I didn't set out to build a carousel maker. I set out to kill a specific decision: "what do I post today."
That decision is small on any single day and brutal in aggregate. It's the reason so many small business Instagram accounts go quiet for three weeks, then post twice in one day to compensate, then go quiet again. Nobody decided to stop caring about the account. They just ran out of willingness to make that one decision, over and over, forever.
So the plan for Kohapu was: read a business's existing posts and offer, generate a batch of text-forward carousel scripts based on real content pillars and let the owner just approve, edit or skip. No blank canvas, ever.
The early decision I thought was final
Early on, I made a deliberate call: no AI-generated images. There were four reasons and at the time they all felt solid:
- Cost. Image generation at scale doesn't pair well with a generous free tier.
- The format already worked. Bold typography on a solid background is a proven carousel style. It doesn't read as "less professional." It reads as clean.
- Simpler approval. Deterministic, templated rendering means no "this one came out weird, try again" moments.
- The actual differentiator isn't visual. Kohapu's value is removing the daily decision, not making the prettiest slide. Pretty templates are a commodity. Any competitor copies that in weeks.
I was fairly confident in that call. Then I looked closely at what a direct competitor was actually shipping.
What changed my mind
I went deep on a competitor's carousels, not just the output but how they were actually building them. What I found wasn't a generic "AI image per carousel." It was an illustration per slide, semantically matched to that specific slide's content. A magnifying glass over an ingredient label for the slide about label myths. A scale weighing price against purity for the slide comparing the two. That's not something a fixed template system can fake. You either generate something that understands the specific idea on that specific slide or you don't.
I tested the middle ground first: a closed kit of reusable illustrations instead of true generation. It wasn't enough. The precision of a slide-specific illustration was doing real work. No amount of clever template-matching got close.
So I reversed the decision. Not because the original reasoning was wrong in the abstract (it wasn't), but because the thing I'd assumed was a nice-to-have turned out to be load-bearing for how good the content actually felt.
The part that made the reversal safe: the real cost wasn't what I feared
The reason I could actually make that reversal without wrecking the economics: once I mapped the real cost of a per-slide illustration, using a cheap style tier and a fallback chain for the text generation instead of a single expensive model, a full 5-6 slide carousel landed at roughly $0.03 to $0.08. Even for a business posting daily, that's a couple of dollars a month, not a line item that breaks a generous plan.
If I'd made the reversal on instinct alone, without checking that number first, it would have been the wrong call. The number is what made it the right one.
Where that leaves Kohapu today
Every carousel Kohapu generates now ships with real AI illustration, generated automatically alongside the text. It's not an extra step and it's not gated behind a separate click. See the full approval flow this feeds into, from the weekly batch down to scheduling and publishing.
The approach kept evolving after that, too. Once I prototyped clean typographic layouts for the slides that carry the main points (numbered points, quotes, stats, comparisons) and compared them side by side with illustrated versions, the structured versions won on those slides. So today the cover and a few key slides carry the illustration, while the slides in between are built for reading. Same lesson as the original reversal, applied one level deeper: decide by what makes the carousel better, not by what sounded right on day one.
The core bet hasn't changed: the hard part was never making a slide look nice, it's deciding what goes on it in the first place. I was just wrong, for a while, about how much the illustration itself mattered to whether that decision felt safe to approve.
I'm building Kohapu in the open. If you want to follow along or you're tired of staring at a blank Instagram composer every morning, give it a try.