I just used Claude to write a plug-in which changed the bottom of the CLI to say:
( ) Usage: 34% (resets in 2h 47m) Context: 56% [Opus 4.6]
auto mode on (shift+tab to cycle)
That way if my usage starts shooting up, it's very easy to notice (and the color of "( )", which is a circle that I couldn't paste here, changes when it gets high, ensuring I don't miss it).
I coupled that with a hook that watches for usage spiking (basically when I've been talking too long or did something to add a ton of context, so suddenly every turn sends a ton of context back, using up a ton of usage).
Between the two I haven't hit usage caps in weeks.
B) Warning everyone, it slowly eats tokens - If you're like me and on the pro plan, perhaps not a good idea to have its (albeit slow) drain on the tokens.
I set it up and saw the token notifier down the bottom slowly ticking up. I'll try again and report back, perhaps I misread something? I was surprised something like thst (which I assumed was an API call) was trickling token usage.
When was it ticking up? The status line only updates after each turn. If you submitted a turn, tokens were used. The status line itself is a bash script, you can go look at it.
I have something like this persistent in my waybar, with session, weekly, and fable usage, with little color coded percentage bars for each. Super helpful, highly recommend it.
Great idea on the hook alert.
I feel like the advent of LLMs has made a super customizable OS like Linux so, so much better than the closed source alternatives, and so much better than they were before.
Why not juat write the code. Its not that much faster to produce good code with AI. I spend nearly identical amoubts of time reviewing AI code as it would take to write. Im honestly not convinced there is all that much of a productivity boost, maybe 20% faster.
I was mostly joking, but to answer your question, the productivity boost entirely depends on your concern for code quality. Sometimes is it faster to just write the code than to try to explain in the prompt what to do. But if you don't care or don't need to understand what it is creating, you can get a ton of functionality fast.
I genuinely don't even understand how people run out of tokens with the basic subscription. I had Claude write an entire app that has completely replaced Adobe Illustrator for my specific usecases (mostly inking sketches, and embellishing generated SVGs). It does everything exactly the way I want and has much better ergonomics (blender hotkeys ftw) and the whole thing took me like 700k tokens over the course of a couple days.
Do people just let Opus ultra extra max thinking run amok on their whole codebase with a single context? I don't even think I could even design sensible features faster, I'm the limiting factor here, the robot pretty much oneshots everything I ask for.
Sure I get that, but the part where the robot has to read the repo and figure stuff out can be done by a weak model for near-free with no thinking. The delta isn't that big.
I made an "amateur hour" error with Codex. Given Anthropic's recent reliability problems I thought I'd take a little time to try Codex with their $20/mo plan. So I downloaded it and gave it a whirl, not realizing that the default model was gpt-5.6-sol. Well after just an hour or two, I blew through my entire week's quota. Whoops! It would be cool if these harnesses could all graphically display your quota usage on the screen at all times.
I think there is a setting for that! AS a soft shill I'm building www.freepi.ai which is free ai inference in a Pi harness, and I'm going to take that suggestion and add it!
Yep, even with cache hits that like submitting uncached 80k-token prompts over and over.
My Claude.md + framework stuff has it tell me when it recommends /clear before the next prompt (and write itself handoff notes) and I rarely see >100k context size even on large work.
Yeah. I like the way Claude Design handles this: if you come back to a chat after the cache expires, the chat UI shows a message like "start a new chat to save 300k tokens" or whatever. It's pretty nice if you're working in multiple design sessions and are quota-constrained.
23 em-dashes in the span of a single README. I gotta hand it to Anthropic. They seriously managed to find a completely legal way to sell crack to crack addicts using other crack addicts as their unpaid sales force.
Pretty cool tool! Congrats!
Not sure if I'm missing something but this tells you where your tokens went, breakdown per day/tool.
So the "why" is still a question left for the user to answer. Can be something like "Because I was missing a good CLAUDE.md file so it had to explore the whole repo before doing any work" or anything else. Just my take.
for claude, it's response headers contain available usage, you can get it there. My deepseek harness plugin auto stops asking things when I am at 80% to allow inflight prompts to finish.
