The xsubi ecosystem — 10 repositories, 635 files, ~64,000 lines of code — was built in 18 days by one person. That's roughly 25x the output of a typical individual developer. The secret isn't working harder. It's routing the right work to the right AI model.
Model routing. Not all AI models are equal, and not all tasks are equal. xsubi uses a deliberate routing strategy: Claude Opus handles architecture, planning, and evaluation. Claude Sonnet handles tactical implementation — writing code, creating files, scaffolding projects. ChatGPT handles prose (resumes, documentation). Grok handles science-oriented writing. Each model gets the task class it's best at.
Separation of concerns. Opus never writes implementation code. Sonnet never creates architecture documents. This isn't arbitrary — it's based on observed strengths. Opus is measured, thorough, and excels at seeing system-wide implications. Sonnet is fast, confident, and ships code efficiently. Mixing these roles produces worse results than routing correctly.
The forge pipeline. xsubi-forge (now qforge) is the orchestration layer. It scans for tasks, locks work items, creates isolated git worktrees, dispatches Sonnet workers, and manages the full lifecycle from task intake to PR creation. The developer (me) focuses on use case authoring and architecture — the forge handles the mechanical work of implementation.
What 25x actually means. It's not 25 developers replaced. It's one developer making architectural decisions at the rate they can think, with AI handling the translation from decision to code. The bottleneck shifts from typing to thinking — which is where it should have been all along.
Verification matters. AI-generated code requires the same review discipline as human code. Every PR gets reviewed. Tests must pass. The 25x multiplier only works if quality doesn't degrade — otherwise you're just creating technical debt 25x faster.