Base44 Let Claude Fable 5 Rebuild Its System Prompt, Unsupervised, for Four Hours

Base44, the vibe-coding platform whose customers range from non-technical small businesses to teams building full SaaS products, handed Claude Fable 5 a job it had previously reserved for its most senior engineers: rewriting the system prompt that varies by hundreds of permutations depending on user tier and app type. According to Yoav Orlev, Base44’s first employee and current product lead, the model asked clarifying questions for about an hour, then ran unsupervised for four hours, delivered 90-95% of what was needed, and the result shipped that same afternoon via A/B testing. It also independently flagged a gap in Base44’s own eval suite — missing cache-hit testing — that could affect cost at scale. Orlev’s framing: it felt like working with a senior engineer, not a junior one, because earlier models got stuck and repeated the same failed fix rather than searching the codebase for where a problem was likely already solved.

That description of “senior-engineer-caliber” trust isn’t an isolated claim. The same week, Anthropic detailed two large-scale internal migrations run the same way: Bun co-founder Jarred Sumner used Claude Code to port Bun from Zig to Rust — roughly a million lines of code in under two weeks, with 100% of Bun’s existing test suite passing in CI before merge — and Anthropic Labs co-lead Mike Krieger migrated a Python codebase to 165,000 lines of TypeScript over a weekend using hundreds of agents and eight phase gates. Base44’s four-hour unsupervised run and Bun’s two-week migration are different scales of the same pattern: work that used to require a scarce senior engineer’s judgment call, delegated and checked rather than hand-written.

For companies evaluating how much scope to hand an agent, the operative detail in both cases isn’t the model finishing the work — it’s what happened around the edges: Fable 5 catching its own gap in Base44’s eval coverage, Bun’s port passing the existing test suite before anyone reviewed a line. The verification step, not just the output, is what turned “unsupervised” into something a team was willing to ship same-day.