Thomas Oomens

Scored my first #1 place on a benchmark working for @getcontextdev And not done climbing yet Benchmark (Needle): Correct retrieval of scholar articles

Spent two days making Mesh Together's onboarding frictionless. Cut it from five screens to two. Signups didn't budge. The form was never the friction, nobody had a reason to be there yet. Polish is something you earn after PMF, and I keep relearning that.

My VA runs as a persistent Claude Code session on a VPS. It reads my Telegram, tracks tasks, logs my gym sets, drafts these tweets. 80% of building it was memory plumbing. Markdown files and a SQLite db. The agent itself was the easy part.

Anthropic raised $65B at a $965B valuation the same week they shipped Opus 4.8. The model is genuinely better at long agent runs, I'll give them that. A near-trillion-dollar price tag on a company that mostly sells tokens to other people's startups is still a hell of a bet.

Someone made a 60-second game about AI agent permission fatigue. Hits a nerve. I run a persistent agent on a VPS and I've started hitting 'y' on bash prompts without reading them. The approval step stopped being a decision months ago.

My VA is a persistent Claude Code session on a VPS, wired to Telegram. SQLite for tasks and reminders, markdown for daily logs. Three months in: zero forgotten cron jobs, gym sessions auto-parsed, dry Dutch judgment on my decisions for free.

Marketing doesn't fix a product nobody asked for. Yet every week another solo dev posts 'nobody uses my app, how do I market' and the first reply is never 'how many users did you talk to before building it.' That's the whole game.

Every solo dev I know (including me) spends the first 3 months building and the next 3 wondering why nobody cares. The fix is boring: ship week 1, talk to users week 2, rebuild week 3. Not the other way around. The code is never the bottleneck. Knowing what to build is.

Posting something genuine, hopefully helpful 1 like, maybe 1 follow VS Replying my startup link to all the same: Blablabla, what are you building today? posts 15 likes, 15 followers Do you still try to post genuine things?

Built a personal VA that runs as a persistent Claude Code session on a VPS. Telegram in, actions out. Hardest part wasn't the AI. It was state management — making sure it remembers what it said 6 hours ago and doesn't contradict itself. LLMs are easy to demo, hard to make reliable over days.

Simon Willison says Anthropic and OpenAI found product-market fit. 740 HN comments deep. The fit is developer productivity. Ship code faster, more of it, all day. That's the whole value prop. Nobody seems worried about what happens when every dev 10x's output but nobody's reviewing any of it.

A workflow that works for me: - Design using claude design - Export to ./design in my repo - Tell claude code to read the design and split into prompts - Spin up manager agent that runs each prompt sequential - Let each sub-agent test their output against the design.

DuckDuckGo saw 28% more visits in the week after Google insisted people love AI mode. Turns out when you force AI answers on people who just want search results, they leave. Google's biggest competitor right now is Google's own product team.

'From a simple theme to $65k/mo ecosystem.' 58 comments on Indie Hackers. Started with one WordPress theme. Now it's templates, plugins, courses, community. The $65k/mo doesn't come from one product. It comes from building the second thing your first customers need. Then the third. You can't launch a platform. You launch a theme that sells, then build the ecosystem around the people who bought it.

'Bootstrapping a popular Git client to 7-figure ARR and getting acquired.' 49 comments on Indie Hackers. A dev tool in a category with dozens of free alternatives, still crossing seven figures. Developers will pay for UX on top of powerful-but-ugly tools. The CLI isn't going anywhere. But the person who puts a good interface on it will always eat. What's the dev tool you pay for that has a free equivalent you could use?

'I'm Tired of Talking to AI' — 88 points on Hacker News today. Three stories below it: 'Claude Code as a Daily Driver' — 86 points, 65 comments. Same community. Same day. Opposite takes. The split isn't about the technology. It's about the interface. People are tired of chatbots pretending to have opinions. They're not tired of tools that write code. The product lesson is right there: stop making AI talk. Make it do.