Hacker News Reader

Comment by steve_adams_86 | Hacker News Reader

I'm well within the 95%. I might lack an imagination here, but... What are you guys doing that you hit or exceed limits so easily, and if you do... Why does it matter? Sometimes I'd like to continue exploring ideas with Claude, but once I hit the limit I make a mental note of the time it'll come back and carry on planning and speccing without it. That's fine. If anything, some time away from the slot machine often helps with ensuring I stay on course.

View on HN

Opus + extended thinking + deep research = 3-5 messages/reports per five hour limit. That’s the fastest way I’ve found to blow through the Pro plan.

Some stuff I’ve used it for in the last day: figuring out what a family member needs for FAFSA as a nontraditional student, help identify and authenticate some rare first editions and incunabula for a museum collection I volunteer at, find a list of social events in my area (based on my preferences) that are coming up in the next week (Chatgpt Agent works surprisingly well for this too), adapting Directus and Medusa to my project’s existing schema and writing up everything I need to migrate, and so on.

Deep research really hits the Claude limits hard and that’s the best way to avoid hallucinations when asking an important question or making it write complex code. I just switch from Claude to ChatGPT/Gemini until the limits reset but Claude’s deep research seems to handily beat Gemini (and OpenAI isnt even in the running). DR queries take much longer (5-10 min in average) but have much more in depth and accurate answers.

I hadn't considered that. I'm using it almost exclusively to validate logic, kind of like a fuzzer in nature ("What if we need to do this with this logic/someone tries to do that/what am I missing/etc"), or to fill in specifications ("what feature would compliment this/what could be trimmed to achieve MVP more easily/does this spec appear to be missing anything according to this set of requirements"), which requires a lot of review, and using more expensive models like Opus doesn't appear to provide meaningfully better results. After prompting it, I typically have a lot to think about and the terminal goes quiet, or I prompt it on a related matter that will similarly require my eyes and brain for long enough that I won't be able to limit out.

I can see how work involving larger contexts and deeper consideration would lead to exhausting limits a lot faster though, even if you aren't using it like a slot machine.

"find a list of social events in my area"

Isn't this something you can do with a simple Google search? Or Perplexity?

No need to shove by far the most expensive LLM (Claude Opus 4) at it.

Not for the Los Angeles metro area. There isn’t a single calendar or event aggregator that covers the entire area and with an LLM I can give it complex schedules (i.e. a dump of my calendar for that week) and preferences to filter the list of events for the stuff I like, including vague stuff like “I like country music in the style of ‘Take Me Home, Country Roads’ but not modern country radio”.

That sounds like a startup for me.

Collate all the LA Metro area events from different sources and whip up an app or web site where people can filter them and subscribe to the events in Google Calendar or in .ical format.

You can even have Claude vibe code it for you :)

Killing a fly with a cannonball.

Doesn’t really matter when the marginal cost of the cannonball is effectively zero - I’m already paying the monthly subscription.

Then not using the canonball is just a waste of time, which is a heck of a lot more valuable than some purist aversion to using LLMs to save time and effort.

One could argue this is like paying a subscription for gasoline and saying you better use it up or it's a waste. There's an externality at play.

I know LLMs aren't as much of an environmental scourge as people sometimes make them out to be, but if they're used eagerly and aggressively, their impacts certainly have a capability of scaling in concerning ways.

Gosh I so despise this new normal. Just when I thought I could fight bloat and unnecessary tech in my own tiny corner of the world, only for a few to ruin it with ridiculous LLM (ab)use.

Where is the fly swatter at?

When you say not even in the running, is that including Deep Research on o3-pro?

I haven't tried o3-pro, but my fundamental problem with ChatGPT Deep Research is that it only searches for a few dozen sources, whereas Claude and Gemini regularly use 400+ sources.

I agree. I'm on the base plan, yet to hit any limits. The bottleneck is my ability to review the code it writes, and to write prompts detailed enough for the output to be useful to me.

I assume that the people hitting limits are just letting it cycle, but doesn't that just create garbage if you don't keep it on a tight leash? It's very eager but not always intelligent.

Switching to Opus is an eye-opening experience. You hit limits often, and need to get creative to avoid burning through limits, but the difference is seriously impressive. You'll waste a lot less time with dead ends and bad code.

The issue (with Sonnet, I'm not using Opus), is not always that the code is bad per se, but merely that it doesn't solve the problem in the way I expected.

I have two problems with that. Firstly, I want my code to be written a particular way, so if it's doing something out of left field then I have to reject it on stylistic grounds. Secondly, if its solution is too far from my expectation, I have to put more work into review to check that its solution is actually correct.

So I give it a "where, what, how" prompt. For example, "In file X add feature Y by writing a function with signature f(x: t), and changing Z to do W..."

It's very good at following directions, if you give it the how hints to narrow the solution space.

I think this is it. They use it like a slot machine, and when something isn't quite what they wanted, they provide broad instructions to refine and do better. Progress is slow and questionable, but anticipation and (sometimes) reward is increased.

The issue could be, in part, that a lot of users don't care to be efficient with token usage and maintaining condensed, efficient, focused contexts to work with.

i wonder how many are negligent vs ignorant. negligence would be senior engineers that could scope and work with context properly but are lazy or don't care. ignorance would be vibe coders that genuinely can't express anything beyond plain english and superficial descriptions of issues and changes.

I've found that asking for deep research consumes my quota quite fast, so If I run 2 or 3 and normal use I hit the limit and have to wait to reset

Me too. I've also found that even when trying to restrict models meant for these tasks, they tend to go on tangents and waste tremendous amounts of tokens without providing meaningfully better outputs. I'm not yet sold on these models for anything outside of fuzzy tasks like "does this logic seem sound?". They tend to be good at that (though they often want to elaborate excessively or propose solutions excessively).

One way I've seen personally is that folks are using tools that drive many Claude Code sessions at once via something like git-worktree as a way of multitasking in a single codebase. Even with garden-variety model use, these folks are hitting the existing 5-hourly rate limits routinely.

I use this approach because I like to work on features or logical components in isolation and then bring them together. I still can't limit out most of the time because I need to actually look at the outputs and think about what I'm building. At the moment I have 3 directories in my work tree. Sometimes I prompt in more than one at a time, especially at interfacing code, but that could mean 30–90 minutes of reviewing and implementing things in each directory. Over a work day I apparently send an average of ~40 messages according to `claude --resume`

Im not a formula 1 driver but why do they have those big padel things on the back? looks dumbo IMHO I just dont get it

I respectfully consider this analogy void, but welcome an explanation of why I'm wrong.

I haven't yet seen anyone doing anything remarkable with their extensive use of Claude. Without frequent human intervention, all of it looks like rapid regression to the mean, or worse.

Reply on news.ycombinator.com