Mastering ChatGPT (OpenAI) · Custom GPTs, canvas and agents · lesson 12 of 19 · 14 min

Canvas: editing writing and code side by side

What canvas is

Canvas opens a document or code editor beside the chat. Instead of regenerating an entire answer to change one paragraph, you edit directly, highlight a section and ask for a change, or use built-in shortcuts. It is available across plans, and you can open it by asking ("open this in canvas") or from the tools menu. Canvas keeps version history so you can step back, and you can copy, download or share the result.

Use canvas for longer writing (proposals, articles, reports, scripts, SOPs) and code (scripts, formulas, small apps). Use normal chat for short answers.

Writing workflow: structure first

Structure: "Create an outline in canvas for a 1,200-word proposal: headings and one-line purpose per section."
Fill: "Draft section 2 only, using the attached case notes."
Target: highlight a paragraph and ask: "Make this more concrete with one example from the notes" or "Cut by 40% without losing the numbers."
Shortcuts: use suggestions for editing, adjusting length, changing reading level and final polish, then review each change.
Check: "List every factual claim in this document with its source in my notes; mark any without a source."

Targeted edits preserve the parts that already work. Full rewrites risk losing them.

Reviewing suggested edits

Canvas can suggest edits inline. Review them individually: AI edits sometimes change meaning, soften commitments ("will" becomes "may"), or introduce facts. Accept what improves clarity; reject anything that alters facts, tone you need, or legal wording.

Code in canvas

For code, canvas offers actions such as reviewing code, adding comments or logs, fixing bugs and porting to another language. It can also run some code (for example, Python) so you can see output. A safe loop for non-developers:

In canvas, write a Google Apps Script that copies new rows from the "Leads"
sheet to the "Follow-up" sheet when column F says "Hot". Add comments
explaining each step in plain English, and include instructions for testing
it on a copy of the spreadsheet first.

Then highlight a line you don't understand and ask, "Explain this line." Test on a copy, never on live data.

Facts, quotes and approvals

Never let AI invent quotes. Any quote attributed to a real person (your CEO, a client, a partner) must come from them and be approved by them.
Never let AI invent figures. Ask for [DATA NEEDED] placeholders.
Legal and compliance wording (terms, disclaimers, regulated claims) should be reviewed by the responsible person, not "polished" by AI.

Worked example: a press release

A PR consultant in Dubai drafts a product-launch press release in canvas. Structure first (headline, subhead, dateline, body, boilerplate), then section-by-section drafting from the client's brief. ChatGPT suggests a punchier CEO quote; she rejects it and sends the original to the CEO for approval, who edits it himself. She uses the reading-level shortcut on the body, rejects one suggestion that changed the launch date wording, and exports the final version for client sign-off.

Hands-on

Draft one real document in canvas using the structure-first workflow. Use at least three highlighted edits, check the version history, and note one suggestion you rejected and why.

Canvas vs Claude Docs vs Google Docs

Canvas is ideal for drafting and iterating with AI. For collaboration with colleagues, comments and approval workflows, you will usually move the final text into your organisation's document tool (Google Docs, Word, Notion). Keep the approved version there as the source of truth.

Five highlight prompts worth saving

| Highlight + prompt | When to use it | |---|---| | "Make this concrete: add one specific example from my notes." | Abstract paragraphs | | "Cut by 40% without losing any numbers or commitments." | Wordy sections | | "Rewrite for a busy CFO reading on mobile." | Executive summaries | | "Flag any claim here that needs a source; don't rewrite." | Before approval | | "Explain this line in plain English." | Code you don't understand |

Save them in your prompt library so every document gets the same quality pass.

Pitfalls

Regenerating the whole document for a one-paragraph change.
Accepting all suggested edits without reading them.
Letting AI "improve" quotes or legal text.
Running canvas-written code on live data.

How to measure success

Fewer full rewrites, faster time from outline to approved draft, and zero invented quotes or figures in anything you publish.

Video lecture: Canvas: editing writing and code side by side

Lecture coming soon · 16 chapters · about 8 minutes. Read the full transcript below.

Canvas
Why canvas matters
What canvas is
Structure first
Review suggested edits
Code in canvas
Facts, quotes, approvals
Simple example
Worked example: a Dubai press release
Where the final version lives
Pitfalls
Try this now
Key idea
Watch me do it, part 1
Watch me do it, part 2
Recap and next step

Lecture transcript

Canvas

Have you ever asked ChatGPT to change one sentence and received a completely rewritten document? Canvas fixes that. It gives you an editor beside the chat, where you can make targeted changes to writing and code. In this lecture you will learn a structure first writing workflow, how to review suggested edits safely, and how non developers can use canvas for small scripts.

Why canvas matters

Why use canvas? Because regenerating a whole document to fix one paragraph is how good work gets lost. The paragraph you loved gets rewritten along with the one you wanted changed. Canvas lets you edit like a professional editor, one section at a time, with version history if you need to step back. Think of it as the difference between repainting a whole room and touching up one wall.

