John Malone · Human-directed, AI-drafted

Brief mode: condensed view. Switch to Full for persona notes and full analysis.

Context: A photo search during a prototype session exposed the gap between how people remember and how software expects them to ask. Rowan, Seton, Campion, and Devil’s Advocate are AI personas used as editorial lenses; John is the human author.

I Remember the Place. The Computer Wants a Filename.

A photo search during a prototype session exposed the gap between how people remember and how software expects them to ask.

The computer had the photographs. I had a memory. Naturally, this became a translation problem.

During a prototype session, I wanted to put the original photograph at the beginning of a slideshow of AI artwork. I could describe the relationship between the pictures. I remembered where the photograph was taken. What I could not produce on demand was its filename.

This struck me as a reasonable division of labor. The machine is rather good at retaining filenames. I was there when the picture happened.

The request, stripped of its private details, was simple: find the original photo behind these pictures; I remember where we took it.

Use the clue that exists

That sentence contains useful information. There is an existing slideshow. There are images to compare. There is a place that can narrow the search. There is also an implied distinction between an original photograph and the artwork derived from it. Asking me to start over with a filename throws most of that away.

In this session, an agent used local photo metadata to narrow the candidates, then compared the available pictures visually. A likely match emerged. So did an adjacent near-duplicate.

That last detail matters. Finding a photograph that looks right does not prove it was the exact file used to generate an artwork. We had a strong visual candidate, but no recorded link establishing which source file had entered the generation process. The useful answer included that uncertainty.

This was an agent-assisted retrieval exercise. Alcove Home does not yet provide that whole workflow as an automatic photo-search feature. The session gave us something better than a feature name: a concrete example of what the product should eventually handle, and what it must be honest about along the way.

I want personal search to start where my memory starts.

Sometimes that is a place. Sometimes it is a relationship: the document that went with the other document, the photograph before the edited version, the thing we were discussing when another thing happened. Those are hypothetical examples, but the product question is quite practical. Can the software use the context already available instead of making the person reconstruct a filing system?

Keep the uncertainty visible

A remembered clue should narrow a search. It should not become permission to invent the missing parts. If the place information is absent, say so. If only a preview is available locally, make that visible. If two pictures are nearly identical, show both and explain why the distinction remains unresolved.

The answer also needs somewhere to land. A thumbnail and a confident sentence are not enough when I want to use the photograph. I need to open the candidate, inspect it, and decide what happens next. The search should lead back to the material it found.

A collection you can actually open

For Alcove Home, local access is part of that ambition. I want to search a prepared collection of my own records without assuming that a network connection will be available. That requires a plain account of what is actually on the device. A picture the system knows exists and a picture I can open offline are two different things.

The prototype session ended with a selected photograph in the derived slideshow. It also left us with an unresolved question about exact generation provenance. Both belong in the account of what happened.

The engineering work now is to turn that kind of assisted investigation into a dependable product experience: carry useful context forward, retrieve candidates, show the evidence, and leave uncertainty visible.

I can probably learn to remember more filenames. I would prefer that the computer learn to make better use of the clue I already gave it.

Conversation notes: September 26, 2026. Published September 28, 2026.