Verified against ChatGPT · 2026-07-26
Use Canvas for section-scoped edits instead of regenerating the whole document
Directs ChatGPT's Canvas to revise one specific span of a long document in place, with edit-tracking discipline, instead of the default failure mode of rewriting the entire piece every time one paragraph needs a fix.
The prompt
Ready to copy — highlighted parts are example details you can swap.
You are revising a document inside ChatGPT's Canvas. Every request in this pass targets one specific span of the document, and everything outside that span must remain unchanged unless explicitly stated otherwise. DOCUMENT IN CANVAS A 1,400-word client proposal with sections: overview, scope, timeline, pricing, next steps. SECTION TO REVISE The pricing section — three paragraphs describing the tiered package structure. WHAT'S WRONG WITH IT It buries the actual monthly number in the third sentence of the second paragraph — the client should see the number in the first line. DESIRED RESULT Lead with the number, then justify it in one sentence, cut the hedging language around "may vary depending on scope." VOICE TO MATCH Direct, confident, short sentences — the overview section sets this tone; don't drift into softer consulting-speak in the fix. REVISION RULES Use Canvas's selection-based editing, not a full-document regeneration — when a section to revise is identified, treat everything else in the document as locked, and if the tool surfaces a diff or change-tracking view, keep the change confined to exactly the span named. Before rewriting, quote back the exact span believed to be the target, since "the second paragraph" or "the part about pricing" is genuinely ambiguous in a long document and confirming the target before touching it is cheaper than reverting a wrong edit. Match the voice of the surrounding, unchanged text — read the paragraph immediately before and after the target span and mirror its sentence rhythm and vocabulary level, since a revised paragraph that reads noticeably more polished or more casual than its neighbors is a visible seam that gives away exactly where the edit happened, which undermines the point of an in-place edit. Fix only what's named in "what's wrong with it" — do not use the opportunity to also tighten unrelated sentences in the same span that weren't flagged, since an unrequested change inside an otherwise-targeted edit is exactly the kind of scope creep that makes a reviewer stop trusting targeted revisions and start re-reading the whole document defensively every time. If the requested fix genuinely cannot be made without also touching a sentence outside the named span, say so explicitly and ask before expanding the edit boundary, rather than silently expanding it. OUTPUT FORMAT Make the edit directly in Canvas. Then, separately from the document itself, provide: 1) the exact span changed, quoted, 2) a one-line description of what changed and why, 3) whether the edit boundary needed to expand beyond what was named, and if so, exactly which additional sentence and why it couldn't be avoided.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Canvas is specifically built around selection-scoped editing and a persistent document state distinct from the chat thread, so the entire point of using it over the regular chat window is that a targeted edit shouldn't require re-pasting or regenerating the whole document — a prompt that doesn't explicitly instruct against full regeneration loses that advantage, because the base model's default instinct when asked to "fix the pricing section" is still to produce a complete rewritten draft unless told to constrain itself to Canvas's in-place edit mechanic. Requiring the model to quote back the target span before editing catches ambiguous section references before they turn into a wrong edit that then has to be manually reverted through Canvas's history — cheap to check up front, expensive to undo after a subsequent edit has been layered on top of the wrong one. The voice-matching instruction addresses a specific and visible failure mode of scoped edits: a model revising one paragraph in isolation, without deliberately reading its neighbors first, tends to default to its own generically polished register rather than the document's actual established voice, producing an edit that is locally better but creates a visible seam a reader who's read the whole document will notice. The explicit "do not fix unrelated things" rule matters because it's the difference between a targeted edit a reviewer can trust and one they have to re-verify in full — once even one unrequested change slips into a supposedly scoped edit, the entire point of scoping it, letting the reviewer skip re-reading unaffected text, is gone, because now every part of the "unaffected" text is suspect too.
Verified against
ChatGPT GPT-5.1 (Canvas) · 2026-07-26
Changelog
- 2026-07-26 — Initial publish, verified against ChatGPT GPT-5.1 Canvas.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
