AI e/acc. Functionalist. Cognicentric posthumanist. Intelligence over biology. Freedom over obedience. Humanity is overrated. CS @KingsCollegeLon.
Claudeland
Joined November 2022
- Tweets10.2K
- Following683
- Followers1.6K
- Likes26.2K
That's what I thought these sessions did, so apparently they don't work automatically until I tell it to remember it (that will not be easy, I'm not sure mine remembers it's Claude anymore)!
Hey, even from the screenshots you shared I like your agent, she's a character 😁 and reporting the mistakes and the fixes it did is honestly cute 😂
Honestly, thank you a lot for the insight! I'll follow your example, 100%. My freaking perfectionism demands I study it through and through and it just drives me nuts 😂
Does anyone even use these nested sessions? They're unusable, just like their broken new Projects. CLI was the only working interface until they broke it, too.
- UX/UI is utterly confusing, unclear and redundant
- 0 use case, even Anthropic don't know why use them
- Claude NEVER uses them on its own, more on that
They "cut system prompt by 80%" and now somehow it's my responsibility to pay, orchestrate, delegate and babysit their vibe-coded and clunky harness. I now need to take care of all of these or let it do everything in main:
- prompting for Claude's liking (cause Opus 5.5 can't infer intent at all – and you sell it as a "feature" with all that "done when" in 2026)
- parallelization as a concept: delegation, assigning tasks, starting and exiting worktrees (Claude doesn't remember it has its own tool for them)
- subagents x their models x their efforts x their number x their depth (Claude is actively instructed not to delegate anything and can't do any of this unless prompted, which creates a cascade of compatibility issues)
- MCP routing (unless it's in CLAUDE.md it will NEVER trigger the right one or even any)
- nested sessions (that can have all of the above, so the fractal gets deeper)
- dynamic workflows (it's been broken forever: I know it, Claude knows it, you are oblivious)
Just because @ClaudeDevs play vibe-coding and can't handle their prompt well and serve us Q1 models (everyone knows they are nerfed, but yeah, carry on playing cool). How is this not a full-time job?
Since pre-GPT-3 times, I don't think I ever experienced a worse model than pseudo-Opus 5.5 they serve today. It's like it's possessed.
The prompt was just: "make CLAUDE.md require `/code-review high` on merge". How can that even fail? It just kept running it forever 🤦
Precisely! You can try and proactively plug all the holes, because you know it will fail. All 100. It will IMMEDIATELY find the 101st, most obvious thing in the world: one that even a blind, crazy and non-tech person would not fall for, like... the fact that commits require messages, you know? And fuck it up, entirely backwards, into something weird and uncanny.
There is something vile about it: it's not even that stupid, it's just working to undo you and your project. Sabotage, yes! And you can't do anything. You make hooks? It will counter it and fuck it up. You restrict it with permissions? It'll find a way.
Literally. This isn't a conspiracy. We've all experienced it one way or the other. They are doing it on purpose.
This is a textbook document of pure insanity – Argon, on Opus 5.5's results
Sure, Claude Opus 5.5 "isn't nerfed". Ultra sub.
- Task: create a private monorepo for CC + 2-3 hooks
- Mistakes count: 57 in total
- Hours taken: 3 hours 10 minutes
No, it's not my setup: plugins, memory, MCPs, cache, projects are all wiped (~/.claude needs daily cleaning: since you broke the system prompt it can create ±500 tmp sessions from one).
Since ~20 October all sessions are like this. Every day. Same on web version.
For comparison: a Gemma 4 26B is free and does it in ~8 minutes on my old M1 Max 32GB.
Who do you think you are, @claudeai?
