jusasiiv

I wonder why having the usage limits for subscriptions was not enough. I see that Opus 4.7 now takes 7x usage so admittedly the limits would go over very fast. In the end this will end up costing more for users but I understand that they cannot operate on loss.

Maybe the root issue is that Anthropic is operating at a loss on their subscriptions and everyone else who tries to add Claude on top of their service will have a though time turning a profit.

Very good tips for anyone working with bigger tasks. Since memory management for LLMs is still kind of an unsolved problem (yes we have RAGs etc. but they are just a crutch right now) to get the best results you still need to know how to manage context yourself.

One thing I like to do for larger tasks is to first create a separate document and then prompt using the document, usually new session for each task. AI is good at slicing the task for itself and as a human in the loop its easier to understand the steps its performing and what is happening.

There are probably a million ready made solutions already but manual labour for context management still seems to get the best results for me personally.

I disagree. UX could be improved but for an alpha version I found out it was easy to use. The main feature "Guess" should be probably more prevalent though. Also for me it did not seem sluggish but that might depend on a lot of factors.

What kind of token usage you have with this setup? Also why both ChatGPT and Claude?

Hah pretty funny. It keeps saying the same things over and over again. Maybe improving the prompt or just a bit bigger context could help even with the smallets models?

By this point if you think that they are not tracking you I dont know what to tell you. Even the APIs, although they say they do not use it for training the models if you opt out, there is no telling what else they do with the data. OpenClaw showed us that the general user does not care about privacy or security as long as they can use an agent to make a note in their calendar.

Imagine how much money the tobacco industry spent on lobbying against rules and regulations (now they spend to money to sell vapes). Alcohol industry has been doing exactly the same... Smoking was cool until the health risks were made clear, alcohol is still cool even though the health risks are basically as high. Its just that drinking has been so ingrained in our culture that it will be a slow ship to turn and the industry will 100 % do everything to prevent that as much as possible. But I do have hope, some statistics say that drinking is not as popular with young people anymore.

This is from a software development view, but I keep seeing this problem where the lead is promoted from a well performing individual. Leading a team is greatly different than shipping code so that drives the new leads to focus on learning how to manage that part of the process, usually leaving less time to work on the code. After a while there start to be a disconnect between the lead and the code base causing friction on performing as a developer. So why do we keep promoting the great coders to team leads where their output as a developer drops dramatically?

Lots of questions already about the project so Ill ask about your vibe coding experience. After the project has grown have you found it harder to upkeep the project with AI? Also did you use any structuring while working with Claude/Codex like first planning the tasks on separate documents, or did you just work with prompts and the context that were generated on the sessions?

Seems interesting. Ill give it a try on my agent, memory is definitely an ongoing issue. How long have you been running this in a continuous state? Also have you tried other LLM's and seen a difference on how well they can use it?

You state that this wraps Zeroclaw but would there be any restrictions on running this with something like NanoClaw, PicoClaw or any of the million other OpenClaw derivatives?