Editing content

Ask for the change you want in plain language — a new article, a reworded section, updated pricing copy, a fixed typo across every page. The assistant handles the mechanics; a few rules keep it safe.

The editing flow

• It reads before it writes. Before editing any document the assistant queries its current state — the exact document, its fields and shape — so edits are grounded, never guessed.
• It confirms the target once per task. Which environment (production, or staging when connected) and draft vs published. Say it in your request — "as a draft", "straight to live", "on staging" — and that's the answer; it won't re-ask for every document in the same task.
• It knows today's date. The real current date and time is injected into every run, so publish dates, "today", and relative dates ("next Friday") come out right.
• Slugs and validation follow your content model — lowercase-hyphenated slugs from the title, required fields included on create.

Drafts vs published

A draft is saved in your CMS but not live — it stays unpublished until someone promotes it. A published write goes live immediately (on Contentful-style static sites, after the ~1–2 minute rebuild). Where your CMS supports draft versions of published documents (Sanity, Contentful), edits can be staged on the draft and promoted later.

The publish gate

Any write that changes live content — publishing a draft, editing a published document directly, or creating a document as published — requires the requesting user's role to have publish permission for that content. Users without it aren't stuck: the assistant makes the change as a draft and offers to send an approval request to a colleague who can put it live.

Verification

After every change the assistant re-reads what it wrote and confirms the fields hold exactly the intended values before reporting success. If something doesn't match, it says so and fixes it — it's not allowed to declare success on unverified work.

Custom style & tone prompts

On plans that include custom prompts (Growth and above), admins can train the assistant on your voice under Style guide in the portal: a tone of voice, a content & writing style guide, and a list of content to avoid — plus reference URLs that exemplify your writing, and uploaded brand documents (text and Markdown files are read directly into the assistant's guidance). Everything you set becomes part of the assistant's standing instructions for every team member's content work, and the avoid-list is binding: the assistant refuses conflicting requests and proposes an alternative.

Deletes

Deleting is destructive, so the assistant describes exactly what would be removed — which document, which environment — and waits for an explicit "yes, delete it" before acting. Deletes are permission-checked and audit-logged like every other write.