> ## Documentation Index > Fetch the complete documentation index at: https://docs.polymet.ai/llms.txt > Use this file to discover all available pages before exploring further. # Prompt the agent > Describe what you want, attach screenshots and files, answer the agent's questions, approve its actions and stop it when you need to. The generation input at the bottom of the chat is where you tell the agent what to build or change. Each prompt you send starts a generation: one run of the agent, from your prompt until it stops. Every plan can prompt the agent. Each generation uses credits, by how much work it takes; fixing errors is free. See [How credits work](https://docs.polymet.ai/plans/credits.md). ![The generation input with its buttons: agent connectors, attach, model, context, accept all, web search and voice](https://mintcdn.com/polymet/TfIzLh80_L4bW9xQ/images/docs/canvas/generation-input.png?fit=max\&auto=format\&n=TfIzLh80_L4bW9xQ\&q=85\&s=dd974a357a7bf27c6f1e7ae12017aba1) ## Send a prompt 1. Click into the generation input and describe what you want. Enter and Shift+Enter start a new line. 2. Press Cmd+Enter (Ctrl+Enter on Windows), or click the send button. The agent's work streams into the chat: the files it reads and writes, and what it decided. When it finishes, the canvas shows the result. A prompt can be up to 15,000 characters. Past that, the input says how many characters you are over and won't send. You can't send while you have unsaved changes in the code editor. Save or discard them first. ## Attach screenshots and files Click the file icon at the bottom left of the input, drag files onto the input (it shows **Drop files to attach**), or paste them. | You can attach | Limit | | - | - | | Images, such as screenshots or sketches | Up to 3 per prompt | | PDFs, and text or code files: .txt, .md, .csv, .json, .xml, .yaml, .html, .css, .js, .jsx, .ts, .tsx and similar | Up to 3 per prompt in total. Each PDF up to 5 MB, each text file up to 500 KB | If several files together are too large for the model to read, the input says **Attachments are too large to fit the model's context — remove one or use smaller files.** ## Dictate a prompt Click the microphone and speak. You can record for up to 30 seconds; the button counts down. Click it again to stop. Your words are added to the prompt, so you can edit them before you send. ## Suggested prompts On the projects page, under the generation input, three buttons offer ideas: **Design a product**, **Design a component** and **Design from image**. Click one to see three example prompts, then click a prompt to put it in the input. Some come with an image attached. ## Answer quick questions Before a large piece of work, the agent may ask a few questions first. They appear above the generation input under **Quick questions**, with a dot for each one. 1. Pick an option for each question, or type into **Or type your own...**. 2. Add anything else in the generation input. It reads **Add additional notes... (optional)**. 3. Click **Submit**. If you don't answer for 5 minutes, the agent continues with the default answer for each question. For the last 30 seconds it warns you: **Continuing with the highlighted defaults in 30s, pause to edit your answers.** Click the pause button to stop the countdown and take your time. The countdown also pauses by itself when your organization is out of credits, or when today's free credits are used up. ## Plans for larger requests For a bigger request, the agent writes a plan first and works through it. The chat shows **Created plan**, and the plan appears in the file list under **Plans** with its progress. See [Pages, components and files](https://docs.polymet.ai/canvas/files.md#plans). ## Approve the agent's actions Some actions wait for your approval before they run, for example: * deleting a file * running a query on your database: **Accept this SQL query on your database?** * changing your database's structure: **Deploy database migrations now?** * adding or removing backend functions: **Deploy backend endpoints now?** or **Delete these backend endpoints?** * calls to your [agent connectors](https://docs.polymet.ai/connectors/agent-connectors.md) Click the approve button (**Accept**, **Deploy** or **Delete**) or **Reject**. The arrow next to the approve button offers **Accept all in this session**, which approves every later action, calls to your agent connectors included. You can also turn this on before you start: click the fast-forward button in the generation input, or press Shift+Tab while typing. Its tooltip reads **Accept all actions, including calls to your agent connectors, which can spend credits there**. Calls to outside tools can cost money with those tools, so turn it on only when you are happy for the agent to act without asking. ## Generated images When a design needs a picture, such as a hero photo or an illustration, the agent can make one. The chat shows **Generating image**, then **Generated image** with its shape, the prompt it used and a preview. Click the preview to see it full size. ## Stop a generation While the agent works, the send button turns into a red stop button. Click it to stop. * If a teammate started the generation, Polymet asks first. **Stop generation** stops it for them too; **Keep it running** leaves it. * If you have queued messages, choose **Keep queue** to keep them for later, or **Discard queue** to drop them. See [Threads, context and queued messages](https://docs.polymet.ai/canvas/threads-and-context.md). A stopped generation still uses credits for the work it did before you stopped it. See [How credits work](https://docs.polymet.ai/plans/credits.md). ## When something breaks When a screen fails to build, an **Errors to fix** card appears in the chat, listing each file with a problem. Click a file to jump to it, or open it to read the error and copy it. * **Fix errors — free** asks the agent to fix them. Fixes don't use credits, but your organization needs a balance above zero to start one. * **It's okay** hides the card. The same choice appears on the broken canvas node itself. See [The canvas](https://docs.polymet.ai/canvas/canvas.md#when-a-screen-shows-an-error). ## Related * [Models and thinking effort](https://docs.polymet.ai/canvas/models.md) — the model menu and the web search button * [Threads, context and queued messages](https://docs.polymet.ai/canvas/threads-and-context.md) — sending while the agent works, and slash commands * [Select and edit with AI](https://docs.polymet.ai/canvas/select-and-edit.md) — point at elements instead of describing them * [Write better prompts](https://docs.polymet.ai/canvas/prompting-guide.md) — what to put in a prompt * [How credits work](https://docs.polymet.ai/plans/credits.md) — what a generation costs This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.