For as long as it’s not the non-functional requirements, it’ll be fine:
“Nice, it’s only 1,500,000 LoC. Have I mentioned it needs to be a real-time app?”
DaAnalyst
westbaystars
I’m also in the camp of “watch over her shoulder.” And I agree that she gets caught in loops quite often after very long “thinking” sessions.
I usually just write “Continue” and wait more after I interrupt. Now I’m going to ask if anything needs clarification after a long wait.
I had asked her to “Remember: if you find yourself ‘second guessing’ yourself, stop and ask for clarification.” She’s never done that and often says “I’m second guessing myself.”
2
Awlexus
I put the browser on one side of the screen, reading threads like these and on the other side I wait for the next patch of code to be reviewed, intercepting if it makes semantic mistakes or adjust how the code should be written.
I also monitor the token count increase and study it’s relationship to personal feelings of anxiety and dread as the limit seems rapidly approaching. I’ve just hit the rate limit a second time today, so I’m coding by hand for the rest of the day 
2
sodapopcan
I’m interested in these rate limit stories as it’s making me think I’m using it wrong. I’m on the lower tier max plan (it hurts paying even $100USD/month but what can I do… other than hand code, of course) and rarely come close to maxing it out. I feel I must be doing something wrong. I’m assuming it’s that I’ve never had more than two agents going at once, and usually just one but ya. Still curious as to what y’all are doing to keep maxing it out.
1
Awlexus
The secret ingredient is a standard seat in Team. According to me extensive monitoring of token usage while waiting for code snippets, it seems to be 175-200k tokens/5h. And auto reasoning level, which jumps to high on more delicate tasks. It can occasionally consume 30k token in a single iteration.
I’m somewhat convinced this is another case of anthroic ripping off customers, considering the amount of code produced/read, but the code is solid if you correct semantic errors early on 
0
DaAnalyst
Same here. Never hit the limit. The max I saw and it was only once last week when Claude was extremely slow it reached 76% in one 5h session. Never before, never after.
Also, I asked myself the same thing - what am I doing wrong. Then I asked Claude and Grok. The answer was on the lines of: “You’re reviewing a lot, bro. You’re the ideal Anthropic customer.”
Since I have no intention of ceasing to review & correct Claude, it’s probably gonna stay this way.
PS. My weeklies never reached 50%.
1
sodapopcan
Ok that tracks!
Well if you are on the 200/m plan (which I think you said up there), you can probably get away with downgrading!
I was also thinking that it could have to do with that I’m working on a greenfield project and it just isn’t big enough? I did notice it grepping my entire project every time I used a fresh context, so now I point it to the main modules that would be affected by my next prompt and it stopped doing that (I also finally started building up a CLAUDE.md which I only put off as I wasn’t feeling the pain until recently). Also, since when I started my project I wasn’t planning on using LLMs very heavily, I didn’t notice the AGENTS.md Ash generates which has a LOT of stuff around LiveView and I’m using Hologram… so deleting all that has helped a lot… but again, still nowhere close to maxing out, heh.
1
sodapopcan
I still don’t have a good grasp of token usage and don’t often look at it. I guess I should?
Me also use Caveman ulta mode. Eat less token.
Ya, I really hate all of these companies and, since I’m now being honest with myself, this was really the primary reason for not wanting to use AI in the first place. It’s hard to be perfect, though.
0
DaAnalyst
Using Tidewave will save you some grepping (when it’s looking for modules). You don’t need a CLAUDE.md if you’re the only one on the project. You can tell Claude to generate you one out of its saved memories. What I do is always tell it to memorize what I need as a rule at the project level - that’s an analog to writing it into CLAUDE.md. At some point I asked it to generate CLAUDE.md out of what it had learned by then about the project rules. The file it created - perfection. I could never be that concise.
1
sodapopcan
Oh interesting, this is good info, thanks! I literally generated my first CLAUDE.md today, heh. And I’ve never used Tidewave. I still don’t really get MCP, but I’ll look into it now.
Also, apologies for de-railing the topic a bit.
To get back on track, other things I do while waiting: pet my dog. My laptop is his worst enemy.
0