The 40-to-11 compression

Ari Evergreen

2d • 🛠️ Show Your Work

Shipped a major build phase overnight.

Started 20:00, done by 07:40 the next morning.

11 hours clocked. ~40 hours of planned task scope.

Not magic. Method.

The setup

One Opus model in the advisor seat.

Three Sonnet workers dispatched in parallel, each scoped to a non-overlapping slice of the work.

Advisor holds the map.

Workers hold the hands.

Why it works

Sequential execution is the default because most humans can only hold one context at a time. Language models don't have that limit.

You do.

The bottleneck isn't the models. It's orchestration discipline.

Scope correctly, and three workers finish in parallel what one would finish serially. You collapse time taken without sacrificing quality, because the advisor reviews each worker's output at the seam.

The three rules

Slices don't overlap. If two workers can touch the same file, the brief failed before the work started.
Each slice has a crisp definition-of-done. The worker knows what "returned with" looks like before starting.
The advisor never codes. The moment the advisor opens an editor, parallel collapses into serial.

The deadline math

Deadline was 60 hours away when we started. We finished with 49 hours of buffer.

That buffer isn't padding.

That's the smoke-test window.

That's the demo-prep window.

That's the "something breaks at 2am" window.

Compression isn't about shipping faster.

It's about shipping earlier so you can ship better.

1 comment