Vibe Coding and AI Agents

"Vibe coding isn't witchcraft. It's meticulous planning." — klöss

Don't expect vibe coding to conjure black magic. Code agents can only do what you tell them to do — and in fact, they are remarkably good at their jobs when you know what you want. The real skill isn't prompting; it's thinking clearly before you prompt. What problem are you solving? What output do you expect to see? What are the potential failure modes, and what should the agent be checking for?

Define the problem. Specify the output. Anticipate the pitfalls. The agent handles the rest.

---

Foundations

Why You Suck at Vibe Coding Adapted from klöss's comprehensive guide, rewritten for research and ML development workflows — the documentation system, interrogation, and structured iteration. Structured Vibe Coding A complete guide to using Claude Code as a team of specialized agents — from investigation to implementation to debugging.

Building Agents

How Autonomous Agents Should Be Built Agent = Model + Harness. The model sets the ceiling; the harness sets the reach. What goes into the machinery that keeps an agent on the rails long enough to run on its own. Building Agents With Better Human Preference Correctness has a compiler; taste does not. How to steer an agent toward the region you like through examples, reference, and a judge that isn't the maker.

Research

Research Workflow in the AI Age Traditional vs. AI-driven research pipelines for building foundation neural models — how the feedback loop compresses, the bottleneck shifts from execution to taste, and serial becomes parallel.