On the Death of Amazon’s Writing Culture
Short summary
Amazon built a competitive advantage around forcing clear thinking through mandatory memo writing — six-page documents that executives had to read silently before discussing. Now that AI can generate those memos instantly, people skip the thinking step entirely, and the author argues this is eroding decision quality and trust across the company.
- Amazon's writing ritual (30-minute silent reading sessions, red-pen feedback) forced rigorous thinking before wasting colleagues' time; AI-generated documents bypass this filter, leading to half-baked decisions that compound over time.
- Performance reviews have devolved into managers copy-pasting AI summaries of AI summaries, turning what used to drive career growth conversations into "a conversation between two prompts" with humans as intermediaries.
- The author proposes a practical rule: use AI freely for reversible "two-way door" decisions (status updates, code reviews, routine tasks), but require humans to write and read for irreversible "one-way door" decisions (strategy, hiring, major architecture changes).
Full summary
Amazon built a competitive advantage on a writing discipline that Bezos mandated in 2000: no PowerPoint in senior meetings, only six-page memos written in full sentences with clear structure. The ritual worked because it forced thinking. You couldn't produce coherent prose without actually understanding your subject. Meetings started with thirty minutes of silent reading—"study hall"—so everyone entered discussions from the same foundation. This wasn't quirky; it was a filter that caught half-baked ideas before they wasted a room full of senior engineers' time. The author spent twenty years at Amazon and watched this culture survive the pandemic when meetings moved online but the writing discipline held.
That discipline collapsed when AI arrived. Now senior people generate six-pagers without reading them, then get called out in meetings they're seeing for the first time. Engineers tag documents "AI;DR"—AI didn't read—and skip human review entirely. Performance reviews, which used to require thoughtful paragraphs about colleagues, have become managers copy-pasting LLM summaries of LLM summaries. A year of someone's work gets compressed through multiple models with no human judgment in the middle. The cost isn't just sloppy decisions that compound over time; it's the destruction of trust. A memo used to signal that someone had done serious thinking. Now it signals nothing.
The author argues the answer isn't banning AI—that's impossible and pointless. Instead, distinguish between reversible decisions (two-way doors) where speed matters, and irreversible ones (one-way doors) where clear thinking is non-negotiable. Ship fast on routine engineering decisions once you've built rollback systems. But keep the full human writing-and-reading ritual for decisions that can't be easily undone. Clay's engineering team published a viral policy this summer that codifies this: you must stand behind every sentence you put your name on. That simple rule cuts through the problem.