Who owns the way an inspection is written?

WINspect is the product WIN Home Inspection's inspectors use to run an inspection and write the report at the end of it. Roughly 11,000 of them, across about 900 teams.

The remark library was large and it was genuinely good — expert-authored, reviewed, and structured to cover the range of things a house can be wrong about. The product exposed it the way the organisation understood it: category, then subcategory, then a list. The legacy system had 491 subcategories. The redesign had already cut that to 78, moving the leftover variation into contextual Abouts — attribute sets hung off each subcategory.

So by the time I picked this up, the tree had been pruned and search existed. Inspectors kept navigating anyway.

Most of this list is not design constraint in the abstract. It is the shape of a product mid-migration with a deadline:

An inspector standing in a property thinks: I saw a crack in a window. The library thinks: Exterior, then Windows, then a list of forty things. Those are not the same sentence, and everything hard about this project lives in the distance between them.

Navigating wasn't stubbornness or a habit nobody had broken. It was the only route that didn't require guessing what the system had decided to call the thing they had just looked at. Category navigation is slow and it is certain. A search box that comes back wrong twice makes slow and certain look like a bargain.

Which meant the fix I had been reaching for — better ranking, a more prominent entry point, fewer taps to the same place — was aimed at the wrong thing. None of it closes a gap between two ways of describing the same crack.

Search-first is one sentence: when an inspector knows what they observed, they should be able to start there. Not at the top of a tree, and not by first learning how the organisation files things.

The temptation with a principle like that is to enforce it — make search the only door and let the tree wither. I chose not to, and it is the decision I would defend hardest.

Making search fast enough to win on merit is a harder brief than making it mandatory. It is also the only version where the inspector stays in control, which is the whole ideology.

A plain list of matching sentences would have been a worse product than the tree. The tree at least told you where you were. So a result card carries what an inspector needs to judge it without opening it: the branch it was filed under, the status it will carry into the report, where it applies, the wording itself, and how often it has actually been used.

The branch line is doing more work than it looks. It is also the evidence for the whole approach — a set of results whose paths disagree with each other is a set no single walk down the tree could have produced.

Search-first has an obvious failure mode: it assumes someone is in a position to type. On a roof, in an attic, halfway through explaining a finding to the client standing beside them, they are usually not.

So the surface offers a zero-query path. Suggested observations come up before anything is typed — ranked by what this inspector actually reaches for, which makes the empty state the most useful screen in the flow rather than a blank waiting for input.

The same logic reaches into capture. Suggestions can surface while an inspector is photographing the thing they are about to write up, because that is the moment they know exactly what they are looking at. Search that only exists on its own screen has already made them leave the observation to go and describe it.

And all of it works offline. That was not a nice-to-have — it settled the hierarchy of the whole feature set. Anything an inspector depends on in a basement with no signal could not be the thing that needed a network.

A search that returns good results and then forgets what you did with them has solved a smaller problem than the one an inspector has. They are not browsing a library. They are assembling a document, in order, against a property.

So the surface separates what you can draw from and what you have already used: My Library against Added to Report. It reports where the current inspection has got to — remarks added, issues found — so search is attached to this job rather than floating above the product. Picking a remark is not a search result being dismissed, it is a report getting one row longer.

Downstream, those remarks become the report: gathered in the gallery against their evidence, promoted into the summary when they carry an issue status, and rendered into the document the client actually reads. Designing retrieval without that tail would have optimised the part of the job that was never the point.

Every ranking decision above is downstream of how the library is written, categorised and maintained. That work happens in the SME Toolkit, where remarks carry utilisation counts, company-key context, categories, locations, terms, and draft and template state — and where support and SMEs can see what is being used, by whom, and what has gone stale.

This is also where the ownership question gets answered properly. A search over one central library is a search over somebody else's words. So the sources are separated — an inspector's own library, the network's, and the SME-authored baseline — the filters are additive, and the selection persists. Any remark can be saved into an inspector's own library, where the saved copy becomes theirs to maintain rather than an edit to everyone else's.

The measure was never whether the search screen tested well. It was whether inspectors stopped routing around it — and that is a question only production answers, through what people search, what they reuse, what they shake the phone to complain about, and what Support keeps getting asked.

87% of remarks now enter a report through search rather than being written new. That is the number I would put the whole ideology on: not that search is used, but that the library is being reused instead of retyped, which is what the access layer was for.

AI earned a smaller role than expected, and the trend is the interesting part. It fills the gap when search comes back empty — describe the observation, shape the result, save it to your library so the next similar finding is a search rather than a generation. Its share of new remarks fell as the libraries matured. Not because generation got worse; because there was less left to generate.