Find the real constraint
Code is cheap now. What slows an organization down is the system around the code: how it decides, defines and approves work. That is where I start.
The problem
When an engineering organization is slow, people reach first for a better model, more headcount or more licenses, and those are usually not the answer. Code is cheap now. Two things are scarce. The first is deciding what to build and writing it down as a spec with tests, so the result can be checked against it. The second is doing that check once the work is built.
When a team looks like it is failing, the cause is almost always the operating model, not the people.
How I approach it
I start by measuring the system, not the people.
- Waiting against building. Every piece of work records the time from brief to production, split into waiting and building. If waiting is more than half, that is the main finding.
- The full cost of an outcome. People time plus AI spend, counted together. When I claim a gain, I count the time it gave back to the team minus what it cost to get.
- Quality next to throughput. Next to every throughput number I show two quality numbers. The regression rate is how often a change breaks something that used to work. Time-to-green is how long it takes to get the tests passing again after they fail. They sit beside the throughput number because a team can ship more by shipping worse, and more code does not mean more value.
- What I refuse to count. Lines of code, story counts, ticket counts, and scores on individuals. I report trends rather than snapshots.
- Numbers computed by code. Each number has one written definition, and code computes it, so nobody can argue a number up or down by hand.
- Evidence a skeptic accepts. I only put a number in front of a skeptic if it would convince one. Early results only show a direction. When something else changed at the same time, for example a team got the new tools in the same quarter it gained two engineers, I say so before anyone else does, because either change could explain the number.
What it produces
A measured answer to where the organization’s time goes, before it buys anything.
In the Luci House
The AI-staffed team I run every day as a working prototype.
Every step on a request is dated, so the time it takes is measured, not estimated: a median of 3.6 hours from the engineer's first note on an item to live. The numbers are counted by a script from the house's own records, each with one written definition, and the chart of commits each week is drawn with the verified releases of the same week beside it. How work moves
Figures as of November 20, 2026.