Task estimator

Estimates from your measured history, not from the model's intuition about how hard something looks. The model's intuition is the thing that is broken: developers using AI tools believed they were 20% faster while measuring 19% slower (METR 2025). An estimate that is not anchored to recorded actuals just reproduces that gap in writing.

So this skill does three things: it keeps a ledger of what tasks actually cost, it quotes new tasks against the empirical distribution of similar past tasks, and it scores its own past quotes so the bias is visible.

Depth on the evidence is in references/: estimation-science.md (the classical results — EBS, reference classes, anchoring, flow metrics), ai-era-evidence.md (what changed once an agent was in the loop), measurement-model.md (the phase taxonomy, the measured constants, and the calibration math the scripts implement). Read them when a rule here looks arbitrary — every rule traces to a result there. Do not re-derive the theory during an estimate; execute the steps.

Pick a mode