I have two kinds of hobby code projects;
1) where I enjoy choosing my tools and writing everything as perfect and beautiful as I can,
and
2) where I’m just enjoying hacking or learning something, and also the fact that I can write as much spaghetti as I want.
I love both 😅
Two things I love most about working with LLMs;
1) it's immediately clear if your code is unintuitive or written in a non-obvious way as LLMs struggle with it no matter how much you try to document it
2) following best practices finally pays out in full 🎉
Claude Design model selector shows me Opus 4.8, Opus 4.7, and Opus 3 (!?) but not 4.6 -- which is the only usable one from the latest models 🤔
The product itself is not that great but the results I get from it are AWESOME. Superb tool for a talentless designer like me.
The killer feature of Pi is that it can modify and debug itself. All software should have this.
Would happily pay to be able to say "fix the ts/eslint/prettier/oxfmt Vue configs - format on save broke again" in VS Code and Zed. So annoying to waste hours on editor configs 😩
I feel like the greatest limitation is the Claude Code harness itself, it's just hands down the WORST available tool at the moment. Combined with ridiculous pricing and usage limits, Opus 4.8 won't be in my daily arsenal - gpt-5.5 is just better. But good for docs and plans! ✌️
First few 5h sessions w Opus 4.8:
1) Extremely annoying and hard to work with due to unpredictable behavior. Ask it to "load plan implementation skill", it loads it and then starts using git and searching for plans, picking one *randomly* and just start doing stuff. (!!?!?) ..
3) it does feel marginally smarter, tho. Haven't found ANY real world tasks so far that it didn't know how to solve after letting it do its thing. I appreciate that it can dive extremely deep in a Sentry issue, pull relevant bits from two repos and prod db and then just "Fixed."
@badlogicgames as long as we still see if it was a partial or full read. that is kind of importan info to know as some models do idiotic things like read 200 lines of a 500 line skill and then keep failing bc they missed most of the instuctions 🫠
Feeling vindicated bc I never took the time to learn cookiecutter properly, always kept manual template repos.
Now I don't have to anymore; it's so fast to let GPT-5/low loose on a handwritten template repo and tell it exactly what you want 🎉
WKND project coming up!
I remember a time when Netflix had "share your logins with a friend" ads. Getting the same vibes with this Anthropic shit. They have enough corporate clients now to not care about smaller clients anymore.
CC is now by far the worst harness but Opus 4.6 is still my fav model 🤷♂️
It's bonkers to think back how AI workflows have changed in the last couple of years. Getting a function to copy-paste felt awesome a year ago before CC era started. No I'm bummed when _some_ of the code produced by a 80 min agentic session is sub-par 🤯
Slopping w Claude, part n
Me: hey Claude (Sonnet / High), wasnt this issue already solved?
Claude: nope
Me: properly recheck, what were the last commits about?
Claude: related but didn't fix this
Me: ok, fix it
Claude, ~1h later: the PR wont merge bc main has the code already 💀