What canvas is

Canvas opens a document or code editor next to the chat. You can type directly, highlight a section and ask for a change, or use shortcuts. It keeps version history so you can step back, and you can copy, download or share the result. Open it by asking, open this in canvas, or from the tools menu. Use it for longer writing and code, and normal chat for short answers.

Structure first

The best writing workflow is structure first. Ask for an outline with headings and a one line purpose for each section. Then draft one section at a time from your source material. Then make targeted edits by highlighting, make this more concrete with one example from the notes, or cut by forty percent without losing the numbers. Use shortcuts for length, reading level and polish. And finish with a fact check, list every factual claim with its source in my notes, and mark any without a source.

Review suggested edits

Canvas can suggest edits inline, and you should review them one by one. AI edits sometimes change meaning, soften commitments so that will becomes may, or slip in new facts. Accept what improves clarity. Reject anything that alters facts, the tone you need, or legal wording.

Code in canvas

For code, canvas can review code, add comments or logs, fix bugs and port to another language, and it can run some code, like Python, so you see the output. A safe loop for non developers is to ask for a small script, for example a Google Apps Script that copies hot leads to a follow up sheet, with plain English comments and testing instructions. Highlight any line you do not understand and ask, explain this line. And always test on a copy, never on live data.

Facts, quotes, approvals

Three firm rules. Never let AI invent quotes. Any quote attributed to a real person must come from them and be approved by them. Never let AI invent figures, so ask for data needed placeholders instead. And legal or compliance wording should be reviewed by the responsible person, not polished by AI.

Simple example

A simple example. Draft a cover letter in canvas. The opening and closing are good, but the middle paragraph is vague. Highlight just that paragraph and ask, make this more specific using the requirements in the job advert. Only that paragraph changes, now mentioning the exact tools and results the advert asks for. The rest of your letter stays exactly as it was. That is the core canvas skill, and it scales to proposals, reports and code.

Worked example: a Dubai press release

A PR consultant in Dubai drafts a product launch press release in canvas. She builds the structure first, headline, subhead, dateline, body and boilerplate, then drafts section by section from the client's brief. ChatGPT suggests a punchier CEO quote. She rejects it and sends the original to the CEO, who edits it himself. She uses the reading level shortcut on the body, rejects one suggestion that subtly changed the launch date wording, and exports the final version for client sign off. When the client's legal team asked for a change to the boilerplate, she did not regenerate anything. She highlighted the boilerplate section, pasted the approved legal wording, and asked canvas only to fix the formatting to match the rest of the release. Legal text stayed exactly as legal wrote it, and the rest of the document stayed untouched.

Where the final version lives

Canvas is ideal for drafting and iterating with AI. For collaboration, comments and approval workflows, you will usually move the final text into your organisation's document tool, such as Google Docs, Word or Notion, and keep the approved version there as the single source of truth.

Pitfalls

Four pitfalls. Regenerating the whole document for a one paragraph change. Accepting every suggested edit without reading it. Letting AI improve quotes or legal text. And running canvas written code on live data. Success looks like fewer full rewrites, faster approval, and zero invented quotes or figures in anything you publish.

Try this now

Try this now. Pick one real document you need this week, a proposal, a process document, a script. Open canvas and build the outline first, then draft one section at a time. Make at least three highlighted edits using the prompts from the lesson. Open the version history once so you know how to step back. And when canvas suggests an edit that changes meaning, a date, a commitment or a quote, reject it and write down why. That note is your reminder that you are the editor.

Key idea

Here is the key idea behind the structure first workflow. Changes are cheap early and expensive late. Moving a section in an outline takes five seconds. Moving it in a finished document means rewriting transitions, fixing references and rechecking the flow. So spend your attention on the outline, where changes are nearly free, and use highlighted edits for polish at the end.

Watch me do it, part 1

Let me draft the Dubai press release. I ask for an outline in canvas, headline, subhead, dateline, three body sections and the boilerplate, with a one line purpose for each. I move the customer benefit section above the technical details, because journalists want the story first. Then I paste the client's brief and ask for body section one only. The draft appears in canvas under the right heading. Nothing else in the document has been touched, which is the whole point of working this way.

Watch me do it, part 2

Now targeted edits. I highlight the long second paragraph and ask, cut by forty percent without losing any numbers. Done. Canvas suggests a punchier CEO quote. I reject it and put a placeholder saying CEO approval required, then email the original to the CEO. I apply the reading level shortcut to the body. One suggestion changes will launch on the third of March to is expected in March, which changes a commitment, so I reject it. Finally I open version history to confirm I can step back if needed, and export the draft for client sign off.

Recap and next step

Recap. Use canvas for long writing and code. Work structure first, make targeted edits, review suggestions individually, and keep real people in charge of their quotes and legal wording. Your next step: draft one real document in canvas with at least three highlighted edits, check the version history, and note one suggestion you rejected and why.

Key takeaways

Canvas provides an editable panel for longer writing and code alongside the chat.
Structure first, then fill sections, then make targeted highlighted edits.
Review suggested edits individually; they can change meaning.
Never let AI invent quotes or facts; real people must approve quotes attributed to them.

Try it

Draft one real document in canvas using the structure-first workflow. Use at least three highlighted edits and note one suggestion you rejected and why.