I run Claude Code daily. I have for months.
My subjective experience: I move faster. Ideas hit the editor quicker. Boilerplate disappears.
My measured output: harder to quantify than I expected.
That dissonance is the whole story.
METR ran a controlled study — 246 real tasks, 16 experienced engineers. Developers were 19% slower when using AI tools. They believed they were 20% faster. The perception-reality gap was not small. It was 39 points in the wrong direction.
Faros AI tracks engineering metrics across thousands of real teams. 93% AI tool adoption. ~10% org-level productivity gain.
No one is lying. The speed feeling is real. The output gap is also real.
Here is what I think is happening:
AI tools optimize for the phase of coding that feels hardest — the blank page, the boilerplate, the lookup. They don't optimize for the phase that actually compounds — review, refactoring, architectural coherence.
So we ship faster. We review less. We accumulate debt we don't see yet.
RDEL's taxonomy of AI debt names three distinct types: technical debt (duplicated code, security holes), cognitive debt (code you no longer fully understand), and behavioral debt (systems that drift from their specifications).
The teams ahead of this curve are not the ones with the most AI tooling. They're the ones who treat AI output as a first draft, not a final answer.
Speed is the input. Understanding is still the job.
Where does your team sit on this gap — measuring output, or measuring speed?