Work with the creation chat

Choose Plan, Design, or Ask

The mode menu beside the model controls changes how the conversation approaches your request. It is separate from the creation type, which describes the output, such as a prototype or motion design. Start by choosing the result in Creation types, then choose how you want to work.

1. Choose Plan when you want an editable plan before implementation. Review its scope and decisions before handing it over to Design.
2. Choose Design to create or change concrete outputs, such as a page, prototype, animation, slide deck, or project file.
3. Choose Ask for answers and discussion without assuming a design artifact must be created.

Design is the default. These are conversation modes, not security-permission levels. A request such as “explain this layout” belongs in Ask; “change this layout and save the result” belongs in Design. Switching a mode does not by itself rewrite the existing project.

Send a brief you can check

1. Describe the task, audience, required content, visual direction, dimensions, and behavior. State what must remain unchanged when modifying an existing work.
2. Review the selected agent, model, mode, and context chips before sending.
3. Send once, then follow the conversation and file changes. The rotating example text in an empty composer is a hint, not a request you have already entered.

Update the existing booking page. Keep the layout, logo, and sample classes. Add required-name validation, show an inline error for an invalid email, and display a confirmation state after valid submission. Do not add a backend or change other pages.

For later revisions, refer to the exact screen, file, scene, or element. Avoid asking for several unrelated changes in the same follow-up if you need to review their effects separately.

Add context deliberately

Use Add context or the attachment entry to provide reference documents, screenshots, or assets. In versions with the @ picker, you can search available workspace context, project files, plugins, Skills, enabled MCP servers, and connectors. The actual list depends on what is installed and available.

1. Choose the file or resource needed for this request. Confirm that its chip or mention appears in the composer.
2. If several resources have similar names, check the selected item before sending. Add a sentence explaining how the agent should use it.
3. Remove an unwanted context chip before sending. Removing it cancels that context selection; it does not delete the original file or uninstall the plugin.

A reference is not a complete requirement. Say whether it provides facts, visual style, source content, or a file to modify. For reusable preferences across tasks, see Memory.

Working directory is not a file import

Home’s Working directory lets a new task use a selected local folder as external context. It does not import all of that folder into Design Files, move the project, or replace the project’s own directory. Use the project file workflow when a file needs to appear in the project.

Choose the directory before starting the new project and verify the path. The current project composer does not provide the same working-directory picker; do not search for an identical control at the bottom of an existing conversation. If a directory cannot be accepted, check that it still exists and is accessible before retrying the selection.

An attached file, a linked folder, and the project’s generated files have different roles. When asking for an edit, name the intended file and check where the result was actually written.

Review the agent and model

Open the model control beside the composer to inspect the current selection. Available models depend on the active agent or configured provider. Where the agent and model expose reasoning options, select from those offered; an absent reasoning selector is not proof that your configuration failed.

Changing the selection does not regenerate an existing result. Check it before the next request. Do not assume the same model names, reasoning levels, or service options exist across OpenDesign Cloud, local agents, and BYOK. Setup and credential checks are covered in AI connections and Local agents.

Understand requests sent during a run

When the current conversation is busy, another submitted request enters that conversation’s queue. It is not a second task running in parallel. Look for the pending-request list rather than repeatedly sending the same instruction.

1. Write the next change and send it once. Check that it appears as a pending request.
2. Use Edit on a queued item to bring its text, attachments, comment attachments, and context back into the composer. After editing, Send updates that queued item rather than creating an extra copy.
3. Use the drag handle to change order, or delete a pending item you no longer want.

Keep the order meaningful: a request to refine a page should follow the request that creates it. Check restored context after editing, especially if a referenced resource has since become unavailable. Queue controls appear when there are pending items; the figures on this page do not claim to demonstrate a completed queue run.

Stop and send immediately have different effects

The running composer provides Stop. Use it when you want to interrupt the current generation, then inspect the conversation and files before deciding what to do next. Stopping is not an undo action: already created or changed files can remain.

The Send action on a queued item is more consequential than simply changing its position. If the agent is still working, it interrupts the current run, brings that queued item to the front, and starts it through the normal queue path once the conversation is ready. It does not keep both runs going.

If you want the current task to finish, leave the request queued. If you deliberately need a replacement task now, use the queued item’s Send action and then review the partially completed work.

Answer the agent’s questions

The agent can present a question form to clarify your request. Depending on the question, choose a fixed option, select several options within the allowed limit, enter a custom answer, or provide the requested file. Multi-question forms may show one question at a time.

1. Read the question and any required marker. Answer it, then advance to the next question if the form is stepped.
2. Use a custom answer when the available choices do not fit. Check selection limits before submitting.
3. Submit the form to continue. If you choose Skip or Skip all, that is recorded as a deliberate skip, including for questions that would otherwise require an answer.

Skipping does not mean you approved one of the suggested choices. The agent may continue with assumptions, so review its next result. Submitted forms remain as a record of your answers; they are not the same as a still-editable draft. Give a correction in a new message when you need to change a previous decision.

Recover from a failed request

1. Read the error and identify the cause: sign-in, provider configuration, network, model access, quota, or a failed file or media operation.
2. Resolve that cause and use the recovery action actually offered by the client. Some failures offer authorization followed by retry; others allow a manual retry or a different recovery route.
3. Check the existing files and preview before repeating the whole brief. For one failed asset or step, ask to preserve the rest and retry only the affected part.

There is no universal Retry button for every failure. A hard quota restriction may require waiting or resolving account access first; repeatedly sending the same request does not remove the restriction. Retry is also not a rollback guarantee. Keep the error text and occurrence time for Troubleshooting.