Chat & sessions

Everything you do in Harriet Desktop happens in a chat session inside a workspace. This page covers starting sessions, reading streaming responses and the execution plan, and how workspaces and templates keep different jobs separate.

Workspaces come first

A workspace is a folder you choose and share with the app: an existing project or notes folder works fine. The AI reads and writes inside that folder, which is what lets it do real work — summarize a quarter's invoices for finance, draft vendor emails from a notes file for operations, assemble a renewal brief for sales — instead of only answering questions.

Keep unrelated jobs in separate workspaces so files, sessions, and context stay tidy.

Sessions

A session is one conversation with its own history. Use Create new session to start a fresh task in the current workspace rather than piling unrelated requests into one thread. Old sessions stay in the workspace, so you can come back to yesterday's analysis without re-explaining it.

1. Pick or create a workspace. Choose the folder that holds the files this job needs.
2. Create a new session. Each distinct task gets its own session.
3. Describe the job. Be concrete: name the file, the output you want, and any constraints. Your assigned skills apply automatically when they match.

Finding your way around the sidebar

The sidebar lists your workspaces in a fixed order, with only the one you are working in opened. Selecting another workspace opens it and closes the last one, so the list stays the height of a single workspace however many you have. To keep a workspace open alongside the current one, click its chevron — pinned workspaces stay open until you close them again.

A closed workspace still tells you what changed inside it. It shows a count of unread tasks, or a red count when something in it is waiting on you, and a moon when a cloud workspace is asleep. Sleeping is normal: cloud workspaces pause when idle and wake when you open one.

Two sections sit above the list:

• Needs attention appears only when something is blocked, and gathers it from every workspace at once: a task waiting for your approval, a question the AI asked you, a run that failed, or an event that needs review. Cloud tasks appear here even while their workspace is asleep, and opening one wakes it.
• Recent lists the last few tasks you worked on, in any workspace, so you can pick up where you left off without opening anything.

Both sections can be folded away with their headers, and stay that way until you open them again.

Searching across workspaces

The search box at the top of the sidebar finds a task by its title, by a filename it produced, or by a phrase said inside it — useful when you remember what a task was about but not which workspace it was in. Results are grouped by workspace, with the one you are in first, and each transcript match shows the sentence it matched.

Use the All workspaces and workspace buttons to switch between searching everywhere and searching only where you are. Arrow keys move through results, Enter opens one, and Escape clears the search and brings the workspace list back. The same search runs from the keyboard with ⌘K (Ctrl+K on Windows).

Task names are searchable everywhere, including cloud workspaces that are asleep and ones you have never opened on this computer. Harriet keeps a list of your task names so the sidebar can find them without having to wake a sleeping workspace.

Searching by a phrase from inside a conversation works for workspaces you have opened on this computer. The conversations themselves stay in the workspace rather than being copied to Harriet, so a workspace you have not opened here is findable by name but not yet by its contents. Open it once and it catches up.

Streaming responses and the execution plan

Responses stream in live, so you see the AI's answer as it is produced rather than waiting for a finished block. On longer tasks the app also shows an execution plan: the AI's todo list, rendered as a timeline you can watch. While the AI is working out its approach the session shows a Planning status; as it works through the list, the app reports progress in the form "2 out of 5 tasks completed".

The plan is your early-warning system. If an IT admin asks for a device report and the plan shows the AI heading toward the wrong data, stop it and redirect before it finishes, not after.

Templates for repeated work

If you run the same workflow often — a weekly pipeline summary, a monthly close checklist — save it as a template and re-run it instead of retyping the prompt. Templates are stored locally on your machine and are not shared with anyone. To give the whole team the same capability, ask your admin for a skill, which travels with your profile to every device.