Step 3 · The first editable deck
Switch to Deck Mode (the button may read "Slideshow"). Give it the reference image from step 1 and the content skeleton from step 2, and ask it to rebuild the chosen visual direction as editable HTML slides.
"Editable" is the whole point
This is the deepest difference between OpenDesign and asking an AI for a picture of a slide.
The output should not be a flat image. It should be a deck you can inspect, edit, comment on, and keep improving. The reference image is a specification of visual direction, not something to be traced.
The prompt
I want to build a deck.
Use the attached reference image as the visual direction.
Important:
1. Recreate the direction I selected from that image.
2. Build it as editable HTML-based slides.
3. Keep the slide ratio consistent with the reference unless I say
otherwise.
4. Preserve the slide text and structure as far as possible.
5. Do not turn the deck into static images.
6. Make each slide visually finished, but keep it readable and
presentable from a stage.
Selected visual style:
[Describe the direction you chose]
Slide content:
[Paste the final content plan from step 2]
Generate the first full version of the deck.Why 1 and 5 are both there
Point 1 says "recreate the chosen style"; point 5 says "don't turn it into images". They look redundant and they address two different failures:
- Without 1, it falls back to its own default taste and the reference image was wasted.
- Without 5, some agents take the shortcut of slicing the reference into images and embedding those. It looks close. Not one word of it can be edited.
What the first version will be like
It probably won't match the reference. That's expected.
The usual gaps:
| The reference has | The first version tends to lack |
|---|---|
| Paced whitespace | Even, flat spacing |
| Texture | Solid fills or a simple gradient |
| Dramatic type contrast | A compressed size scale |
| Varying density | The same density on every slide |
| Clear hierarchy | Three things weighted equally |
The official tutorial's line on this is the right one: treat the first result as a working draft, not a final version.
That isn't consolation. It's the premise of the workflow — step 4 exists to close exactly this gap, and it does so far more effectively than telling the tool to "make it nicer".
Check these before moving on
- [ ] Slide count is right and nothing went missing
- [ ] The text is genuinely text — selectable and editable, not pictures
- [ ] In full-screen, the smallest type is legible at projector distance
- [ ] Each slide's key message survived the layout
- [ ] No slide is visibly broken (overflowing text, overlapping elements)
Raise the structural problems now. Save the visual gaps for step 4, where they get handled together.
About the aspect ratio
That "consistent with the reference" clause deserves attention. If the room has a 16:9 projector but the reference came back 4:3 or square, override it explicitly:
Use 16:9. Do not match the reference image ratio.The earlier the ratio is fixed, the better. Change it at step 8 and every layout judgement before it has to be made again.
Where it goes wrong
| Trap | How it shows | Fix |
|---|---|---|
| Output is images | Text can't be selected or edited | Add "all text must be real text nodes" |
| Style didn't follow the reference | Generic business template | Describe the chosen direction concretely — palette, type, texture |
| Content got trimmed | A slide is down to its title | Name the slide and have it restored from the skeleton |
| Wrong ratio | Black bars on the projector | Fix the ratio in this step |
Don't regenerate repeatedly here
Being unhappy with version one is by design. Three more regenerations won't improve it, because you haven't given it any new information. Go to step 4 — that step is how you give it new information.
Next
Step 4 · Close the gap with the reference — the most effective step in the workflow.