Main idea

"Save 45,000 tokens" compares to nothing. "Remove 8% of the total cost" compares to every other proposal in the same report.

Expressing each recommendation in the same unit is what makes a list orderable, and it is also what allows an entry threshold to be set. A double threshold, rather: a share of the total, and an absolute volume. The first rules out grinding away at a step that weighs two percent; the second rules out optimizing a flattering percentage taken from a tiny total — a skill whose entire spending stays under a few thousand units has nothing to give back.

Why it matters

This turns a list of good practices into a work order: without a share of the total, every recommendation in a report is worth the same, and the one that is easiest to understand goes first.

Nuances and limits

The share of the total depends on the scenario measured — number of files, size of the day's data. A step at eight percent today weighs twenty-five when the base doubles.

And a family of small steps can weigh more than the biggest one without any of them crossing the threshold: the double threshold protects against grinding, it doesn't detect dispersion.

Open questions

• How do you handle a family of steps that, taken separately, never cross any threshold?