Mike Taylor

<start of post>

I spent the weekend trying to break the new Claude model and I think I finally found its limit.

I started with my usual "PowerPoint benchmark"—asking it to generate a structured deck outline from a messy transcript. Usually, models hallucinate the slide count or miss the nuance of the speaker's tone.

Claude didn't just get the count right; it suggested three slides I hadn't even thought of that actually improved the narrative flow.

Then I got ambitious. I fed it 40,000 words of raw interview notes from a project I'm working on and asked it to find the "hidden thesis" that connects them all.

It cost about $14 in tokens and took less than two minutes. The result was so coherent it felt like I was reading a summary written by a senior editor who had spent a week with the material.

Accuracy of technical terms (95%+)

Tone consistency (Matched my previous writing)

Logic gaps (None found)

Funnily enough, when I showed the results to my co-founder, he thought I’d spent all Sunday writing it myself. I’m not sure if I should be impressed or worried about my job security!

If you're going to test this, don't waste it on simple summaries. Give it something messy, something you've been procrastinating on because it's too much "grunt work."

<end of post>