# Preferences

Preferences control how the operator works with you. They help you choose an approach to planning, execution, model use, and interaction without repeating those choices in every message.

## Choose an approach

For unfamiliar or consequential work, begin with a plan and review it before execution. For a well-understood task, direct execution can reduce back-and-forth. Approval preferences determine which actions require review within the controls available to that resource.

| Choice | Purpose |
| --- | --- |
| Plan first | Scope, approach, or consequences need agreement. |
| Execute the task | The desired result and boundaries are already clear. |
| Review critical actions | Routine work can proceed, but consequential changes need attention. |
| Review all changes | You want to inspect changes before they happen. |

An execution preference never grants access to another person's resource or bypasses platform restrictions. A model choice also does not authorize a connected account.

## Use the Composer controls

Open the relevant control directly or choose it from the [slash command menu](/docs/composer/slash-commands).

| Control | Decision | Guide |
| --- | --- | --- |
| `/model` | Use Auto or choose an available model | [Model selection](/docs/composer/controls#model-selection) |
| `/permission` | Which actions need approval under the available policy | [Permissions](/docs/concepts/permissions) |
| `/plan` | Whether to establish an approach before execution | [Plans](/docs/operator/plans) |
| `/steer` | How to deliver another message while work is running | [Queue or steer](/docs/composer/messaging#queue-or-steer) |

Permission choices include **Autonomous**, **Critical actions**, and **All actions** where supported. They govern review within the current authorization boundary; Autonomous does not give the Operator unrestricted access. Read the control's scope before changing it, especially in a shared Workspace or Agent.

Planning and approval solve different problems. A plan helps establish what should happen. Approval controls determine when an action needs your review. Enabling Plan mode does not approve the resulting Plan.

## Example: explore before changing

Suppose you want to reorganize a customer portal, but have not settled on its navigation.

1. Enable Plan mode from `/plan`.
2. Describe the audience, existing structure, and desired improvement.
3. Review the proposed approach and clarify unresolved choices.
4. Approve the intended Plan revision when you are ready to proceed.

For a later, well-defined correction, such as adjusting one label, you may prefer direct execution within the same access and approval constraints. Choose according to the task rather than maintaining a permanent instruction to plan every small change.

## Example: guide work already underway

If a draft is heading in the wrong direction, use steering to give timely guidance: “Keep the customer examples, but shorten the technical explanation.” If the request should happen afterward, queue it instead: “After this draft, create a three-sentence summary.” See [Messaging](/docs/composer/messaging) for delivery behavior and stopping work.

These messages guide the current task. Save a preference or instruction only when you want the choice to shape future work too.

## Apply preferences at their owner

Use Workspace **Preferences** for eligible Chat defaults. Review an Agent's own configuration for specialist work, or a Project's settings for development. A Chat does not have a separate library of persistent operator configuration.

Before changing a shared default:

1. Confirm the active Workspace or resource.
2. Identify which future work should use the change.
3. Save the choice and try a small representative task.
4. Compare the result with your intent before applying it more broadly.

Changes guide future runs; do not assume a running task has acquired newly selected capabilities. Give task-specific direction in the conversation, and use [Instructions](/docs/configuration/instructions) for guidance that should persist.