100x Coding with Claude

Reading the article it mostly sounds like someone pretty new to an ecosystem leveling up to an ok level - so more like 0.1x developer (newb) to 1x developer than 1x to 100x.

There's a lot of value in skipping months of familiarizing oneself with a certain framework/language and immediately being productive but it doesn't read like the output of someone like Fabrice Bellard (a 100x in my book).

The trouble is large context uses a large number of tokens. I paid for pro but quickly ran out of tokens by lunch doing complex code. I soon closed the account and went back to open ai. Personally I think it's better but not much use in a professional environment.

ps. Not 100x if you are a professional. Probably 100x for a beginner. Great all the same.

> how to only focus on the hard development challenges and hopefully reduce the development time. You know, every developers dream.

And then the rest of the post describes the most tedious hand-holding of software that I can imagine. I'm glad I don't spend much time on boilerplate, so I get to skip this.

Max context of claude is a lot higher than 5 files when you use projects. However, the struggle i’ve found with projects is integrating changes rapidly. You end up reuploading files constantly.

Cursor.sh is a pretty good alternate but I find it a bit buggy at times. I think for me right now AI is great at:

- Quickly rolling an implementation of something in a completely unfamiliar API/language

- Debugging “conceptual” errors, where the problem is me not understanding something

- Playing “spot the typo”.

- Writing boilerplate you know must have already been done a million times.

It’s not so good at:

- Zero shot a conceptually elegant data schema or project structure

- Debugging obscure errors

- Debugging errors with misleading context provided by the user (aka, your hunch and supplied context is completely wrong and the AI gets confused as well)

> 10x developers will turn 100x developers. 1x developers will need to turn 10x and if not, they will perish.

If code behaves anything like games, we're going to knock humans out of developing entirely. I'm forecasting that "developer" will disappear as a job title and become a function of what we currently recognise as product. I would be a lot more comfortable in my career prospects if I thought being a 10x developer is good enough.

Once these models become good enough that human review of the code reduces the quality - which will probably take a little while, but looks like it is incoming - then we can probably engineer out the monkey in the middle copy-pasting code.

Zed (https://zed.dev/) does pretty well at IDE LLM integration. Haven't done anything extremely ambitious with it but the editing experience is much smoother than copy pasting from a browser.

this is what cursor is for! cursor.dev