Feedback loop is all you need
Agentic coding works when the loop has ground truth. Prompts help, but verification is what makes the system improve.
blog.michaelzuo.vip
Notes on AI systems, software engineering, and the operating taste required to ship real work with agents.
$ load latest-thinking
ok 13 posts indexed
ok 8 topic filters online
new Feedback loop is all you need
Agentic coding works when the loop has ground truth. Prompts help, but verification is what makes the system improve.
Correlation explains why AI coding can feel magical and unreliable: the system, not the model alone, turns probability into control.
Capability rankings keep changing. If switching the leading model is hard, the bug is your architecture, not your prompt.
Prompt engineering tunes the mean. Narrowing the distribution takes a different toolkit — one that came out of factories a hundred years ago.
The model always hands you a systematic-sounding story. It usually isn't the truth. What FBI cognitive interviewing teaches you to do next.
Four rules from Karpathy's skills repo — and why they're the same rules you'd teach a junior engineer.
AI creates more jobs than it destroys — history proves it. Here are the 5 skills that matter for what's coming.
What if your project could watch the open-source community, learn from it, and upgrade itself — without you doing anything?
I gave my AI agent a design language and 6 core pages. It built 20+ production pages on its own. I didn't touch a thing.
Figma Code Connect + custom strategies working in layers — combining the best tools to maximize AI coding pipeline output.
AI doesn't replace engineers — it replaces the version of engineering we grew up doing. Here's how to make the shift.
Don't ask designers to change how they work. Absorb Figma's messiness in the engineering layer instead.
Velocity measured effort. FPI measures trust. Here's why it's the only metric that matters when AI writes your code.