Demo
The same question, asked twice.
Two sessions recorded on 22 August 2026. One is an ordinary chat pointed at an empty folder. The other is the same assistant, same question, pointed at a DeskRoot workspace. Nothing was typed into either but the question.
16 seconds. Waiting is sped up where the label on the clip says so.
Typed into both, word for word: Write the update for Thursday's leadership meeting.
What the regular chat did
It could not start, and it said so. The folder was empty and nothing was connected, so it reported having nothing to base an update on and asked where the content should come from: paste it, connect a tool, or take a template.
That is the correct answer. It is also the whole problem. The assistant is not being unhelpful; it genuinely does not know the project, the format, the audience, or what was decided last time, because nobody has ever told it.
What the clip does not show: this session then ran on for another two and a half minutes and produced a two-page Word template full of [What must be decided], [Name] and [Date], then asked for the content again. Just under three minutes from question to a blank form, and the writing is still yours.
What the DeskRoot chat did
It read five files before it answered, and named every one of them on screen: the standing rules, the profile and communication files, and two memory files. Nobody mentioned any of them in the question.
The draft came back in the three-section format this workspace says to use, decisions first. It treated the Northgate slip as a decision taken in April to absorb rather than escalate, not as fresh bad news. It named the reporting cleanup as unowned instead of quietly dropping it or inventing an owner.
The part worth watching for is what it refused to do. The workspace holds no entries after 9 April, so rather than filling the four-month gap with something plausible, it marked the draft unverified and listed exactly which figures need checking before the update is sent. About 36 seconds from question to draft.
What was changed in these recordings
Both are real sessions. Neither is a reconstruction, a mock-up, or a second attempt after a better answer. Everything done to them is listed here.
- Waiting is sped up, and only waiting. Each sped stretch carries its multiplier on screen while it is running. No other part of the timing was touched.
- One local file path is blurred in the DeskRoot clip. It showed the folder location on the machine that did the recording.
- The app's chat sidebar is cropped out of frame in the DeskRoot clip, because it listed unrelated conversations.
- The workspace belongs to an invented person at an invented company, with invented projects, so that no real person or business appears on this page. The session, the files it read and the answer it gave are not invented.
- Nothing was re-recorded to get a better answer. The regular chat's reply is what it actually said.
The difference is not a better model
Both sessions ran the same assistant on the same day. The only thing that changed is whether the folder it was pointed at had standing rules and context in it. That folder is what DeskRoot is: a short set of rules the agent reads before it does anything, plus files that hold who you are and what you have already decided.
How it works · The standing rules file, published in full · Pricing