The post is a blunt plea for smaller pull requests. It argues that AI agents can now generate thousands of lines of code in one shot, but that does not make those changes reviewable. The complaint is not just about line count. It is about authors treating review as a mandatory gate instead of a handoff to another human who needs a clear story, sane scope, and enough context to judge risk. People strongly agreed with that framing. The center of gravity was that AI has made code generation cheap, while understanding, sequencing, and verifying changes is still expensive and very human.
The most useful comments pushed the conversation past "big PRs bad" into workflow. A lot of people said small PRs were always extra work, even before AI, and that many engineers only know the right decomposition after the feature works end to end. The practical answer was not "design every slice perfectly up front." It was to let the model or the engineer build a rough whole, then do a second pass to split it into coherent commits or
stacked PRs that each build, explain one idea, and can stand on their own. Several people said this already works if you invest in prompts, local tooling, or task systems that force feature slices instead of generic task lists. Others were harsher. If the author cannot review and explain the change themselves, they are outsourcing system understanding to the model and hollowing out the team’s knowledge base.
That led to a broader point about what review is for. In open source, maintainers can just say no. Inside companies, approvals often exist because of compliance, release process, or blame assignment, so teams end up pretending to do review while rubber-stamping giant changes under deadline pressure. Many commenters saw the real failure as organizational. If leadership wants AI speedups but keeps the same human review model, review becomes the bottleneck. If leadership drops review without replacing it with stronger validation, ownership and code quality collapse. The practical middle ground people kept returning to was early socialization of design, feature flags, stacked PRs, and stronger standards for both code and
PR descriptions. The mood was frustrated, but not anti-AI. It was anti-slop, anti-theater, and very clear that faster code generation does not remove the need for human judgment.
getsmall.xyz