Summary
The author argues that 90% of people using AI follow the same pattern (prompt → accept → ship) and produce indistinguishable output. The three durable moats in the AI age are taste (knowing what good looks like), distribution (earned trust at scale), and high agency (the drive to figure things out without a roadmap). These are the only dimensions AI cannot replicate because they require accumulated experience, consistency over time, and intrinsic drive respectively. The window to develop them before they become table stakes is roughly 12 months.
作者指出 90% 的人以相同方式使用 AI（生成 → 接受 → 發佈），產出千篇一律的內容。AI 時代三個持久護城河是：品味（知道什麼是好的）、分發（建立規模化信任）、高能動性（面對不確定性仍能推進）。這三者都無法被 AI 複製，因為它們需要積累的經驗、長期的一致性以及內在驅動力。建立這些優勢的窗口期大約還剩 12 個月。
Key Points
- AI defaults are convergent: shadcn/ui 90.1%, Tailwind 68.4%, Vercel 100% — models are probability machines defaulting to training data averages
- Consumer trust drops 33% when people learn content is AI-made, even when output is objectively better (+19% CTR)
- Taste = knowing what to reject, not just what to want; it’s built through deliberate study, not tutorials
- The 80/20 rule: let AI handle 80% (research, drafts, boilerplate, structure); apply human judgment to the last 20%
- Distribution moat: trust runs on a different clock than technology — AI can compress creation from days to minutes, but trust still takes months/years
- High agency: “a multiplier amplifies whatever you bring to it” — curiosity + AI = 10x leverage; passivity + AI = nothing
- A “taste skill” for Claude Code with 400 tokens of explicit design rules (specific fonts, colors, avoid-list) demonstrably changes output quality
Insights
- The study showing AI ads outperformed humans by 19% on CTR but dropped purchase likelihood 33% when disclosed is a striking data point: performance and trust are decoupled
- “Functional is free now. Remarkable still costs something.” — this is the core economic shift; the floor has risen, the ceiling has not
- The convergence problem is structural, not incidental: any AI trained on internet data will statistically default to the most common patterns, which is why taste skills that encode explicit anti-defaults are so valuable
- High agency as the foundational moat is underrated — you can develop taste and distribution, but without agency you never start; it’s the prerequisite that makes the other two accessible
- The author’s personal story (built more in 3 months than 2 years, then went back and found half was mediocre) is a useful calibration for anyone in the “intoxication phase” of AI adoption
Connections
- Taste
- AI Productivity
- Distribution
- Lessons from Building Claude Code How We Use Skills
- Skills
- Context Engineering
Raw Excerpt
“the tool isnt the advantage. everyone has the same tools now. the advantage is knowing what to do with them. and that requires something AI cannot give you: taste earned through deliberate study and honest self-assessment of your own work.”