🚫 Say no to vibe-coding — be the Manager of AIs instead 🤖
19 May 2025I’m convinced that AI tools will soon dominate most engineering workflows, not by replacing engineers, but by becoming the tool we all need to master.
Like any tool, you don’t get good at it just by reading about it. You practice! So I set myself a new challenge:
How far could I go in a single day(ish), using only an LLM, to build something I actually needed?
🛠️ The result:
✅ A complete library to control MKS Servo 42D stepper motors over CAN
✅ Full test coverage
✅ A working simulator
✅ Lots of documentation
Here’s what I learned:
1️⃣ Don’t just vibe-code.
Manage the AI like a team of junior devs: they are fast, helpful, but directionless without your leadership.
2️⃣ Start with crystal clear goals.
If you don’t know what “done” looks like, the AI won’t either 😬.
3️⃣ Use AI for brainstorming and planning, not just coding.
Ask it to explore ideas, structure the repo, and outline the roadmap before writing a single line. (Re)Upload those docs to the LLM whenever needed.
4️⃣ Set standards from the start.
I asked the LLM to follow the Google Python Style Guide, suggest best-in-class tools, and write tests with extensive comments to explain the reasoning behind each one. That set the tone for a clean, maintainable codebase.
5️⃣ Context window is queen.
Break the work down, just like you would for any task. Start fresh sessions when things get fuzzy (or “bug-loop”) and treat them like clean branches.
6️⃣ Be the boss.
Don’t settle for “almost right.” Push for the exact result you need. Ask again. Iterate. Good results come from persistence.
This wasn’t magic or snake oil at all! But the result of a solid strategy, clear structure, and learning to treat the LLM like a powerful collaborator.
Next up: real motors! Time to chase the bugs hiding under all that suspiciously perfect-looking code. Because let’s be honest, who really writes flawless code on the first try? 😄