The only time I've ever managed to burn through a quota that fast (on the cheap plan) was with an open-ended request to check a codebase for any defects or deficiencies. It dispatched five Fable subagents.
I just checked the repo. Nice work
A suggestion:
1. Adding which repo/project was where my most tokens were consumed. I checked the image in the repo README, didn't see that graph.
Take a 2 million line of code repo and ask it to convert from programming language A to B and statically verify every logic flow with an AST for each lang, and then get it to write a test harness that hits 100% code coverage, screenshots every major state and compares screenshots for each state between the old and new. Request it uses an agent team with up to 20 agents. You will use that in a few hours easily.
last 7 days I've used ~4.8B tokens. most of that was during a 2 day span when codex was using one of my harnesses to iterate over a codebase to improve performance of some parts of it - fully automated.
Of course, it's impossible to know for sure what was LLM processed or not, but some of your posts (like this one) have been getting classified that way.
I just used Claude to write a plug-in which changed the bottom of the CLI to say:
That way if my usage starts shooting up, it's very easy to notice (and the color of "( )", which is a circle that I couldn't paste here, changes when it gets high, ensuring I don't miss it).I coupled that with a hook that watches for usage spiking (basically when I've been talking too long or did something to add a ton of context, so suddenly every turn sends a ton of context back, using up a ton of usage).
Between the two I haven't hit usage caps in weeks.
This is actually a built-in Claude Code feature: `/statusline` will ask how you want it to look and set it up for you.
A) Oh nice! that's cool
B) Warning everyone, it slowly eats tokens - If you're like me and on the pro plan, perhaps not a good idea to have its (albeit slow) drain on the tokens.
I didn't think statusline ate tokens, is this true?
No, when you run it Claude just writes a bash script used to generate the status line. Not sure what they're talking about.
I set it up and saw the token notifier down the bottom slowly ticking up. I'll try again and report back, perhaps I misread something? I was surprised something like thst (which I assumed was an API call) was trickling token usage.
When was it ticking up? The status line only updates after each turn. If you submitted a turn, tokens were used. The status line itself is a bash script, you can go look at it.
I have something like this persistent in my waybar, with session, weekly, and fable usage, with little color coded percentage bars for each. Super helpful, highly recommend it.
Great idea on the hook alert.
I feel like the advent of LLMs has made a super customizable OS like Linux so, so much better than the closed source alternatives, and so much better than they were before.
Wanna share?
I didn't know it was still possible to develop software after your ai quota runs out.
You wait until the month resets.
You’re joking but from the readme
> Is it safe to start a big refactor now, or should I wait for my window to clear?
Why not juat write the code. Its not that much faster to produce good code with AI. I spend nearly identical amoubts of time reviewing AI code as it would take to write. Im honestly not convinced there is all that much of a productivity boost, maybe 20% faster.
I was mostly joking, but to answer your question, the productivity boost entirely depends on your concern for code quality. Sometimes is it faster to just write the code than to try to explain in the prompt what to do. But if you don't care or don't need to understand what it is creating, you can get a ton of functionality fast.
Is this the new 10x engineer hammock meme? No one is vesting anymore they just burned their quarterly token quote in 10 days.
Darmok and Jalad, after their quota ran out.
Shaka, when they discovered open models!
omg. so awesome
Shaka, when the bubble burst!
Zinda, his face black, his eyes red…
Local AI has no quote, except cold hard WATTs
I genuinely don't even understand how people run out of tokens with the basic subscription. I had Claude write an entire app that has completely replaced Adobe Illustrator for my specific usecases (mostly inking sketches, and embellishing generated SVGs). It does everything exactly the way I want and has much better ergonomics (blender hotkeys ftw) and the whole thing took me like 700k tokens over the course of a couple days.
Do people just let Opus ultra extra max thinking run amok on their whole codebase with a single context? I don't even think I could even design sensible features faster, I'm the limiting factor here, the robot pretty much oneshots everything I ask for.
Token usage is generally lower in greenfield projects.
Sure I get that, but the part where the robot has to read the repo and figure stuff out can be done by a weak model for near-free with no thinking. The delta isn't that big.
