What the agent knows

On every message the agent gets:

• The open item, as it is now. In a workspace, its parts and assembly; in a schematic, its parts, wires and nets; in firmware, the files and the board; in a scene, its robots, objects, scripts and controls. On a project page, the project's items.
• Where you are. Which page is open, so it can say what to click. It can also open, create, rename and delete items itself.
• This chat. Everything said and done in the open chat. Very long chats are summarised as they grow, so early details can fade; start a New chat for a new task.
• Your attachments. Images, sketches, screenshots, PDFs, text, CSV and video you attached.
• Your own edits. When you change the item yourself, the chat gets a short note, for example "User edited the assembly.", "User renamed 'arm' to 'upper_arm'.", "User edited script of 'bracket'." or "User imported STEP part 'motor'.", so the agent works from your version.
• This guide. The agent can search and read these pages to answer how-to questions.
• Skills. For specialised work, the agent loads a skill with detailed instructions; its name shows in the panel header.

It does not see your other chats, your other projects, or files on your computer that you did not attach.

Seeing what you see

The agent cannot see the screen unless it takes a picture. In a workspace it can take a screenshot of the view, in a scene it can also record a short clip of the robot moving, and in a schematic it can capture the circuit. Those show in the chat. If something looks wrong to you, attach a screenshot or a sketch so it sees the same thing.

Getting good results

• One step at a time: one part, one block of the circuit, one behaviour of the robot.
• Give sizes, units and constraints ("20 mm wide, M3 holes, fits a 9 g servo").
• Point at things: circle them in a sketch, or name them the way the agent named them.
• When it goes in circles, stop it, say what is wrong in one sentence, and roll back with History if needed.