Where Humans Should Spend Their Time

The bottleneck has moved to review

The writer hardly writes code by hand anymore. Most of their work is now reviewing. That includes pull requests (PRs, proposed code changes) from coworkers, and plan files written by AI agents. Because AI often writes both the code and the review replies, people end up relaying messages between AI tools. The writer asks where human time is best spent.

Ideas from other writers

Simon Willison says that since generated code must be reviewed, the real limit is how fast a person can review. Dan Shapiro's Five Levels model describes a path from simple autocomplete (Level 0) to a fully automated 'dark factory' (Level 5). At Level 3, people only read diffs (lists of code changes). At Level 4, people write specs and review plans instead.

Who should do which checks

The writer suggests a simple ranking. Human time is the most expensive, then LLM time, and machine checks are the cheapest. Anything a program can check should be handled by a program, such as the conftest policy tool in the writer's infrastructure repository. People and LLMs should only handle judgment calls. The writer also notes limits: a program may not catch every risk in application code, and if a person still approves every PR, fewer comments may not mean less waiting. Changing approval rules by risk level is a separate idea. Reviewing plans may be better than reviewing long code, because one bad line of plan can lead to hundreds of bad lines of code.

What to watch

• Whether program-based checks can judge risk in application code, not only infrastructure.
• Whether risk-based approval rules, like those in the Claude Code team's article, reduce review time without more incidents.
• Whether plan review works in practice, since plan files may be committed for a while and then cleaned up.
• How teams tell the difference between high quality and missing detection when a check finds nothing.