Revising well — giving feedback the AI can actually act on

It's easy to assume the revision thread works like a chat, where earlier messages are still "in the room." It doesn't, mechanically: each request is processed with exactly two things — the project's current draft, and the single instruction you just typed. There's no memory of what you asked for three messages ago beyond what's already reflected in the document itself.

In practice: "like we talked about" or "the other change I mentioned" has nothing to attach to. Restate what you want, self-contained, every time, even if it feels repetitive.

"Make it better" or "this section feels off" gives the model nothing to act on — it's not that it can't understand the feeling, it's that there's no concrete change described. "Change the CTA in the Quote section to say 'Get a Same-Day Quote' and mention we serve Springfield and Chatham" is directly actionable because it names the section, the exact change, and the specific detail.

Everyday wording, style, and content changes — including adding, removing, or reordering whole sections — get applied directly. The one category it won't touch is inventing a specific value for legal or compliance text, a license/certification number, or a real dollar price: those come back left blank with a plain note explaining why, rather than a confident guess. That's not a limitation to work around — it's the same honesty guardrail the rest of the product runs on, applied to editing.

Every section has two distinct pieces of text: a planning-only description (for your own review, never shown to a visitor) and the real draftCopy that actually renders on the page. If you ask to change what a section "says," that's the draftCopy — worth being explicit about, especially if a section's plan note and its real copy happen to disagree with each other.

Revision instructions are capped at 300 characters — anything past that is silently cut off before the AI ever sees it, with no warning shown to you. A long combined request risks getting truncated mid-thought, so it's worth splitting genuinely separate asks into separate revisions rather than stacking them into one long message.

That said, batching does make sense when the changes are really one idea — e.g. "rename the business to X everywhere" naturally touches several sections at once, and sending that as a single instruction keeps it consistent. The test: if you'd describe it to a person as "one thing," batch it; if it's really two unrelated asks, send two revisions.

Both the per-section "Adjust or remove this section" control and the broader revision thread route through the same mechanism, which is explicitly allowed to add, remove, or reorder whole sections when asked. If one section fundamentally isn't working — wrong angle, wrong tone, wrong emphasis — it's completely fine to ask for that section rewritten from scratch rather than trying to fix it sentence by sentence. Save the line-by-line nudges for sections that are basically right and just need a detail corrected.