I wish I got this level of productivity. I think every article should list exactly what they asked the LLM to do because I'm not getting as much use from it and I don't know if it's because what I work on is rare compared to say website front and backend code and/or if I just suck at prompts/context or I'm using the wrong services or don't have the correct MCPs etc....
I've found if you're working on something where it hasn't seen a billion examples you have to give it additional information like an academic paper or similar. And get it to summarize its understanding every once in a while so you can pick up the idea in another chat (once the context gets too long) without having to explain it all again but to also ensure it's not going off the rails ...as they tend to do.
They know a lot about a lot of things but the details get all jumbled up in their stupid robot brains so you have to help them out a bunch.
Is it possible to view the prompt history? I’ve had extreme levels of productivity and would love to list out how I’ve been using it generally for an article like this but it would be incredibly impractical to log it on the side.
With Claude Code at least, all of the chats you've had are stored in jsonl files on your computer in ~/.claude - I made a little TUI for exploring these in https://github.com/orta/claude-code-to-adium
Personally, I'm less sold on tracking prompts as being valuable both for production cases (imo if a human should read it, a human should have wrote/fully edited it applies to commits/PRs/docs etc) and for vibe cases where the prompts are more transitory
I really want to record a live commentary of me working with claude. Maybe that's something you could think about.
I feel like the results I get are qualitatively superior to anything I've seen anyone I've worked with produce. The fact that it's a lot faster is just gravy on top.
This is the saddest part of this whole thing for me. You consider the prompts and config to be the real source code, but those are just completely lost into the ether. Even if you saved the prompts you can't reproduce their effects.
Then there's the question of how do other developers contribute to the code. They don't have your prompts, they just have the code.
So, no, prompts are not source code, that's why I ask for people to just show the code they are producing and nobody ever does.
I also make my design documents (roughly the prompts generated by the prompts) into committed markdown documents. So I show the second-tier prompts at least, you could consider those an intermediate language representation if you like.
> Then there's the question of how do other developers contribute to the code. They don't have your prompts, they just have the code.
I usually try to commit the initial prompts and adjustments. I don't commit trivial things like "That's not quite right, try doing X again" or "Just run the entire test suite"
