Ask in the AI panel

The assistant works inside the project, so it can read the files you point it at instead of guessing.

How to do it

1. Open the AI panel from the editor and describe the result you want, not just the symptom.
2. Type @ to attach a specific file as context, or select source in the editor and it is offered as context automatically.
3. Attach an image when the configured model accepts images — a screenshot of a broken layout is often faster than describing it.
4. Attach any other file from your computer (a PDF, a spreadsheet, an archive) with the same button or by dropping it on the message box. It is saved in the project's .jointtex/uploads folder, where the assistant reads it; that folder is not compiled, so ask the assistant to copy a file into the project when the document should use it.
5. Start a new session for a new task; each session keeps its own history, and old ones can be hidden or permanently deleted.
6. Open More Agent actions to choose Switch to floating or Dock to the right; Reset position and size puts a floating panel back.

Control what it may change

Editing your files is a permission you grant per session, not something the assistant assumes.

How to do it

1. Use the Agent controls to allow or forbid changes to project files for this session.
2. Turn on Edit to allow changes. Leave Ask approval on to approve each file-changing tool call, or turn it off, shown as Auto, to let them run without a prompt.
3. Leave Track review on so the session's changes are tracked in Review and can be accepted or rejected one by one.
4. Set Thinking in the same controls, from Instant to Maximum where the model offers it, to choose between faster replies and deeper reasoning. It cannot change while the Agent is running.
5. While a turn is running you can answer its question, steer it in another direction, or stop it outright.

Plan before it edits

For a larger task, have the assistant research and write a plan first, and let it change files only after you have read the plan.

How to do it

1. Type /plan followed by the task. Plan mode is read only: the assistant can look through the project but not change it.
2. Read the plan it proposes. Choose Approve and execute to carry it out, or Keep planning and say what to change; each new version is numbered as a revision.
3. Choose Exit plan mode, or type /plan off, to stop planning and return to a normal conversation.

Use slash commands and goals

A message that starts with / is a command to the panel, not a message to the model.

How to do it

1. Type / to list the commands, or /help to see each one described. Commands do not accept images.
2. /new starts a new conversation, /clear starts an empty one and keeps this history, /fork branches the current conversation, and /stop stops the current work.
3. /context shows how much of the model's context is in use, and /compact summarizes older context to make room in a long session.
4. /goal followed by an objective sets a persistent task that the assistant keeps pursuing within a budget of rounds and tokens. Use Pause goal, Resume goal, and Clear goal, or /goal pause, /goal resume, and /goal clear.
5. /skills lists the skills available in the project and /skill followed by a name applies one; /mcp shows external tool connections. Skills and connectors are added from Plugins.

Edit a selection with Quick Edit

Quick Edit works on the text you have selected, in place, without opening a conversation.

How to do it

1. Select text in the editor and open Quick Edit.
2. In Edit mode, describe how to change the selection and choose Generate. Read the change preview, then choose Accept to apply it, Regenerate to try again, or Discard to drop it.
3. Switch to Ask to put a question about the selection without changing anything; with read-only access this is the only mode.
4. In Quick Edit, choose Send to Chat to carry on in the AI panel. To attach a selection to the AI panel as context without opening Quick Edit, choose Add to Chat on the selection toolbar.
5. A selection must be shorter than 32,000 characters.

Ctrl+E opens Quick Edit.

Work through the results

Everything the assistant writes lands in the same places as a human collaborator's edits.

How to do it

1. Read its changes in Review, and use History to compare or restore if a turn went the wrong way.
2. From the Problems panel, choose Explain or Fix on one diagnostic, or Fix all to have it fix every error, with the compile log already in hand.
3. A long turn stops at a step limit; continue it if the work is genuinely unfinished.
4. Retry re-runs a message in a fresh session that ends just before it, so the original session stays intact.