# Mobbin Deep Search - Work OS Addendum

Date: August 9, 2026

This pass focused on the remaining Work OS flows rather than generic AI chat
screens.

## Durable work

- [Relevance AI task timeline](https://mobbin.com/screens/c1486516-90be-4b17-9bb7-f7d5e3290911)
  keeps a task, its background step, and returned work in one chronology.
- [Manus completed task](https://mobbin.com/screens/f7cc29b0-b12c-4154-b870-8f1cb4410d5a)
  promotes the artifact while retaining a compact factual progress record.
- [Elicit research run](https://mobbin.com/screens/232b3136-ccbb-42e7-a45c-412ea436f330)
  names research stages and factual counts instead of inventing percentage
  progress.

Borrowed:

- durable run identity;
- named stages and completed units;
- returned work promoted above the transcript;
- progress that can be inspected after completion.

Work OS difference:

- the run stays attached to the consequential object in the room;
- a foreman rejection remains visible;
- the tape records what the machine deliberately did not do.

## Draft review

- [Rox email draft](https://mobbin.com/screens/3d205a5d-5ead-4787-af2c-9b0a36ed2fc7)
  places an editable draft beside the supporting account context.
- [Notion Mail AI composer](https://mobbin.com/screens/c76eb3bd-f70d-4235-baf4-f24c06421205)
  exposes accept, discard, and retry around machine-written text.

Borrowed:

- editable work before use;
- recipient and message context beside the draft;
- an explicit review boundary.

Rejected:

- a mail-send button inside Work OS;
- treating generated text as approved because it exists;
- hiding requested and actual provider state behind a success message.

Work OS ends at a receipted draft in Gmail. Nathan sends from his own mail
client.

## Learned behavior

- [ChatGPT saved memories](https://mobbin.com/screens/90a8ee61-150e-4cb3-b3af-8c35a451fc82)
  makes retained memories inspectable and searchable.
- [Claude style editing](https://mobbin.com/screens/d7fb1246-8be8-44ed-9b8e-bc5e86fefc7d)
  separates the editable instruction from a preview of how it will behave.

Borrowed:

- visible learned state;
- direct correction;
- a preview of machine interpretation.

Work OS difference:

- each pattern shows its denominator and arithmetic;
- a correction strikes through the old read instead of replacing it;
- learned behavior remains quiet until it is relevant or explicitly inspected.

## Resulting design decision

The research supports the selected one-drawer model. Runs, drafts, and learned
behavior need focused depth, but none needs a new page or a chat-first home.
The Work OS room remains visible and causally connected behind every depth flow.