@jmoiron.bsky.social replied: I've had a similar "humans are bad at code reviews" insight in the past but came to a different conclusion, which is that we think their value is technical but it's actually primarily social. An AI reviewer won't mind the tedium, but I love your insight that they're too deferenti...
jmoiron plays the blues
@jmoiron.bsky.social My father is from a town about 15km away from this vineyard! My uncle has an orchard nearby.
My benchmark for Portuguese whites at this price point is Pera Manca, but this is going on the watch list. Dão has some surprisingly good whites IMHO.
jmoiron/monet 2bb3b89 update lesscss, add a loader to the content-rendered section for slower connections as the wasm is like 6 mib
@jmoiron.bsky.social This happens a lot in software engineering. Normally, experience teaches you humility and you learn to assume you're just missing something.
If, however, you are a narcissist, you will never build the self awareness to score this game appropriately, hence DOGE, Twitter, and the ...
jmoiron/monet a390f01 migrate off of justvector, update fontawesome, minimize the woff to make it much more light weight, update social links and icons
jmoiron/monet 09e9c72 font extraction and combination across different source woffs.. the tool might be done?
jmoiron/monet 20a8101 claude updated its own md spec, started on the tool, we have a cli interface, sample spec, and spec parsing
@jmoiron.bsky.social replied: Finally, there is a lot of hype around AI doing these kinds of tasks in "an afternoon", but outside of running many agents in parallel, I'm not seeing it. It's a fun workflow when the stakes are low, but I'd hate to be in the position where I would feel compelled to use it to hit...
@jmoiron.bsky.social replied: Also, my time estimate for myself would involve some intense work to put that much code down. It would feel good, but it'd be tiring too. Keeping Claude on task is far less taxing work.
@jmoiron.bsky.social replied: If/when? these models are 10x or 100x faster, this will absolutely be quicker than hand-writing code. As it currently stands.. it feels like it's maybe 85% as fast as I could manage for the whole project, but that includes a lot of time waste putting it back on the rails.
@jmoiron.bsky.social replied: Initially I was very impressed, but I made a mistake by giving Claude a black box description of what I wanted, and I spent a few curious hours with it going down a garden path, fascinated to see if it might actually lead anywhere. It didn't.
@jmoiron.bsky.social replied: Claude's performance degrades noticeably as your project grows (it gets distracted easily), so creating software in independent units and then integrating them in a parent unit that is itself simple is key. This is already good practice but it will be even gooder practice w/ LLMs...
@jmoiron.bsky.social replied: The writer of those specifications would really need experience both building the type of software they want Claude to make AND experience with some of the dumb things Claude likes to do to be successful.
@jmoiron.bsky.social replied: I can totally see how an "agent swarm" approach could produce fairly sophisticated software successfully and to specification, with "manager agents" that prevent it from indulging in its bad habits and break it out of cycles.
BUT...
@jmoiron.bsky.social replied: It has some "bad habits" that it will consistently get stuck on, consuming time and precious tokens unnecessarily. The tool I was building was a Go tool, and it kept trying to create little one off files to test with `go run`, which can't work, then back out of that and write a r...
@jmoiron.bsky.social replied: It can be surprisingly insightful about bugs and I think it shines as a debugging companion. Its internal narrative is close to the way I approach programming so it's very effortless for me to follow along.
@jmoiron.bsky.social replied: My usage has coincided with limit changes made by Anthropic, and I've found it fairly easy to hit the 5hr token limit on the $20 plan. If I am using it steadily for ~2h or so, I will hit the limit. I'm not using it consistently enough to hit the longer term limits.
@jmoiron.bsky.social replied: Claude is far superior to the free ChatGPT, DeepSeek, and the open qwen models at generating code that builds. I'd say that roughly ~80% of the time, the code it generates builds, and 80% of the other 20% it fails due to minor errors like imports.
@jmoiron.bsky.social replied: A quick description, this project involves creating a parser for a schema format, and then a validator to apply that schema format to json files and determine if they follow it or not.
I'm not a parser expert, but I've written dozens of parsers so I know my way around.
Anyway, ...
@jmoiron.bsky.social I've been using Claude code $20 edition for the past week to build a simple but non-trivial tool.
I might write up something more structured and with some data when/if I finish, but I have a few observations I'm ready to make about it now.
@jmoiron.bsky.social replied: So I guess I disagree that the pattern is "just bad", so congrats on a real unpopular opinion :) But personally I think it's very inconclusive whether it's worth using and I'm sympathetic to the idea that we shouldn't bother.
@jmoiron.bsky.social replied: By exporting the config struct, this gets messier. You will need a Config.Foo and Config.FooT for the deprecation period, and if it's valid for the user to set these to the zero value, detecting deprecated usage is basically impossible.