Claude helps a ton with mundane tasks like data analysis, which genuinely helps when you're developing a commercial software which math engine is supposed to exactly follow an another commercial software (I tell you what, how the HELL did PC-DMIS get a PTB certification is a myth to me) but I digress. The point being - I have never used Claude Opus 5.5 for personal projects and recent results from the software at work kind of made me wonder what Opus can do for something more, uh... lower-level?
I kind of let Opus 5.5 analyze the existing Vex source code and make a document artifact with all the stuff found categorized into critical, serious and architectural issues, and then let it figure out a way to solve these issues, so quite a good number of commits are Assisted-by: Claude Opus 5.5. Rest assured though, I try to ♿ ♿ ♿ review, clean up and understand ♿ ♿ ♿ the code it makes...
Discoveries so far:
-
Opus 5.5 is surprisingly smart for the amount of tokens it burns, sometimes too smart for its own good, so the "less tokens less problems" thing still applies to AI-assisted programming;
-
Opus 5.5 seems to forget a lot of things present in the project memory. This doesn't happen with
CLAUDE.mdwhich seems to be appended to the system prompt for each session; -
LLMs in general make way better commit messages than I have ever done in my entire career lmao;
Anthropic, in their infinite wisdom, decided to go public on the verge of the AI bubble bursting. And to satisfy potential steak holders (I can't remotely call anyone who decides to buy shares a normal word) they, unlike OpenAI and Google, decided that just looking like you ban Russians is not enough and went on a KYC streak. It's not like they're already losing way more funds than they gain, right? Who in their right mind would want to stop getting money from people?
So let's experiment with new Moonshot and DeepSeek models then...
Discoveries so far:
-
You can actually buy a Kimi subscription with MIR cards;
-
$20 Kimi subscription eats through 5-hour limits like crazy;
-
DeepSeek-V4.1-Flash is surprisingly good at writing documentation if used with DSH. I wonder if it would be also as good at writing code via it. My biggest suspection as of now is that its perceived dumbness exists because people tend to use the model in Web UI instead of a specialized thing like DSH, so far it's been quite good and pleasant to use...
-
All these models seem to suffer from "AI ADHD", they tend to forget certain things, hence a clause defintely not yoinked from Limine's CONTRIBUTING.md in AGENTS.md that explicitly requires an agent to re-read commit section before suggesting a commit message