DEV Community

汪小春
汪小春

Posted on

Why we hardcoded 8 niche presets instead of letting GPT generate slide layouts

Most AI slide tools let GPT decide everything — the layout, the typography hierarchy, the section structure. Each generation is a new design lottery. We shipped the opposite approach: 8 hardcoded niche presets that GPT can fill but not redesign.

This post is about why constraint won over creativity for our slide-generation pipeline.

The problem with letting LLM design layouts

Early prototype: pure GPT layout generation. The model decides:

  • How many sections per slide
  • Title vs subtitle hierarchy
  • Bullet vs paragraph structure
  • Color emphasis
  • Asset placement

Result: each deck looked different from the next. "Different" sounds good until users started telling us:

  • "Why does the second slide have a 5-point list and the third has a 3-bullet hierarchy?"
  • "The font on slide 7 is huge, but slide 8 is tiny."
  • "Looks AI-generated."

The third complaint was the killer. When variance is visible, users default to "this AI doesn't know what it's doing."

The fix: pre-pick layouts, let LLM only fill content

We hardcoded 8 vertical-specific presets:

  • Career: hierarchy of pain → frame → action sections
  • Finance: chart-heavy with bullet clarifications
  • Reading: book cover + chapter quotes + 3-takeaway template
  • Beauty: image-led with overlay captions
  • Health: stats-forward with citation footers
  • Culture: timeline-style with accent imagery
  • Travel: map + photo grid + itinerary breakdown
  • Knowledge: 3-column comparison + "key insight" callout

Each preset is a deterministic layout system. GPT picks the right one based on the input topic, then fills slot content. The structural variance disappears.

Why niche-first, not generic-first

We considered the obvious alternative: 5 universal templates ("clean", "minimal", "playful"). It failed in user testing because:

  • "Clean" doesn't tell you what content goes where
  • The same "minimal" template applied to a finance deck and a travel deck both look generic

Niche-specific templates encode domain assumptions:

  • A finance deck's first slide should be a chart
  • A reading deck's first slide should be a book cover
  • A travel deck's last slide should be a map

These assumptions ride for free with the niche selection — no need to teach the LLM what each genre expects.

What we lose

We lose:

  • Flexibility for niches we didn't anticipate (business pitch, scientific paper, etc.)
  • The ability to experiment with novel layouts mid-deck

For both, our answer is "we'll add presets when there's clear demand" rather than "let the LLM figure it out". The latter is what failed in v0.

What I'd do differently

Each preset has a single "voice" — the same layout system applied throughout the deck. In hindsight, voice should vary by slide position (cover vs body vs CTA) within a preset, not just by niche. We'd ship "preset families" with intra-deck variation rather than treating each preset as a single template.

Try it

If you want to see what 8-niche-preset architecture looks like in practice, AnySlide ships the v1 of this. Free to start (60 credits at signup, daily +10 reset, no credit card).

I'd love to hear from anyone who took the opposite bet (full LLM creativity) — did it pay off?

Top comments (0)