Hacker News Reader

Comment by rcarmo | Hacker News Reader

AI has zero need to see the renders (you are just wasting tokens) and that is why I built the Go version - I patched a lot of the PPTX code in my MCP server, as it happens.

View on HN

I don't think so.

I have had a lot of experience creating and editing PowerPoint slides with Claude recently. It always converts the file to a PDF using LibreOffice and then renders the PDF into images to see if everything went right and that no text has overflowed.

That’s probably because you have (or implicitly created) a skill that does that. My agents know how to check bounding boxes.

while i appreciate that you are working on something to give away for free, providing your own little world of value... your comment makes it sound like you've never made powerpoints before. of course, there are a bajillion powerpoints out there in the wild that layer white rectangles to erase stuff from screenshots of charts, among numerous other atrocities

And if you task Claude with making a moderately sized powerpoint matching existing style guides, it will spend at most 30% of the time on the initial version. The remaining time is spent rendering out slides, looking at them and adjusting them

Of course AI can one-shot slides, but if you want good results where everything is aligned and has proper contrast you need a visual feedback loop

Yes, but automatically generated content doesn’t need to do that - like I wrote in another comment, my agents know how to check bounding boxes.

But bounding boxes don't always match what you want to do visually. Often you need to align to the baseline of text instead of its bounding box. And often you need to compensate for visual weight, making things different sizes so they appear more similarly sized to the eye

I'm not saying a pure API view won't work, but it will be a quality compromise compared to also having visual inspection

Font kerning software takes into account all of those things without writing to a pixel buffer and visually inspecting it.

Not saying that LLMs can do this, but there are algorithms which can, and neural networks are universal function approximators, so it is plausible.

People overestimate what LLMs can see vs what they tell you they can “see”

Reply on news.ycombinator.com