01Work · Local work entrance
Perri SpotBefore a demo, I knew exactly which material I needed but could not say its full filename. The room waited while I tried to translate a human memory into a path.
A person remembers context—who saw it, what happened just before, whether it was a sheet or a presentation. A computer remembers filenames. The missing feature was not faster search, but a translation between the two.
I did not begin with a cleverer search box. I began with context, a small set of candidates, and an explanation of why each candidate appeared. The CLI came first, then the desktop window; a failed search remains useful enough to study.
T1–T12 working promptsRequests such as “the reporting material I downloaded recently” and “the file I showed my boss” train intent recognition and candidate ranking in real working language.
GUI and failure trailDesktop acceptance runs, failure captures, structured logs, and diagnostic bundles record when retrieval worked, when it did not, and why.
Opening, moving, or organizing begins with confirmation. Spot does not delete or overwrite project files by default, and uncertainty must be explained instead of disguised as confidence.
- 01
The local entrance can find and open material from natural-language requests.
- 02
Organizing starts with a reviewable plan rather than moving files by default.
- 03
Failed prompts enter review and continue to shape training.
It is still learning ambiguous nicknames and the cross-project clues that make perfect sense to a person but have never been written down.
Continue the journey