What 900,000 Lines of AI-Built Code Taught Me About Being Wrong

Josh Anderson made his name in dev circles a year ago by publicly failing. He tried building a product, Roadtrip Ninja, using nothing but Claude Code for three months, and watched it collapse under its own weight around 70,000 lines of code — small fixes broke other things, prompts ballooned into essays, and he ended up unable to confidently change his own software. He blamed the tools themselves, arguing the context loss and architectural drift were baked into how these systems work, not something a future update would patch.

He was wrong, at least partly, and he's saying so in public. His new product, Reclaim, has crossed roughly 900,000 lines, almost all written by Claude Code, and it's still shipping cleanly — including a pull request that got zero comments from GitHub CoPilot's automated reviewer, which he initially assumed meant something had broken. Anthropic's November release of Claude Opus 4.5 addressed his complaint list almost item for item: consistent performance across 30-minute autonomous sessions, a jump past 80% on SWE-bench Verified, context compaction so long tasks stop hitting a wall, and a rebuilt Plan Mode that asks clarifying questions before touching code.

But Anderson is careful not to hand all the credit to Anthropic. The bigger shift, he argues, happened before Claude wrote a single line of Reclaim — he designed the architecture himself first, rather than letting the model improvise structure on the fly and auditing it after the fact. His current workflow reflects that: fresh sessions per task instead of dragging context forward, planning-mode conversations where he pushes back on thin two-option answers, and CLAUDE.md standards files scattered strategically through the codebase instead of one bloated master file. A custom validation skill checks every change against those standards before it goes to CoPilot's review, and a human still gets the final sign-off after all of that.

He's not declaring victory. The Stanford research he leaned on last year, showing AI coding gains falling off a cliff well before 100,000 lines, still holds up as a general pattern — he's just sailed nine times past the point where he personally hit that wall, thanks to better tooling and tighter process working together. His read: model improvements raised the ceiling, but owning the architecture is what let him actually build up to it. Skip the second part, he warns, and faster tools just help teams build a bigger mess faster.