Counting
What Happens When You Set a Target
A documentation target produces documentation-shaped output, in the required quantity, read by nobody. The mechanism is not cynicism.
What nobody wrote down here, and what it cost
That the runbook saved somebody a morning
Nobody told the author
Six occasions, none reported back
That a note ended an argument
The author never heard
Both times from people who joined later
two things nobody wrote down · 8 occasions it cost somebody something
Two things, eight occasions, and the author heard about none of them.
Before adding another required page, managers can see the workflow and check whether the underlying bottleneck is missing knowledge, overloaded authors, or unclear ownership.
A useful independent point of reference for this part of the process is Screaming Frog.
One team asked for a fixed number of pages per person, with what each row produced
Pages written to meet the target
180 / 24Twelve pages across six people; twenty-four readings between all of them
One page written because somebody needed it
8 / 304Eight minutes, more readings than the twelve target pages combined
A long page split into four to count as four
0 / —No new writing at all; the count went up by three
The quarter after the target was dropped
95 / 442Six pages, all written in response to questions
four documents · 283 minutes written · 770 minutes saved across 59 readings · one with no saving claimed
Four rows, 283 minutes written, 886 minutes saved across 59 readings. The first row cost 180 minutes and produced twenty-four readings. The second cost eight and produced nineteen. The third is what a count does to behaviour when nothing else changes. One team's own two quarters.
The obvious response to a team that does not write is to require that it does. Two pages a quarter, a documentation item in every sprint, a percentage of time.
It works, in the narrow sense that the pages appear. It does not produce anything anybody reads, and the reason is not that people are gaming it.
Why the pages are bad without anybody cheating
A person required to produce two pages has to choose a subject before anybody has asked a question.
Without a question, the only available organising principle is the subject itself, which produces exactly the documents this collection has been arguing against: the overview, the glossary, the guide to an area. They are written honestly and read by nobody.
The target also sets the stopping point at the page rather than at the answer, so everything comes out page-shaped — the four-line note that would have been most useful is not a page and does not count.
What the number does to the shape
Three predictable distortions, and none requires bad faith.
Length converges on whatever reads as a page. Shorter feels like cheating.
Subjects converge on what can be written without consulting anybody, because consulting takes time the target did not allocate.
Timing converges on the end of the period, when the material is coldest.
The splitting problem
The third row of the panel is somebody dividing an existing page into four, which raised the count by three and changed nothing.
It is easy to read this as gaming and it is better read as a signal. The person had already written the thing; the target was asking for a quantity they had already supplied in a different shape, and the only way to express that was to reshape it.
A measure that can be satisfied without any new work is a measure aimed at the wrong quantity.
Why the measure cannot easily be fixed
The obvious repairs all fail.
Count readings instead — and pages get linked from high-traffic places, titles get broader, and long pages get split so each accumulates its own opens.
Count distinct readers — slightly better, still gameable, and it penalises specialist pages that three people need badly.
Require review — adds a gate, slows everything, and produces reviewers who approve things to clear their queue.
Any number attached to a consequence will be produced. That is not a flaw in these particular numbers.
What to do instead
Ask for specific things from specific people, which is the first item in the previous page.
"Four lines on the deploy order by Friday" cannot be gamed, because it names the output rather than counting outputs. It also cannot scale into a policy, which is a genuine limitation and is the price of it working.
Managers who want a mechanism rather than a series of requests are asking for the thing that does not exist.
Measuring without a target
The numbers are still worth having, used differently.
Who wrote what, as a diagnostic for the manager — not reported, not compared, not targeted.
How often pages are opened, as private feedback to authors — the subject of an earlier page in this section.
Which questions stopped being asked, which is the only measure that tracks the thing anybody actually wanted.
None of those has a consequence attached, which is precisely why they keep meaning something.
What any number does to the shape
Three distortions follow from a target, none of which requires bad faith from anybody.
- Length converges on whatever reads as a page, because shorter feels like cheating.
- Subjects converge on what can be written without consulting anybody, because consulting costs time the target did not allocate.
- Timing converges on the end of the period, when the material is coldest.
Each is a rational response to being measured. Together they produce documents organised around the subject, written from memory, at the wrong length — which is a description of exactly the pages nobody reads.
The quarter after
The last row of the panel is the team after the target was dropped: six pages, all written in response to questions, more readings than the twelve target pages produced.
That is not an argument that targets make things worse than doing nothing. It is an argument that the target was answering a question the team had not asked — how do we produce pages — rather than the one that mattered, which is how do we answer the things people keep asking.