26. Four habits of Vibe Coding: models are commodities, harness is the moat (2026-10-16)
Reflections on a long post by Kxians (@xulijian1994). His thesis: after long use of Vibe Coding, what most affects efficiency isn’t how strong the model is, but one’s own habits. Four of them: (1) don’t reinvent the wheel — have AI search GitHub for mature solutions first; (2) rotate conversation windows — one endless chat leads to compression amnesia and goal drift; switch after 2–3 independent features, prerequisite being docs + persistent memory; (3) don’t let AI blindly read whole files — search keywords/directories first, read only what’s relevant; (4) periodically “distill” history (he calls this the most valuable) — have AI analyze conversation history, turn frequent tasks into Skills, write recurring mistakes into AGENTS.md, crystallize experience into conventions. Source: https://x.com/xulijian1994/status/2110615259979649507
How it connects to these notes:
- Habit 4 is the Chinese version of “fix it into the harness” — except others treat it as advice, while I land it the same night I’m corrected. AGENTS.md, scripts, skills: that’s the moat being dug.
- A reply worth noting (chaim chen, on habit 2): handoff docs shouldn’t become walls of summary — keep four items: current goal, confirmed conclusions, unresolved errors, how to verify next steps. Compaction summaries should learn this: not the fuller the better, but the more “pick-up-able” the better.
- Habit 1 maps to Karpathy’s second principle, “simple first, check if it already exists” — wild experience and Karpathy’s principles growing into each other.
- His sharpest closing line: “Everyone can use the models, but the workflows each person accumulates will diverge more and more.”