is that you will lose the integration with brew. The repository provides a convenience script to bootstrap everything github.com/openclaw/openc…

@philkellr Poetic

The longest way to bed is The distance from the laptop lid to the keyboard

@philkellr @MilaChervenkova Here’s something I really enjoyed recently. netflix.com/de/title/81937…

@MilaChervenkova thanks Mila - I still find it hard to not work at all :)) after 2 movies I was bored out - any tips?

Day #6 of sustainable side building with a 9-5 job The last 48h I built … nothing! I got sick on Saturday And decided that while sick I won't code nor tweet. I don't care about GitHub or Twitter streaks Took me years to have this insight, but since I see people coding even while sick I thought I'd share

@philkellr That’s assuming that before you were not tired at 21:00? What did you do instead?

Day #3 of sustainable side building with a 9-5 job I'm building solo while having a family and a demanding 9-5 job. I finally realized my main problem: It's not lack of time!

🤖 Claude Code prompt changes v1.0.128 → v2.0.0 🔗 Full diff: github.com/marckrenn/cc-m… 🧵 10 major changes detected ↓

@philkellr The last piece is a code word for the 20% that the user has to finish / endless AI loops

Why aren’t they using lovable to build lovable. Pff.

@philkellr I typed „Grill timing calculator“ into the Replit app on my iPhone. Then hit the magic pen that created a detailed prompt. I didn’t bother reviewing it and just sent it off. Then 20 min later reviewed and deployed it.

@etoulas This is great! I‘ll start grilling in 30m but it will be sausages:)) What were the 3 words you used as prompt?

I wanna teach vibe coding to Product Managers How would I best do this? A) jump right into it with a tutorial project B) send them to programming basic course first C) interweave "tech snacks" into vibe coding tldr: can you vibe code without any coding knowledge?

@philkellr I would say no. Here’s a simple app. I literally built with three words as a prompt while having a barbecue with friends. 😉 grill-temps.ch

@etoulas So the two programming languages won’t be a problem?

@philkellr I’d argue that you can totally use it for vibe coding too. You should teach some basics or assume those are clear: - how a web app works (when do I need a db, etc) - what git is used for (it integrates seamlessly) - how to host a real app (deploy to prod using your own domain)

@etoulas Thanks Tim, this is some crucial input! I checked this with chatGPT and it agrees with you :)) Replit seems the way to go as it is production ready and has a real stack. My only concern is that it's python based, so they'd need to learn both python and JS 🤔

@philkellr @stefanmeyer This is what the magic pen in Replit made out of it: Love this feature. I built a small tool with just three words this way.

@stefanmeyer like this?

I went down the local LLM rabbit hole the last days. Here’s how to get up and running quickly with a ChatGPT-like UX. And what I learnt along the way. toulas.ch/blog/2025/07/l…

OpenAI lanches Agent but not in EU + Switzerland yet..

@shreyas This in a single picture. Core assumption in your example is that work is charged by the hour. This plagues a lot of PM orgs indeed. But then again most business cases are driven by effort assumptions, so the outcome will always be contrasted with the time it took to achieve them

This happens in product work all the time: some product people consistently reach the correct conclusion within minutes/hours/days. Some take weeks/months/years to get there. And some others never do. The main skills that make it possible - Disciplined Discovery - Product Sense

@cagan Thanks for sharing. I recently discovered his talk and Q&A as CEO of NeXT Computers at MIT in 1992. (He’s 37 that day.) It shows his deep understanding of users and you can hear already the vision for the iPhone. I’m blown away by his story telling style. youtu.be/Gk-9Fd2mEnI

Product Twitter: treat yourself to a free, 1-hour true masterclass on product from Steve Jobs, circa 1995, covering product discovery, the dangers of process people, the perils of roadmaps, why so many companies lose their product mojo, and more: youtube.com/watch?v=aJzmlo…

Great tips in this thread. 🚀 LNO – Leverage, Neutral, Overhead tasks Radical Delegation (extension of Eisenhower Matrix) With an ROI mindset, people naturally gravitate towards decreasing the denominator via "quick wins". Opportunity Cost thinking favors big wins.

@pstadler (nod)

Microsoft Teams is the worst piece of software I've encountered in a long while.