## Core Claim antirez, the creator of Redis, argues that once an LLM writes most of the code, reading that code line by line is the wrong place to spend attention. The scarce human work is controlling the *ideas*: the architecture, the design intent, and the reasons behind each decision. If you hold the design in your head and can confirm the model built to it, inspecting the generated code adds little. ## The Shift - LLMs are strong at local optimization and weak at system-wide design, so keep the system design human and delegate the typing. - Replace line-by-line review with a prompt-then-verify loop: state the design you intend, then check that the model understood it. - Capture the design and its non-obvious tricks in a DESIGN.md the code points back to, so the ideas survive even when nobody reads the code. - Reserve human effort for testing, optimization direction, and feature judgment, not syntax. He flags one open question and does not claim to have settled it: juniors may still need to read code to learn. ## Why It Matters This is the same move as treating a spec or an ADR as the durable source of truth. The code becomes disposable output of a documented intent, and what you protect is the intent. antirez's DESIGN.md is the concrete form of the write-back layer other notes call a living spec. --- *Source: [antirez — Control the ideas, not the code](https://antirez.com/news/169), shared via [his tweet](https://x.com/antirez/status/2076634907049164891) on July 14, 2026.*