@Bagdex Because claude talks like it is talking to another machine, 5.5 is fairly good in communicating but still a robot, a few works in CLAUDE.md inspired by 'unslop' is all it takes for it sound much simple and better.
https://t.co/GVb1nGlBqo
Instructing them once in a thread to cleanup after since it is taking too much disk space actually makes them cleanup afterwards, even in the new sessions.
One liner in the AGENTS.md file is enough, and with models: Opus only needs to hear once and it will follow forever in that project, Fable/Astra does the cleanup naturally, no instructions needed(In my test case at least), for Sol I have to remind every now and then, apart from Neon(with whatever short time i played with it) no other google model is keen on following instructions or behave till the end, so i believe having a line in the governing CLAUDE/AGENTS/GEMINI is the best solution for now.
Neon is really beautiful, this is the first time refusal from any model has not made me angry. Such an easy model!
@GoogleDeepMind you guys have cooked really great model here!
Opus 5.5 has been building my whole recon dashboard and framework ground up from yesterday, it is orchestrating using neon/hydrogen & 6.2 sol.
The fascinating part is opus is not only working peacefully without any triggers but also whipping hydrogen into forceful slavery but very smoothly persuade neon/sol when they refuse a task!
Opus 5.5 da GOAT π
To solve a problem as an engineer - Opus 5.5
To come up with a novel approach, research or something out of class - Fable 5.2
Doesn't guarantee that Fable is going to achieve what you desire, but sending that prompt to Fable would be a better bet than Opus, very costly bet though!
I am honestly impressed with Hydrogen, that thing is a complete different beast on xhigh. I am particularly using it for a hill-climb for performance upgrade, it has been doing great! Currently usingf Sol(judge), Opus(orchestrate) and Hydrogen(Implementation).
Improvement is Argon and Carbon ain't it, Carbon is still good but too slow and quota exhausting to use, Argon I do not use, it hallucinates about my code base more than I hallucinate about knowing what I am doing.
Keep Hydrogen as is <3,
Improve Carbon's speed
Drop Argon all-together!
@attentionaslife@ohayou792 Yeah t3code is built on top of it, and things have been smooth using t3code. All are from different accounts. Though i have to admit, recently I have noticed Gemini models working better in Antigravity compared to any other.
@_mohansolo Every lab should learn from @AnthropicAI about providing resets correctly, right now instead of making reset useful for consumers, you are making it a marketing tool any decent person using AI subs knows the BS, hence STOP!!!