26 Repos in 29 Days With Claude Code: The Real Failures Weren't Code

Between June 5 and July 3, 2026, one developer using Claude Code shipped 26 repositories, 1,549 commits, and 335 live pages on a deliberately boring stack: static HTML/JS and Astro deployed to Cloudflare Workers, no backend, no database. The notable part isn't the throughput — it's which failure modes surfaced, because none matched the usual 'AI can't write code' narrative.

Where the pipeline genuinely excelled was classical algorithms: the 2048 solver runs a real expectimax search, benchmarked over 250 self-play games at a 69.6% win rate and ~0.5ms per move; tic-tac-toe uses minimax with alpha-beta pruning; pathfinding uses BFS — all textbook, all correct, all shipped in days. Written conventions in shared memory files also let dozens of pages stay consistent without drift.

What actually broke was structural, not syntactic: a new /word-tools/ hub cannibalized the existing /tools/ pages in search; inconsistent URL patterns (flat files vs. directories) caused canonical mismatches and redirect chains, consuming the sprint's two heaviest commit days; source repos and their production mirrors silently diverged; and even the AI-written audit tools produced false positives — phantom orphan pages, fake canonical mismatches — that would have broken a healthy site if trusted without checking the live URLs.

The token economics were the sharper lesson: roughly 93% of cost came from re-reading accumulated context in long sessions, not from generating new code — fixed simply by resetting context per task and using small memory files. Over 28 days the site drew 207 clicks from 7,320 impressions across 257 pages, with the top-performing page after the homepage being the 2048 solver — the one page with real engineering depth, not the highest page count.