Ways to start
The empty panel offers one-click starters — Summarize this doc, Explain this doc, Proofread, Improve the writing — or just type what you want. Three chat modes control how far it may go:Give it context
The agent already sees the open doc. You can sharpen its focus:- @-mention a Salesforce component in your message — mentions become context pills, and the agent fetches that component’s metadata before answering.
- Add the editor selection from the + menu — highlight text in the doc and hand it to the conversation, so “rewrite this section” needs no re-selecting.
- Attach images or diagrams from the + menu.
How edits arrive
When the agent proposes a change, it doesn’t replace your text. Edits appear as inline tracked-change suggestions directly in the document — word-level insertions and deletions, the way a careful colleague would mark up a draft. A review bar counts the pending edits and steps you through them:- Accept or reject each suggestion individually — clicking one shows a card describing the addition, deletion, or rewrite.
- Accept all or reject all in one action.
- The chat shows “Revision ready” with a Show changes shortcut that scrolls the document to the first pending change.
Slash commands
Type / in the agent composer:AI skills
Beyond the chat panel, the editor has lighter-weight AI entry points: select text and run a skill from the floating toolbar, or press space on an empty line to open the inline AI bar and type a freeform instruction. Built-in skills:
Diagram skills — data model (ERD), sequence, integration architecture, flow logic, and more — draft entire diagrams from the doc’s tagged metadata.
Custom skills
Admins can add organization-specific skills under Settings → AI skills — for example “Convert to admin runbook tone” or “Add a compliance disclaimer”. Each has an emoji, a prompt, and a behavior: edit (arrives as tracked changes) or chat (answers without touching the doc).Whether an edit comes from the agent panel, a skill, or the inline AI bar, it lands as a reviewable change — the grounded-or-it-doesn’t-ship rule applies everywhere.