Timers

A running timer records duration rather than reconstructing it. It is the cheapest accuracy improvement available and it has its own failure modes.

What it fixes

Delay disappears. The record is made while the work happens.

Fragments survive. A twelve-minute interruption is recorded as twelve minutes rather than absorbed.

Totals stop being reconciled, because they accumulate rather than being composed at the end.

The distribution becomes realistic, which is visible immediately: real days are messier than reconstructed ones.

How it fails

Forgetting to start. The work happens and nothing is recorded.

Forgetting to stop, which is worse: a timer left running overnight produces a fourteen-hour entry that someone then corrects by guessing.

Switching cost. Moving between tasks means stopping one timer and starting another, and under pressure people stop doing it.

Meetings, where starting a timer is socially awkward and frequently skipped.

Fragmented days defeat it the same way they defeat everything else: a person interrupted constantly cannot operate a timer per interruption.

Making it work

One keystroke to start, from wherever the work happens.

Automatic stop suggestions: a timer running past a threshold, or past the end of the working day, should prompt rather than accumulate.

Idle detection that asks rather than decides. "You were idle for forty minutes — keep, discard, or assign elsewhere?" is helpful; silently deleting the time is not, and silently keeping it is worse.

Easy retrospective correction, with the correction logged.

Calendar integration for meetings, so the awkward case is handled without anyone starting anything.

The idle detection boundary

Worth separating, because this is where a timer becomes a monitor.

Detecting that input stopped is a duration question and is defensible.

Detecting what the person was doing instead is activity monitoring.

A prompt asking the person to classify the gap keeps the judgement with them, which is both more accurate and the version that does not require the surveillance analysis.

Do not log which application was in focus in order to decide whether time counted. That is the line, and products cross it in the name of accuracy.

Rounding and increments

Record to the minute. The increment is a billing decision applied later, not a recording decision applied at capture.

Systems that round at capture lose information permanently and make the neutrality test impossible.

What to measure

Proportion of time captured by timer against entered retrospectively.

Corrections made after the fact, which indicates how well the timers are working.

Entries longer than a working day, which are forgotten stops.

Median entry length. A median of four hours means the timer is not being switched; real work is more fragmented than that.

The forgotten stop

The failure that produces the worst data, and it is detectable.

Query entries longer than a working day.

And entries ending exactly at a round hour, which are usually corrections.

Prompt rather than accumulate: a timer running past a threshold should ask.

Prompt at the end of the working day, which catches the overnight case before it becomes a guess.

Make the correction easy and logged, because the alternative is a person inventing a number to replace a wrong one.

Idle detection that asks

The design choice that keeps a timer from becoming a monitor.

Detecting that input stopped is a duration question.

Detecting what the person was doing instead is monitoring.

A prompt — keep, discard, or assign elsewhere — leaves the judgement with the person who has the knowledge.

Silently deleting the time is unpaid work; silently keeping it is inaccurate.

Never log which application had focus in order to decide whether the time counted. That is the line, and products cross it in the name of accuracy.

Check the difficult case

Use see how the workflow works to frame one representative test for this issue. The useful evidence is not the marketing description itself but the record created when a worker corrects an entry, a manager reviews it and an administrator exports it.