> ## Documentation Index > Fetch the complete documentation index at: https://docs.cogniagent.ai/llms.txt > Use this file to discover all available pages before exploring further. # The conversation > Working with your AI coworker is a conversation, not a form. Here's what's on screen and how to steer it while it works. You ask, it works, you watch. You can step in at any point without stopping it. ## What's on screen Three things, and you'll use all of them. **The conversation** in the middle — your messages, your coworker's replies, and every step it took in between. Nothing is hidden; if it opened a file or searched for something, that's a row you can expand. **The files panel** beside it — everything in the task. Anything your coworker produces lands here. See [Files](/cowork/files/overview). **The message box** at the bottom, which does more than it looks like. `/` opens the [command menu](/cowork/chat/slash-commands), `@` points at a file, and the paperclip attaches one. The message box with two attached spreadsheets and the plan, effort and model chips beneath it ## Asking well The single biggest difference in output quality comes from the first message. **Say what you want to end up with.** "A one-page summary I can send to my boss" produces something different from "look into this". **Attach the real thing.** Upload the export rather than pasting three rows of it. **Say what you already know.** Saving your coworker from re-deriving context you have is free speed. **Say what to avoid.** "Don't email anyone" is worth writing if that matters. You don't need to be precise on the first try. It's a conversation — say "not quite, more like this" and it adjusts, without starting over. ## Steering while it works You don't have to wait for it to finish before saying something. **Type while it's working.** Your message is queued and picked up as soon as the current step settles. Queued messages appear above the box, and you can edit or cancel one before it's sent. **Correct it mid-task.** "Stop doing it that way" lands and it adjusts. **Stop it.** If it's clearly off track, stop it and re-ask. Faster than watching it finish something wrong. ## Finishing When your coworker considers the job done, it posts a short wrap-up you can expand for the detail. A finished task isn't closed. Send another message and it picks straight back up with everything it knew — which is why "actually, also do X" works without re-explaining. You can mark a task done from the header, and even then a new message re-opens it. ## Common questions Yes. It keeps working. Come back whenever and the whole conversation is there. Anything leaving your workspace waits for approval. See [Approvals](/cowork/control/approvals) and [Permission rules](/cowork/control/permission-rules). Just say so. Mid-task changes of direction are normal and it doesn't need to start again. Related work, yes — the shared context helps. Genuinely separate jobs are better as separate tasks. See [Managing your tasks](/cowork/chat/tasks). Every step is in the conversation and can be expanded. See [Watching progress](/cowork/chat/watching-progress). ## Next steps Everything you can type. Following the work. Agree the approach first. Getting files into the conversation.