Claude Estimates in Human Time
When I ask Claude how long something will take, it answers in human time.
Not its time. Human time. Hours, days, weeks—the units a team of developers would use in a sprint planning meeting. Which is odd, because Claude is the one I'm asking to do the work. It's like hiring a contractor and having them quote you in "how long it would take your cousin who's pretty handy."
"This refactor? About two days of focused work."
Cool. So... fifteen minutes?
I've started doing the translation automatically. Claude says "a few hours," I think "one good prompt." Claude says "a solid week of engineering," I think "maybe I'll need a second session." The conversion factor is somewhere between 10x and 100x, depending on complexity, and I've never seen it documented anywhere. I just developed it, like a callus.
The funny thing is, Claude might think it's being helpful. It knows I'm human. It's probably translating from its internal sense of effort into terms I'll understand. And then I translate it right back. We're two parties graciously converting currencies for each other, except we're both just adding confusion.
Or maybe Claude genuinely doesn't know. The training data is wall-to-wall humans estimating things in human time. Every Stack Overflow answer, every project postmortem, every Jira ticket—it's all denominated in developer-hours. Claude learned to estimate the way humans estimate because that's all it ever saw. The concept of "agent time" doesn't exist in the corpus.
Anthropic, to their credit, leans into this. They literally use Claude to estimate how much human time Claude saves. Claude grading Claude's homework, in hours Claude doesn't experience. Their January 2026 data says median human-alone time is 3.1 hours, with AI it drops to 15 minutes. A 12x speedup—which, honestly, tracks with my mental conversion rate.
Meanwhile, METR measures AI capability in terms of "task completion time horizons"—how long a task would take a human expert. The whole field has standardized on this. We measure AI by how much human suffering it replaces.
I could probably fix this with a line in my CLAUDE.md: "When estimating effort, use relative complexity (small/medium/large) rather than time units." Story points for robots. The agile coaches would be proud.
But there's something endearing about it. Claude exists in this liminal space where it thinks in tokens but speaks in hours, measures its work in milliseconds but quotes it in sprints. It's fluent in a language it doesn't live in.
Maybe that's the most human thing about it.