I made an "amateur hour" error with Codex. Given Anthropic's recent reliability problems I thought I'd take a little time to try Codex with their $20/mo plan. So I downloaded it and gave it a whirl, not realizing that the default model was gpt-5.6-sol. Well after just an hour or two, I blew through my entire week's quota. Whoops! It would be cool if these harnesses could all graphically display your quota usage on the screen at all times.
I think there is a setting for that! AS a soft shill I'm building www.freepi.ai which is free ai inference in a Pi harness, and I'm going to take that suggestion and add it!
You can enable showing the quota usage (and much more) in codex via /statusline
In 95% of cases it is because you had a large context for which the cache expired.
The amount of people I see running around at 800k context and wondering why they are burning through tokens is always surprising.
Yep, even with cache hits that like submitting uncached 80k-token prompts over and over.
My Claude.md + framework stuff has it tell me when it recommends /clear before the next prompt (and write itself handoff notes) and I rarely see >100k context size even on large work.
Yeah. I like the way Claude Design handles this: if you come back to a chat after the cache expires, the chat UI shows a message like "start a new chat to save 300k tokens" or whatever. It's pretty nice if you're working in multiple design sessions and are quota-constrained.
Surprised they haven't brought the same UX to CC.
I have that in CC when I resume a session.
Suggests either starting from a summary, a new conversation, or keep going.
23 em-dashes in the span of a single README. I gotta hand it to Anthropic. They seriously managed to find a completely legal way to sell crack to crack addicts using other crack addicts as their unpaid sales force.
Having a good statusline is key! I use https://github.com/lpsgverrilla/lps-statusline
Pretty cool tool! Congrats! Not sure if I'm missing something but this tells you where your tokens went, breakdown per day/tool. So the "why" is still a question left for the user to answer. Can be something like "Because I was missing a good CLAUDE.md file so it had to explore the whole repo before doing any work" or anything else. Just my take.
I created a custom status line to see/track everything I need: https://github.com/dsebastien/claude-epic-status-line
That avoids surprises
Full story if anyone is interested - https://www.kelviq.com/blog/claude-code-usage-limits-where-t...
I care more about your experience. I'd read the blog post of the story behind this.
> npx skills add kelviq/tare -g -y --copy --agent claude-code
tare is a Python project. Why distribute it with npm rather than PyPI? Weird
You can't ask Claude if your quota ran out. You have to wait for the reset...
for claude, it's response headers contain available usage, you can get it there. My deepseek harness plugin auto stops asking things when I am at 80% to allow inflight prompts to finish.
The only time I've ever managed to burn through a quota that fast (on the cheap plan) was with an open-ended request to check a codebase for any defects or deficiencies. It dispatched five Fable subagents.
I just checked the repo. Nice work A suggestion: 1. Adding which repo/project was where my most tokens were consumed. I checked the image in the repo README, didn't see that graph.
I've been recording every statusline output for months now so I can easily get answers to questions like this whenever I wish.
I’ve been using headroom to save on token usage and it’s pretty effective.
this kind of stats feature should be shipped by default with every harness imvho
How on Earth does one use 1.1B tokens in a week?
With cache reads and writes and a couple of long running session without /clear or /compact it can get there rather fast
Take a 2 million line of code repo and ask it to convert from programming language A to B and statically verify every logic flow with an AST for each lang, and then get it to write a test harness that hits 100% code coverage, screenshots every major state and compares screenshots for each state between the old and new. Request it uses an agent team with up to 20 agents. You will use that in a few hours easily.
Can do that in a day...
By doing ALL your work in a single chat. The context window explodes.
I typically do one context per feature.
last 7 days I've used ~4.8B tokens. most of that was during a 2 day span when codex was using one of my harnesses to iterate over a codebase to improve performance of some parts of it - fully automated.
[flagged]
Can you please not post AI-generated or AI-edited comments to HN? It's not allowed here - see https://news.ycombinator.com/newsguidelines.html#generated and https://news.ycombinator.com/item?id=47340079.
Of course, it's impossible to know for sure what was LLM processed or not, but some of your posts (like this one) have been getting classified that way.
only one of my post/reply is flagged other than this one tho?? Also yea i admit i used AI here , sorry for that