Put the paper and implementation side by side: this Claude Desktop use case is worth trying

Developer Kat shared a learning setup: open a book or paper PDF in Claude Desktop's browser panel, highlight passages, and ask Claude directly. At the same time, open a minimal implementation of the idea alongside and play with it while reading. Kat's keyword for this combination is immediate—both ideas and intuition should be immediate.

The video below shows this interface: PDF content on the left, and charts, dot plots, and a control panel on the right.

This workflow isn't complicated. On the left are concepts from the paper; on the right is code or interaction you can touch. When you hit something unclear, immediately highlight it and ask Claude; when you get an explanation, immediately change a line in the implementation and see the result. There's no switching cost between reading and doing, so understanding doesn't easily stop at terminology.

The comments ask more specific questions. Omar asks how to get this implementation running. He says getting Claude to build an MVP from a paper sometimes works and sometimes doesn't, depending on the paper. Kat doesn't give a set of prompts; instead she asks back: what do you think Claude's implementation from the paper is missing? That counter-question is more useful than a technique. First determine whether the gap is in concepts, interfaces, data, or constraints, then decide how to fill it. At least this shows that the focus of that reply is the gap between the implementation and the paper.

Someone else says they plan to build this workflow themselves. tryingEveryThing mentions using the pdf++ and excalidraw plugins, showing that the tool combination can be swapped. Claude Desktop's browser panel is just one approach. A Japanese reply adds: while reading, don't just read—build a small implementation on the spot to verify it, and keep the code you've tried, so later it's easier to turn knowledge into your own tools.

Someone in the comments mentions swarmalators and vortex patterns in the paper, says the patterns look good, and wants to steal this setup. This detail also shows that reading while playing suits papers with visual feedback. When you see a figure, immediately reproduce a simplified version in the implementation; it's faster than guessing from the text.

Next time you read a paper, try a minimal version: open the PDF, open an empty project, and when you reach a key passage, ask Claude to explain it and write a runnable minimal example. If it doesn't run, ask it to compare against the paper and find the differences. Full reproduction is heavy; a small implementation is enough.

Original tweet: https://x.com/poetengineer__/status/2114086088969245098