Estimate

Models misjudge time in both directions. They overestimate their own runtime (median 6.1x for the worst frontier model tested, 5x or more on multi-step agentic runs), recall elapsed time off by 1.7-100x, and default to human-scale minutes for second-scale work (Garikaparthi 2026, arXiv 2604.00010). On human work, Claude's ticket estimates rank tasks about as well as developers do (rho 0.44 against 0.50) but compress: short tasks come out too long, long tasks too short, and ten in-context examples did not fix it (Anthropic, 2025-11-25). So an estimate is built from parts, labeled with its basis, and checked against a log.

Rules for every estimate

Pick the clock. Agent clock or human clock, never both in one number. A number states its unit and its clock.
Give a range: low, expected, high. A point value hides the spread.
Label the basis. measured n=12 means 12 finished rows of this kind and type in the log. assumed means a default or a judgment. Below 5 measured rows the basis is assumed.
Do arithmetic with the script or a calculator, never inline. Complexity labels ("simple", "trivial") do not move a number: count rounds, modules or assumptions instead (one frontier model scored 18% on pairs where the label misleads, in the paper above).
Log it. Before the work starts, run start. When it ends, run finish or actual. See the log.

Pick the branch

Agent runtime (how long will this agent run take?): read agent runtime.
Human dev effort (ticket, feature, how long would a developer need?): read human effort.
Client quote (effort that becomes a price): read client quote.
Reply to the user (stating any estimate in chat): use the format below.