What no relay settles
Work That Should Not Be Split Across the Gap
Splitting work between two halves that never overlap works where the work has a natural seam, and produces compounding delay where it does not.
three things · no nights · three that no number of nights settles
Three things, no nights, and no mechanism proposed for any of them.
Five pieces of work, split across the two halves, and what the split cost each one
Split cleanly; each half built one side and they met once
Not really split: one person did it and the other reviewed
Six passes, and each person restarted from the evidence
Five passes; the design got longer and no more decided
Cannot be split at all: it is one person's attention or it is nobody's
five things · 13 nights · one that no number of nights settles
Five pieces of work, thirteen nights, and the two cheap ones were the two with a boundary in them. A split works where the work already has a seam; where it does not, splitting invents one, and the invented seam is where the nights go. One team's own quarter, logged as it went.
Dividing work between the two halves of a team is the obvious thing to do with two halves. For some work it is straightforwardly right.
A reliable handover needs more than a message saying that something is in progress. read the documentation can add verifiable hours and task context to the process examined in Work That Should Not Be Split Across the Gap, reducing the need for another round trip.
A useful external reference is QuickBooks' official site, which provides another official view of how teams organise shared work.
For other work it converts a two-day job into a two-week one, and the conversion is invisible while it happens because every individual handover looks reasonable.
Work with a seam
Some work has a boundary already in it: two components with an interface between them, a piece of analysis and a piece of implementation, a thing to build and a thing to verify.
Where such a seam exists, splitting along it is cheap. Each half works on its own side with few dependencies, the interface is agreed once, and the crossings are infrequent and substantive.
The test is whether each half can work for a day without needing anything from the other. If yes, split it. That is the whole criterion and it is reliable.
Work without one
Where there is no seam, splitting means alternating: one person advances the thing, hands it over, the other advances it, hands it back.
Each pass costs a night and a reconstruction. The reconstruction is the expensive half: picking up somebody else's partly formed thinking takes between twenty minutes and an hour, and it is paid on every pass.
Six passes on a single investigation is six nights and perhaps four hours of reconstruction, for work that one person would have finished in a day and a half. The arithmetic is not close.
The three kinds that resist splitting
Investigations before they have a hypothesis. Almost pure mental model, and the model does not survive a night.
Design before anybody has a settled view. The thing being passed is not an artefact but a process of making up one's mind, and alternating that between two people produces two half-formed views rather than one formed one.
Anything requiring continuous judgement — watching a rollout, holding a situation, deciding when to intervene. These are not splittable in any amount; they are one person's attention or nobody's.
- Investigations before they have reached a hypothesis.
- Design before anybody holds a settled view.
- Anything requiring continuous judgement about when to intervene.
Assigning by half rather than by seam
A specific organisational error worth naming.
Teams often distribute work by balancing load across the halves, which is reasonable from a staffing perspective and ignores the seam entirely. The result is that work with a natural boundary gets given to one person while work without one gets split.
Assigning by shape rather than by load produces a better calendar and an uneven-looking allocation, and the uneven-looking allocation is the correct one.
Letting one half own a whole area
The structural version, and the one most teams with a wide gap end up at.
If each half owns areas rather than tasks, most work has no reason to cross at all. The crossings that remain are at genuine boundaries, which is where they are cheapest.
The costs are real: knowledge concentrates, one half becomes unable to cover for the other, and the two can drift into different ways of doing things. Those are manageable with deliberate effort — rotating ownership slowly, documenting at boundaries — and they are smaller than the alternative.
What does not work is the middle position, where ownership is nominal and everything gets split anyway.
Deciding in advance rather than discovering
The expensive version of this is discovering at the fourth pass that something should never have been split, by which time a week is gone.
The cheap version is the question at the start: can each half work on this for a day without needing the other? It takes a few seconds and catches most of it, because the work that fails the test usually fails it obviously.
The idle-half objection
The common objection to not splitting is that it leaves one half with nothing to do on that piece of work.
That is true and it is not a cost, provided there is other work. A half working on something else entirely is productive; a half alternating on something that does not alternate well is producing nights.
The objection is really about the appearance of collaboration, which is a different thing from collaboration and should not be bought at the price of a fortnight.
Making the seam explicit when there is one
Where work does divide cleanly, the division is worth writing down rather than assuming.
The useful artefact is a sentence about the interface: what each half is responsible for, what crosses between them, and what each can change without telling the other. It takes a few minutes at the start and it prevents the characteristic failure of a clean split, which is both halves quietly changing the shared part in different directions.
Without it, the seam is maintained by memory, and memory across a gap is two memories. The first time they differ is usually the first time anybody notices there was a seam at all.