# Prompting
URL: /en-US/docs/agent/prompting

How to give Spirit a task it can finish in your workspace.



Talk to the agent the way you would brief a teammate who can read the repo and wait for approval.

## Be specific [#be-specific]

Name the files, the outcome, and what not to touch. A tight prompt lands a small, reviewable diff. A vague prompt burns a turn on discovery.

```txt
Read src/auth/session.ts and fix the expired-token refresh path.
Do not change the login form. Add a focused test next to the existing session tests.
```

## Attach context [#attach-context]

On Desktop, attach open files, selections, or a Browser element when the task is about that exact spot. The agent still has to read the file to see the current contents — attachments are pointers, not a substitute for [tools](./tools.mdx).

## Pick a mode first [#pick-a-mode-first]

* **Agent** — you already know the change
* **Plan** — you want a written plan and [questions](./ask-questions.mdx) before any edit
* **Ask** — you want an explanation only
* **Debug** — you can reproduce a bug and want logs

See [Modes](./modes.mdx).

## Let approval do its job [#let-approval-do-its-job]

On **default** approval the agent stops before editing or running a command. Read the diff, then approve or reject. Do not turn on bypass until you trust the loop. See [Approvals](./approvals.mdx).

Editor inline completion (Tab) is not part of this conversation. Configure it under Settings → Tab.