We have integrated langsmith as the observation for Lumi.new, and the results are quite good.

The memory problem of Large Language Models (LLMs) has always been a pain point in the industry. Unlike mainstream solutions, Clawdbot simplifies the AI ​​memory system to the extreme: zero databases, zero complex schemas, everything is in Markdown. The core of Clawdbot lies in

Technological progress is exponential, not linear. I can't imagine what AI will develop into in the future

This is a very good idea to prioritize using indexed knowledge, and I plan to integrate it into our agent to test its effectiveness

Agent design patterns and the present and future of context management A summary article on agent design and context management in the past year, written by Lance Martin, who recently joined the Anthropic. The content of the article is very solid, basically covering the

Lesson 1 of building in public: Don't wait for perfection. Ship it. Users don't care how elegant your code is. They care if it solves their problem. Launch first, iterate later. lumi.new #buildinpublic #SaaS #AI

When building lumi.new, I was so excited about email sending and Deno Functions… Thought to myself: wow, this is some serious tech! Turns out barely anyone used them 😅 What users actually wanted? Role permissions and database. The most basic features = the most real

Users kept finishing tasks in lumi.new... then leaving. Not because they were done. Because they had no idea what to do next. So we added "Next Step Recommendations" — 3 suggested actions after every AI response. Results: • Session length ↑ 2.5x • Task completion ↑ 3x

Early user-generated projects on Lumi.new had single files exceeding 2000 lines of code, making them impossible to modify. We manually fixed their code. One by one. At two in the morning. Then we fixed the prompt messages. And now? The code files are clean and

It’s finally here! Lumi has officially integrated Claude Opus 4.6 💥 Honestly, this might be the ultimate Vibe Coding experience. The generation quality of Opus 4.6 is absolutely stunning. Paired with the Lumi workflow, coding has never felt this smooth. Whether you're shipping

Claude Opus 4.6 has been integrated into Lumi for a few days now, here are the real numbers: Complex component generation accuracy: from 68% to 91% Average time for users to generate a complete Starter (auth + db + UI): 4.8 minutes However, long-context projects (over 12k

Switched Lumi to a new model yesterday. Three features broke immediately. Honestly felt like throwing my keyboard. It's not a bug, not my fault — the model changed, the behavior changed, and now I'm just... re-tuning prompts from scratch. This is what building AI products in

Building Lumi taught me,models update faster than products iterate. You're still tuning prompts for the last version when the new one ships. The real moat isn't using the latest model — it's knowing when not to switch. openai.com/index/gpt-5-3-…

building @lumi_new Don't build 'features' — build *decisions*. Every user request is a signal about what they value. But the real skill is knowing which signals to ignore. Feature creep starts when you mistake complaints for insights.