Using Usable Chat

Usable Chat is an AI conversation workspace that can combine a model with Usable knowledge, files, tools, Experts, and other capabilities available to your account. Use it for work that benefits from an interactive answer grounded in maintained context rather than a one-off prompt with no shared knowledge.

The exact controls you see depend on your account, selected model, subscription, workspace access, and whether Chat is running in the first-party application or inside a configured embed.

Start a conversation

Open Usable Chat and sign in.
Choose New chat or open an empty chat route.
Review the selected model, mode, tools, and context shown near the composer.
Write a clear request and send it.
Follow the streamed response and any visible retrieval or tool status.

A new conversation can also begin with context supplied by a workspace, fragment, collection, project, or a link from another Usable surface. Check the visible context before sending the first message.

Resume previous work

Select an existing conversation from the conversation list where history is enabled. Resuming restores its message tree and persisted conversation-level state that you are authorized to access.

Before continuing a long conversation, briefly restate the current goal and ask Chat to confirm its assumptions. This is especially useful after a topic change, model change, or automatic compaction.

Temporary, private, or stateless flows may not retain history in the same way. Embedded assistants can also let the host application own persistence instead of Usable Chat.

Write an effective prompt

A useful prompt usually includes:

the outcome you want;
the relevant audience or constraints;
the workspace, fragment, file, or project that should ground the answer;
the desired format;
what should be verified before acting.

For example:

Using the selected product workspace, explain the released invitation workflow for a new workspace owner. Cite the relevant sources, separate confirmed behavior from uncertainty, and finish with a short checklist.

Ask one coherent task at a time. If the request is broad, ask Chat to propose an order of work before it starts using expensive tools.

Understand a streamed response

Chat streams work as it happens. Depending on configuration, you may see:

assistant text arriving incrementally;
retrieval or planning status;
tool and subagent activity;
source or citation blocks;
generated files, images, applets, or canvas actions;
a structured follow-up question;
a stop or cancel control.

Status messages are progress indicators, not the final answer. Wait for the assistant's visible conclusion and verify important claims against the sources it used.

Large tool results or long conversations may be shortened. Released Chat flows show compaction or truncation notices so that missing detail is not silently mistaken for a complete source.

Correct or redirect Chat

You do not need to start over when the answer goes in the wrong direction. Say what is wrong and provide a precise correction:

Use only the selected workspace. Do not use web search. The question is about workspace roles, not organization roles.

If the conversation has accumulated unrelated context, start a new chat or compact it before continuing. If a tool repeatedly fails, disable optional tools or choose a simpler workflow instead of retrying indefinitely.

First-party Chat and developer APIs are different

The first-party web application uses persisted conversations, selected Usable context, Experts, and enabled tools. OpenAI-compatible developer endpoints are stateless unless the caller implements state and supplies the necessary context.

Do not assume that a prompt sent through a generic chat-completions API will automatically inherit:

your current web conversation;
selected Usable workspaces or fragments;
Chat Experts or Skills;
first-party server tools;
project context.

Developer API and MCP details belong in the later Reference and MCP & agents sections.

Troubleshooting

Chat does not respond

Confirm that you are signed in and that the selected model is available. Try a new conversation, then retry with optional tools disabled. If the app appears stale after a deployment, refresh once or reopen the installed app.

The response stops after tool activity

Wait for the released one-shot recovery behavior to produce a visible answer. If the turn remains blank, start a new turn, change model or route, or disable the failing tool path.

A control is missing

The model selector, tools, uploads, reasoning display, and other controls can be hidden or locked by account or embed configuration. A hidden control does not necessarily mean the underlying feature is broken.

The answer ignores my knowledge

Review the active context and continue with Context and grounded answers.

Continue

Next, learn how Context and grounded answers connect a conversation to maintained sources.