Comment by benreesman | Hacker News Reader

Comment by benreesman | Hacker News Reader

Parent Full thread

Sure. I'm experimenting like everyone else, but I mostly use gptel as the primary interaction surface and Claude Code for a range of refactorings and other "more than mechanical, less than creative" edits. Both of these are very (!!!) well complimented by magit, which is so good at AI supervision it seems designed for it, by a genius.

For Claude Code I'm rapidly switching anything I want "vibe coded" into Hadkell for code, Dhall for config, and check-heavy Nix for deploy. Between that and some property tests that I seed and then have Opus elaborate on, you can put Claude Code so restrictive that it just hits the compiler in a loop until useful code comes out. Its trapped: I hoist CLAUDE.md in from the Nix store so it physically can't edit out the brutal prompts around mocks and lies, and -Wall -Werror in GHC gives it nowhere to hide, all it can do is burn tokens and desperately Web Search tool until it gets it perfect ish or I cut off its money because this requires a real LLM minimum and likely a real programmer. If there's a property test welded into the type system it can't even fail to use a parameter: that's an error Claude.

I have a bunch of elisp in // hypermodern // emacs for things like OpenRouter integration and tool use, but frankly, stock gptel is so strong I always wonder if I'm getting my money's worth trying to tune it into the asymptote.

Happy to answer any more questions.

Sounds wild! What have you built this way?

Also as another Emacs user I'm wondering what lesser known packages or elisp snippets do you use? gptel, magit, tramp and org-mode are the usually touted killer features, but what else do you use in the Emacs ecosystem?

Sorry, I saw another commenter ask about the dots but for some reason didn't see this one, all the key files are linked as gists here: https://news.ycombinator.com/item?id=44817968

Let me know if you have any questions (or suggestions for that matter, it's rough in places).

Thanks for sharing!

Share your dotemacs/gptel config? I'm not in love with emacs eider integration. Wondering how to put direct editing/control to the model. Still very cludgy with gptel though I've been using it for months

This is the fairly "cut down" one I alluded to without my mixed-results heavyweight AI integration stuff, this is like a gist of my open dev box session so it's got random shit commented out and stuff, but I think most people's config looks like that point in time:

https://gist.github.com/b7r6/84c6ab80c0b8bd5267b8c436e4d00a8...

https://gist.github.com/b7r6/23cfacbf181c9b0447841c798345a79...

The AI stuff doesn't work without this running:

https://gist.github.com/b7r6/449faab9b5be00867f2e8053c610bdb...

That lets me publish my API vendor keys without issues:

https://gist.github.com/b7r6/fe96bd0cc37d72c1991d84d1984371b...

Sounds cool. What sort of stuff do you develop, and who's paying for it?

I work for a medium-sized proprietary/discretionary fund. AFAIK the principles trade all kinds of stuff, macro stuff. My current job is tuning up the execution on the cryptocurrency adjacent desk, but not like blockchain stuff, it's somewhere in between OG crypto trading stuff and like Wall St. HFT circa 2006-2010 depending on how you measure, it's in the "kernel bypass matters but FPGAs are still exotic" sort of regime, some of it is legacy REST APIs still but FIX 4.2 SBE and other real finance protocols (and real banks and stuff) are starting to be a part of the ecosystem.

I aspire to be a lot faster than this stuff (I've built faster stuff than this) but this is quite a good library (amazingly good by OSS standards, good stuff in this area is rarely OSS, props to the maintainers): https://github.com/crypto-chassis/ccapi, in particular this library does a really good job of being correct across a lot of surface area, it's serious people doing it, and there are forks of it that use DPDK floating around.

If by who's paying for it you mean the big Anthropic bill? My boss's boss is pretty enlightened about the fact that learning how to use AI well is expensive, so when I'm on a tight schedule I get a pretty forgiving budget for the model fees. It's a pretty serious perk in the sense that it's really expensive to master using these things :)

Thanks.

I have a couple more!

I take it this is all back-end work? Have you tried out one of the Haskell-y front-end languages? Elm?

> Both of these are very (!!!) well complimented by magit, which is so good at AI supervision it seems designed for it, by a genius.

Can you expand a little on this point?

I very much recommend just watching some of the great `magit` videos on youtube, but later on when I have time I'll do a little `asciinema` of like, a Claude Code interaction and reviewing / piecewise incorporating the bots changes, so if you check back here tonight or tomorrow latest I'll do a little demo.

Did you get the chance to record this? I'd quite like to see it.

Reply on news.ycombinator.com