How to Work and Compound with AI
Eugene Yan
The article describes practices for working effectively with AI models by organizing context as infrastructure, encoding preferences as configuration files, building verification systems, and delegating increasingly larger tasks. Key concrete details include maintaining directory structures like ~/src and ~/vault, creating per-project CLAUDE.md files as behavioral contracts, and running three to six parallel sessions simultaneously. As a result, workflows become more efficient and scalable, with the bottleneck shifting from task execution to writing clear specifications and reviewing outputs.
Why it matters
Context as infra, taste as config, verification for autonomy, scale via delegation, closing the loop.
