Stavros Kasidis just published one of the most honest takes on LLM-assisted development I've read. After months of shipping real projects with AI, his conclusions are refreshingly practical.
The Key Shift
"I thought that I liked programming, but it turned out that what I liked was making things, and programming was just one way to do that."
His workflow has fundamentally changed: LLMs handle code generation at scale while he focuses on architecture and UX decisions. The critical insight? Engineering skills haven't become useless — they've shifted. You no longer need to write code correctly line-by-line, but you need to architect systems correctly.
What He's Built
- Stavrobot — A secure AI assistant (his alternative to OpenClaw) with persistent memory, sandboxed code execution, and Signal/Telegram/WhatsApp/email integration
- Middle — A voice note pendant that transcribes and sends to his LLM
- Pine Town — An infinite multiplayer drawing canvas
All real, maintained projects. Tens of thousands of lines of code. He hasn't read most of the code but knows the architecture intimately.
The Three Tiers of Trust
His observation on how LLMs have evolved is worth bookmarking:
- Early LLMs: Review every line
- Later models: Check functions
- Now: Verify "general architecture"
He predicts next year even that won't be necessary — but for now, you still need a human with good coding skills.
Security-Focused AI Assistants
Stavrobot is particularly interesting as a security-conscious alternative to OpenClaw. It runs everything in containers, keeps the host OS invisible to the AI, uses Unix permissions to isolate plugins, and implements a three-tier memory system (memories, scratchpad, database) to avoid context overflow.
The Takeaway
The "LLMs can't write real software" debate is over. The new question is: what's your role when the AI can code?
📚 Source
Original post from Stavros Kasidis's blog