Skip to main content
The chat is not a search box and it is not a genie. It drives a fixed set of typed operations against your content, and a request works when it maps onto one of them. Once you can feel where that edge is, the hit rate goes up sharply.

What it can actually do

Everything below is a real operation, which is why the editor never has to guess whether it is allowed: And the hard edge, which no phrasing gets around:
  • It cannot change your components, layout code, routes, or anything else in the repository.
  • It cannot edit a field your developers did not declare as content. If the badge colour on a pricing card is not in the block’s schema, asking nicely will not surface it — that is one line for a developer to add.
  • It cannot invent a block type that does not exist on your site.
When a request lands outside that boundary, the chat says so rather than silently doing something adjacent.

Scope it by selecting, not by describing

The biggest single improvement to your results costs nothing: turn on the element picker (the cursor button in the composer), click the thing you mean, and then type. Without a selection you have to write “in the second testimonial card, the one from the logistics customer, change the quote so it mentions onboarding speed.” With a selection you write “mention onboarding speed.” Shorter, unambiguous, and much less likely to edit the wrong card.

Be specific about the outcome, not the mechanism

The model is better at what you want the page to say than at how you imagine it being built. Requests that name tone, audience or intent work well: “rewrite for a technical buyer”, “cut this by half”, “make the CTA sound less pushy”.

One request at a time, mostly

The planner handles compound requests — “change the headline and add a CTA at the bottom” — but every extra clause is another chance to get one of them wrong, and a partially-wrong plan is more annoying to unpick than two clean ones. Batch when the parts are related; split when they are not.

What things cost

Chat runs on your organisation’s own API key, and the price of a turn varies by more than you would guess from the length of what you typed. From a measured run on the demo site: Two things generalise from that, even though your numbers will differ. Whole-page rewrites — translation above all — cost roughly ten times a single-block edit, because they rewrite every text-bearing block rather than the one you were looking at. That is the one request worth being deliberate about. Nothing in the interface warns you. Some requests are genuinely free. Adding a standard section is answered without calling a model at all. The Properties panel is always free. If you know exactly which field and exactly what value, typing it there costs nothing and cannot be misread.

When the site has no API key

If the header shows Add an API key, chat still answers, but with a much smaller brain: a short list of literal phrases rather than a model. The suggestion chips on the first screen change to the three it can genuinely execute, and a sentence of your own invention will usually come back asking for a key. Everything else — the preview, the Properties panel, click-to-select — is unaffected.

Teach it your voice once

Rather than repeating “keep it in our house style” on every request, put it in the site’s editorial brief — an overview, a tone, and a list of constraints, in Site settings. The planner reads it on every turn. The editorial brief

When a result is wrong

Three things, in order:
  1. Undo it. The undo button on the message that caused it, or Ctrl+Z.
  2. Say what was wrong in the next message — “too formal, and keep the original headline”. It has the context of what it just did.
  3. Give it a thumbs-down. The Bad response button on a message takes an optional note, and those reports are what the prompts get tuned against.
If it is consistently wrong in the same way on the same site, that is usually a content-model problem rather than a phrasing problem — worth passing to whoever maintains the integration, with chat troubleshooting.