DEV Community

Cover image for The silent layout bug in AI-generated slides
rivi mizuiro
rivi mizuiro

Posted on

The silent layout bug in AI-generated slides

I often generate slides by summarizing documents or PDFs.

The workflow itself is convenient, but I kept running into the same issue:
parts of the generated slides were silently cropped.

What made this tricky was that the slides usually looked fine during editing and review.
The overflow only became obvious after export — or worse, during the actual presentation.

After missing this a few times, I realized the problem wasn’t how I generated the slides,
but how hard it was to notice when something was already broken.


Why layout issues are easy to miss

Slide layouts depend on many factors:
screen size, font rendering, code block wrapping, and export targets.

Even with careful review, it’s surprisingly easy to miss small layout failures.
If everything mostly looks fine, our attention moves on.

This problem becomes worse when slides are generated automatically.
There’s often no strong “this looks wrong” moment, and broken output can slip through silently.


The real problem: detectability, not fixing

At some point, I realized the hard part wasn’t fixing layouts.

It was noticing failures early enough to matter.

Before fixing anything, you need a reliable signal that something is wrong.
Without that signal, both humans and automated systems tend to miss problems.


A small experiment

To explore this idea, I built a small CLI tool that tries to detect layout overflows
in Slidev presentations.

It’s intentionally heuristic-based and far from perfect.
The goal isn’t to guarantee correctness, but to make layout failures
machine-detectable early in the workflow — for example, in CI or automated pipelines.

If you’re curious, the repository is here:
https://github.com/mizuirorivi/slidev-overflow-checker


Takeaway

This project reminded me that many real-world problems aren’t about generating better output,
but about making failures visible.

Especially in AI-assisted workflows, the absence of clear failure signals
can be more limiting than generation quality itself.

I’m curious how others handle layout or visual validation
in generated documents or presentations.

Top comments (0)