# Chat The chat panel is the main place you direct an agent. Type your task in the input box at the bottom and send it; the agent's responses, tool calls, and progress appear in the panel above. ![Chat panel](images/chat_panel.png) --- ## Model picker The model picker sits in the chat input toolbar, to the left of the send button. For Claude agents it lists each model family (Fable, Opus, Sonnet, and Haiku) as a "latest" entry, then the current and previous generation; Haiku has no previous entry yet. A "latest" entry runs the newest model your Claude CLI knows, so it moves to each new model as it ships; a numbered entry stays on that version. Click the picker to switch the model for the current agent session. The change takes effect on the next message you send. You can set the default model for new agents under **Settings > Claude**; until you do, new agents use the model you picked most recently, or Opus 5.5 if you haven't picked one yet. ![Model picker](images/model_picker.png) --- ## Mentions and attaching context The **+** button on the left side of the chat input toolbar opens a menu for adding context to your next message. The same menu opens when you type `@` in the input. From it you can: - **Files & folders** — point the agent at a specific file or directory in your repo. The agent loads it into context. - **Skills** — start a slash command. (You can also type `/` directly.) - **Images** — attach a screenshot, design reference, or other image. You can also drag and drop images onto the input, or paste them directly. Attached files appear as small pills above the input; click the **×** on a pill to remove it. Images shown in the chat — whether you attached them or an agent produced them — can be clicked to open at full size. To save one out, right-click the image (either the thumbnail in a message or the full-size view) and choose **Copy Image** to copy it to your clipboard. ![Plus-button menu in the chat input toolbar](images/attach_context.png) --- ## Tracking context usage After each agent turn, the turn footer shows how much of the model's context window the conversation has used — for example, **45% context**. Click the percentage to see a token breakdown. To free up context, use `/compact` (summarize the conversation so far) or `/clear` (start a fresh conversation). See [Slash commands](#slash-commands) below for details. --- ## Plan, fast, and effort The right side of the chat input toolbar — between the message input and the send button — has three controls that shape how the agent handles your next message. **Plan mode** (checklist icon) — Toggle on to make the agent plan before it acts. It reads the relevant parts of the codebase, writes up a plan, and waits for your approval before making any changes. Useful for larger tasks where you want to review the approach first. Toggle off to return to the default "act first" behavior. **Fast mode** (lightning-bolt icon) — Toggle on for faster output from the same model, at a higher per-token price; useful for quick iteration. It appears only for models that support it: Opus 5.5, Opus 5, and Fable 5.2. On Claude subscription plans, fast mode draws on usage credits even when your plan has usage left. You can set the default for new agents under **Settings > Claude**. **Effort level** (brain icon with a fill bar) — Click to pick how much thinking the agent budgets for each step: **Low**, **Medium**, **High**, **Extra High** (the default), and **Max**. Lower effort is faster and cheaper; higher effort is slower but more thorough. --- ## Slash commands Type `/` in the input box to open the command and skill picker. It lists Sculptor's built-in commands, the bundled skills, and any skills or commands you've installed locally — all sorted alphabetically. ![Slash command picker](images/slash_commands.png) **Conversation commands** act on the current agent session: - `/clear` — clear the conversation context to start a fresh task. - `/compact` — summarize the conversation so far to free up context. - `/context` — visualize what's using context, to decide whether to compact or clear. - `/copy` — copy the last response to your clipboard. **Skills** run as full agents with their own tools. A few general-purpose ones — `/batch`, `/btw`, `/loop`, `/simplify` — handle work you'd otherwise re-prompt for, and two skill bundles (`sculptor-workflow` and `sculptor-experimental`) make up Sculptor's engineering workflow. See [Skills](skills.md) for the full reference, and the [Claude Code commands reference](https://code.claude.com/docs/en/commands) for everything else. --- ## Agent tasks For complex work, the agent breaks the task into discrete steps — a todo list it plans and works through. While the agent is running, its progress shows in the status indicator above the chat input (for example, **1 / 8** with the current step). Hover over or click the indicator to expand the full list, where each step is marked pending, in progress, or done. If the agent goes off track, send a correction in the chat input to redirect it; the task list updates to reflect the new plan.