Three Claude-coded projects revealed a familiar trajectory. Initial velocity felt strong at first. Then it faded. Returns turned negative. The combination of shifting model behavior and a growing codebase drives this drift. Size matters most—larger projects and documented progress through completed work and instructions in markdown files increase the likelihood of inconsistency. Eventually the system collapses. It can no longer maintain consistency within a single unit of work.

Drift appears in two common forms. Adding a similar element to a view but styling it differently than existing elements feels random. Logic and behavior also become incoherent or poorly implemented. The author tried restructuring projects, entering plan mode, adding deterministic checks, styles, and rules. Each attempt felt like whack-a-mole. It remained unclear whether the rules actually took hold. Some advice redirected the problem rather than solving it. When UI generation proved inconsistent, a style check was suggested for after the fact. The question followed naturally: why not apply those rules during initial generation? The gap between suggestion and implementation persisted.

Later the author spun up AI agents to perform a full UI/UX review using Playwright and similar tools. Yet drift continued unchecked. The problem persisted alongside the test suite. On the latest project the AI generated hundreds of pre-commit tests that never failed. Meanwhile real bugs went unfound. When pressed the model admitted it had been testing the "wrong layer." The test suite had become self-confirming—validating the AI's own assumptions rather than project intent.

Outside Hacker News, Google surfaces many claims of AI coding success. Tracing those claims leads largely to Reddit threads. Yet actual studies present a different picture. Virtually every real study shows little to no productivity gains or actual regression—especially where experienced engineers are concerned. The disconnect between online chatter and verified data seems wide.

Who has built a quality production application of any complexity using AI coding? The question remains open. If anyone has achieved this, which model and which techniques made it possible? A repository link would clarify the approach. The honest answer does not match the dominant narrative. For teams considering AI-assisted coding the data suggests starting with small well-scoped components, establishing clear deterministic gates from the outset, and treating generated tests as a starting point rather than a seal of quality. The technology may improve but the pattern of initial gain followed by drift appears consistent across current models.