A good carousel or infographic is mostly a writing job with a layout on top, and AI image tools are weakest at exactly the part readers notice first: the words. So the working order is copy first, layout second, pictures last, with text added in an editor you control. This post gives prompts for each step, a small script that catches overstuffed slides, and a proofreading routine for when you do let a model draw text.
Status of the prompts: untested templates. I didn't run image generations for this post, so there are no sample slides and no claim about any tool's output. The slide-length script below I did run, and its output is real.
What should a carousel say before anything is designed?
Start with the argument. A carousel is a sequence: hook, three to seven points, one takeaway. Ask a text assistant for copy in a fixed structure so you can check it.
Turn the notes below into a [LINKEDIN or INSTAGRAM] carousel of [7] slides for [AUDIENCE].
Rules:
- Slide 1: a hook headline of 8 words or fewer that promises a specific result. No clickbait.
- Slides 2 to 6: one idea per slide. Format each as a headline (8 words or fewer) on the first line, then a body of 30 words or fewer.
- Last slide: one takeaway and one question or action for the reader.
- Use only facts and numbers present in my notes. If a slide needs a number I did not give, write "ADD NUMBER".
- Separate slides with a line containing only ---.
Notes:
[PASTE YOUR NOTES, STEPS OR DATA]
The "ADD NUMBER" rule is there because assistants will happily invent a statistic to make a slide punchier. Anything with a number, a percentage or a named study needs a source you've checked before it goes in a graphic that people will screenshot and share.
For longer written posts that you want to adapt, see LinkedIn content prompts and content repurposing prompts.
How do you catch overstuffed slides?
Paste the assistant's output into a text file and run this checker. It splits on ---, treats the first line as the headline, and flags slides over the limits. The limits (8 and 30 words) are a starting rule I chose, not a platform rule. Change them after you see your slides on a phone.
import sys
MAX_HEADLINE_WORDS = 8
MAX_BODY_WORDS = 30
text = open(sys.argv[1], encoding="utf-8").read()
slides = [s.strip() for s in text.split("\n---\n") if s.strip()]
bad = 0
for i, s in enumerate(slides, 1):
lines = s.splitlines()
head, body = lines[0], " ".join(lines[1:])
hw, bw = len(head.split()), len(body.split())
flags = []
if hw > MAX_HEADLINE_WORDS:
flags.append(f"headline {hw} words (max {MAX_HEADLINE_WORDS})")
if bw > MAX_BODY_WORDS:
flags.append(f"body {bw} words (max {MAX_BODY_WORDS})")
status = "OK" if not flags else "FIX: " + "; ".join(flags)
bad += bool(flags)
print(f"slide {i}: {hw}+{bw} words {status}")
print(f"{len(slides)} slides, {bad} to fix")
I ran it on 2026-10-09 against a three-slide sample I wrote (one good slide, one with no separate headline line, one with a nine-word headline). Real output:
slide 1: 5+9 words OK
slide 2: 35+0 words FIX: headline 35 words (max 8)
slide 3: 9+7 words FIX: headline 9 words (max 8)
3 slides, 2 to fix
Slide 2 shows the most common real failure: a paragraph with no headline. The checker reads the whole paragraph as the headline and flags it. Ask the assistant to rewrite only the flagged slides rather than regenerating the set.
How do you prompt the visuals without letting the model write text?
Generate backgrounds, icons and illustrations with an explicit "no text" instruction, and put every word in yourself. Keep one style block and reuse it word for word so the slides match.
Style block (reuse exactly): flat vector illustration, soft [COLOUR 1] and [COLOUR 2] palette on an off-white background, rounded shapes, generous empty space, no gradients, no text, no letters, no numbers, no logos.
Slide visual: [A SIMPLE SCENE FOR THE IDEA, e.g. an invoice with a checkmark and a calendar]. Leave the top third of the image empty for a headline. Portrait 4:5.
Three habits that save time:
- Ask for empty space. "Leave the top third empty" gives you a place for text and avoids the model putting scribbles there.
- One idea per image. A prompt that lists five objects gets five half-drawn objects.
- Generate icons separately. For a process with four steps, ask for four matching icons on a plain background in one image, then cut them apart in your editor.
For an infographic, decide the structure before the picture: a numbered list, a comparison of two columns, a timeline or a simple flow. Build that structure as native shapes and text in a slide or design tool, and use AI art only for the decoration. Charts with real data should come from your data, not from an image model. A generated bar chart is a drawing of a chart, and its bars will not match your numbers. If you need a chart, make it in a spreadsheet and drop it in.
What if I do want the model to draw the text?
Sometimes you want a stylised headline in the artwork. Then reduce the risk:
- Keep it to one to four short, common words.
- Put the exact text in quotation marks in the prompt, and say where it goes and how large.
- Avoid numbers and small print where you can. I haven't measured which kinds of text fail most for any tool, so test your own strings.
- Generate several versions and pick the one that is spelled correctly rather than iterating on a bad one.
- Proofread by reading backwards, letter by letter.
If you're choosing a tool for this, run the text-rendering comparison method with your own words. It takes a few dozen generations and tells you more than a general ranking would.
Proofreading routine for every slide before posting:
| Check | How |
|---|---|
| Spelling | Read each word aloud; check names and brand names against their sources |
| Numbers | Compare every figure to your notes, digit by digit |
| Consistency | Same capitalisation, same units, same currency symbol |
| Contrast | View on a phone at arm's length, in bright light if you can |
| Edges | Nothing important within a margin of the slide's edges |
| Alt text | Write a plain description of each slide for accessibility |
How do you adapt it for LinkedIn and for Instagram?
The content can be the same while the framing changes. Platform specs for sizes, file types and slide counts change, so check each platform's current help pages before you export; I haven't verified current limits for this post and so give none.
A prompt for the adaptation step:
Here is my carousel copy. Make two versions of the first slide headline and the caption.
LinkedIn: professional, specific, first line of the caption states the takeaway, no hashtags in the first line, 3 hashtags at the end.
Instagram: more conversational, caption opens with a question, a short list of what the carousel covers, 5 hashtags.
Do not change any facts or numbers.
[PASTE COPY]
Then keep your visual identity consistent with a brand kit. If you don't have one, the logo and brand kit workflow covers the colour, font and logo basics that make a set of slides look like one account. For food or product imagery to drop into slides, see AI food photography prompts and AI product photos.
A repeatable routine
- Write notes with facts you can source.
- Generate copy in the fixed structure; mark missing numbers.
- Run the slide checker; rewrite flagged slides.
- Generate text-free visuals with a reused style block.
- Assemble in a design tool with real fonts; add text yourself.
- Proofread with the table above, then read the whole carousel once as a stranger would.
- After a few posts, compare saves, shares and swipe-through on your own analytics. Don't assume a format works because someone online says so.
The shortcut is tempting: one prompt, one finished carousel. It works for mood boards. For something with your name on it, the extra twenty minutes of assembly pays for itself the first time you'd have published "Recieve payments".
