Developer says partner's 5,236-line AI-generated payments PR was never tested once

What happened

A developer has published a firsthand account of what happens when AI-generated code ships without human verification — and the subject is a payments backend, not a side project. Writing on Bear Blog in a post that reached the front page of Hacker News, the author describes discovering that a business partner had built the product's entire payments module in a single day using an AI coding assistant, delivering it as a pull request with 5,236 added lines, documentation included.

According to the post, the two had agreed to build that part of the system together: the author handles the frontend, and the payments work was meant to be a joint effort. Instead, the partner worked alone with a model the author characterizes as an agreeable assistant that never pushes back — in place of the one person who would actually consume the code and knew what the API needed to do.

When the author finally pulled the branch and exercised the endpoints, the module did not work. The partner had never run them. By the author's estimate, testing would have taken roughly two minutes, since the generated endpoints had already been collected in Bruno, the team's API client.

The fingerprints of unreviewed generation

The documentation accompanying the pull request carried recognizable markers of machine-written text: it explains basics such as an error producing a 400 response — information any working developer already has — while omitting what a human would actually need, like where to obtain API keys. The author reads the redundancy as output written for an unknown audience, and the gaps as material nobody reread, inferring that the partner never opened the document his own tooling had produced.

A warning that predates the pull request

The author says this outcome was foreshadowed. At the start of the project, they warned the partner about letting AI assistance stand in for real programming and got confident reassurances in return. Their argument is an atrophy analogy: just as muscles waste away during long bed rest, a developer's ability to do the work degrades without use, and the untested pull request was the moment the weakness showed — like a sedentary player who finds himself out of breath two minutes into a pickup game.

The post ends on a resigned note. The author suggests this may be their last project after this one, half-joking about retraining as a yoga instructor, and mourns that peers who enjoyed programming by hand have largely moved on to prompt-driven workflows.

Why it matters

This is one anonymous anecdote with no independent verification, and the author's broader claim — that AI assistance erodes cognitive ability — is an opinion, not a measured result. But the concrete detail is instructive: money-handling code, 5,236 lines, zero endpoint tests, and documentation nobody read. AI assistance collapses the cost of producing code; it does nothing to reduce the cost of verifying it, and payments backends are precisely where unverified code fails expensively.

The failure was caught only because a co-founder happened to spend two minutes running the endpoints. A solo team, or one where the other partner was too busy to check, could plausibly have merged it. Whatever position a team takes on AI-generated code, the operational lesson here is uncontroversial: generated output needs the same gates as human-written code — tests, review, and at least one person actually running it — and auto-generated documentation still needs a reader. There is also a subtler cost on display: the partner skipped the collaboration itself, replacing a colleague with real requirements and context with a system that agrees with everything. The fastest way to produce code turned out to be the slowest way to ship working software.