Canva Recontent

Canva re-content

Needs the Canva connector: copy-design, read-design, edit-design.

A finished deck is a template: someone already solved the layout, the rhythm, the progressive reveals. Re-contenting keeps that solution and replaces only what it says.

Text and speaker notes go through the API cleanly. Images, page deletion and animations do not — those become a gaps file the human finishes by hand. A run that pretends otherwise ships a broken deck.

Before you start

Confirm the design is one the user owns, has a licence to reuse, or was given by its author for this purpose. Templates and a colleague's shared deck qualify. If provenance is unclear, ask once.

Steps

1 · Copy first. copy-design with the source id; every later operation targets the copy. The original stays as found, which is what makes a bad result cheap to throw away and rerun. Done when you hold a new design id.

2 · Survey before you touch anything. One read-design with filter.fields:["presenter_notes"] across the full page range shows which pages are already converted — cheap, and it catches resumed work. Then one read-design with filter.fields:["design_content"] (no transaction) returns plain readable text for many pages at once. Field-scoped reads stay small; the full element JSON of step 4 does not. Done when you can name what is on every page.

3 · Write the page map before editing. A table: page number, what it currently holds, what it becomes, which source content fills it. Show it to the user. Done when every page in scope has a row, including the ones you plan to leave alone.

4 · Edit one page at a time. Read the page with open_transaction:true, run scripts/canva_peek.py on the saved payload for a compact locator listing, then batch replace_text plus replace_speaker_notes.

Every operation reports applied_unverified, which says only that the call was accepted. The response also carries a thumbnail URL rendering the page with the uncommitted edits on it. That image is the evidence; the status field is not. Done when the thumbnail shows the new text with nothing clipped or overlapping.

5 · Commit every five pages. finalize:"commit" with an empty operations array, then reopen a transaction. Long runs hit session limits, and uncommitted work dies with them. Done when the commit returns and every page behind you is in the gaps log or verified complete.

6 · Log every gap as you meet it. Append to the gaps file — see templates/manual-gaps.md. Done when every image you left, page you would have deleted, and element the API refused has a row naming the page and what to do.

Rules that bite immediately

• Locator ids are full paths. PAGE-LBxxx, or PAGE-GROUP-LBxxx for anything inside a group. A bare element id fails with "Page ID did not match the expected pattern", which names the wrong thing and costs a retry.
• Read one page per call. Four pages of element JSON runs past 60,000 characters and the call fails outright.
• Reach for replace_text first. It inherits the element's existing styling. add_text lands at 16px and needs a format_text pass behind it.
• Keep every position, size, colour and font as found. Move something only when the thumbnail shows it clipped, and log the move.

references/api-reference.md holds the rest: silent failures, element types that refuse edits, what replace_text does to list markers, and the operations that do not exist.

Long decks

Edit responses return the whole page document — roughly 12,000 tokens each. Past about fifteen pages this fills the context window faster than the work progresses. Dispatch one subagent per section so the responses land in their context, not yours. references/orchestration.md has the briefing pattern and what each agent must report back.

Finishing

Hand back three things: the design URL, the gaps file, and a one-line honest statement of what is not done. The gaps are the deliverable's other half — a user who thinks the deck is finished will find the original author's logo on a slide during their talk.