> Then a brick hits you in the face when it dawns on you that all of our tools are dumping crazy amounts of non-relevant context into stdout thereby polluting your context windows.
I've found that letting the agent write its own optimized script for dealing with some things can really help with this. Claude is now forbidden from using `gradlew` directly, and can only use a helper script we made. It clears, recompiles, publishes locally, tests, ... all with a few extra flags. And when a test fails, the stack trace is printed.
Before this, Claude had to do A TON of different calls, all messing up the context. And when tests failed, it started to read gradle's generated HTML/XML files, which damaged the context immensely, since they contain a bunch of inline javascript.
And I've also been implementing this "LLM=true"-like behaviour in most of my applications. When an LLM is using it, logging is less verbose, it's also deduplicated so it doesn't show the same line a hundred times, ...
> He sees something goes wrong, but now he cut off the stacktraces by using tail, so he tries again using a bigger tail. Not satisfied with what he sees HE TRIES AGAIN with a bigger tail, and … you see the problem. It’s like a dog chasing its own tail.
I've had the same issue. Claude was running the 5+ minute test suite MULTIPLE TIMES in succession, just with a different `| grep something` tacked at the end.
Now, the scripts I made always logs the entire (simplified) output, and just prints the path to the temporary file. This works so much better.
> Claude is now forbidden from using `gradlew` directly, and can only use a helper script we made. It clears, recompiles, publishes locally, tests, ... all with a few extra flags. And when a test fails, the stack trace is printed.
I think my question at this point is what about this is specific to LLMs. Humans should not be forced to wade through reams of garbage output either.
Humans have the ability to ignore and generally not remember things after a short scan, prioritize what's actually important etc. But to an LLM a token is a token.
There's attempts at effectively doing something similar with analysis passes of the context - kinda what things like auto-compaction is doing - but I'm sure anyone who has used the current generation of those tools will tell you they're very much imperfect.
That may help with tokens being "ignored" while still being in the context window, but not context window size costs and limitations in the first place.
In my experience, it's the old time-invested vs time-saved trade off. If you're not looking at these reams of output often enough, the incentive to figure out all the flags and configs for verbosity to write these script is lower: https://xkcd.com/1205/
And because these issues are often sporadic, doing all this would be an unwanted sidequest, so humans grit their teeth and wade through the garbage manually each time.
With LLMs, the cost is effectively 0 compared to a human, so it doesn't matter. Have them write the script. In fact, because it benefits the LLM by reducing context pollution, which increases their accuracy, such measures should be actively identified and put in place.
The way I've solved this issue with a long running build script is to have a logging scripts which redirects all outputs into a file and can be included with
```
# Redirect all output to a log file (re-execs script with redirection)
source "$(dirname "$0")/common/logging.sh"
```
at the start of a script.
Then when the script runs the output is put into a file, and the LLM can search that. Works like a charm.
This has been my exact experience with agents using gradle and it’s beyond frustrating to watch. I’ve been meaning to set up my own low-noise wrapper script.
This post just inspired me to tackle this once and for all today.
If you get desperate, I've given my agent a custom $PATH that replaces the forbidden tools with shims that either call the correct tool, or at least tell it what to do differently.
~/agent-shims/mvn:
#!/bin/bash
echo "Usage of 'mvn' is forbidden. Use build.sh or run-tests.sh"
That way it is prevented from using the wrong tools, and can self-correct when it tries.
> Then a brick hits you in the face when it dawns on you that all of our tools are dumping crazy amounts of non-relevant context into stdout thereby polluting your context windows.
Not just context windows. Lots of that crap is completely useless for humans too. It's not a rare occurrence for warnings to be hidden in so much irrelevant output that they're there for years before someone notices.
The old unix philosophy of "print nothing on success" looks crazy until you start trying to build pipes and shell scripts that use multiple tools internally. Also very quickly makes it clear why stdout and stderr are separate
Also becomes rapidly apparent that most modern tooling takes an extremely liberal view of logging levels. The fact that you’ve successfully processed a file is not INFO, that’s DEBUG.
"Finished conversion of xyz.mp3 to xyz.ogg" is valuable progress information to a regular user, not just to developers, so it belongs in INFO, not DEBUG.
I suppose this is subjective, but I disagree. If I want to know the status of each item, I’d pass -v to the command. A simple summary at the end is sufficient; if I pass -q, I expect it to print nothing, only issuing a return code.
BATCH=yes (default is no)
--batch (default is --no-batch)
for the unusual case when you do want the `route print` on a BGP router to actually dump 8 gigabytes of text throughout next 2 minutes. Maybe it's fine if a default output for anything generously applies summarization, such as "X, Y, Z ...and 9 thousand+ similar entries".
Having two separate command names (one for human/llm, one for batch) sucks.
Having `-h` for human, like ls or df do, sucks slightly less, but it is still a backward-compatibility hack which leads to `alias` proliferation and makes human lifes worse.
Something related to this article, but not related to AI:
As someone who loves coding pet projects but is not a software engineer by profession, I find the paradigm of maintaining all these config files and environment variables exhausting, and there seem to be more and more of them for any non-trivial projects.
Not only do I find it hard to remember which is which or to locate any specific setting, their mechanisms often feel mysterious too: I often have to manually test them to see if they actually work or how exactly. This is not the case for actual code, where I can understand the logic just by reading it, since it has a clearer flow.
And I just can’t make myself blindly copy other people's config/env files without knowing what each switch is doing. This makes building projects, and especially copying or imitating other people's projects, a frustrating experience.
How do you deal with this better, my fellow professionals?
Software folks love over-engineering things. If you look at the web coding craze of a few years ago, people started piling up tooling on top of tooling (frameworks, build pipelines, linting, generators etc.) for something that could also be zero-config, and just a handful of files for simple projects.
I guess this happens when you're too deep in a topic and forget that eventually the overhead of maintaining the tooling outweights the benefits. It's a curse of our profession. We build and automate things, so we naturally want to build and automate tooling for doing the things we do.
I don’t think those web tooling piles are over-engineered per se, they address huge challenges at Google and Facebook, but the profession is way too driven by hype and fashion and the result is a lot of cargo culting of stuff from Big Dogs unquestioningly. Wrong tooling for the job creates that bubble of over complicated app development.
Inventing GraphQL and React and making your own PHP compiler are absolutely insane and obviously wrong decisions — for everyone who isn’t Facebook. With Facebook revenue and Facebooks army of resume obsessed PHP monkeys they strike me as elegant technological solutions to otherwise intractable organizational issues. Insane, but highly profitable and fast moving. Outside of that context using React should be addressing clear pain points, not a dogmatic default.
We’re seeing some active pushback on it now online, but so much damage has been done. Embracing progressive complexity of web apps/sites should leave the majority as barebones with minimal if any JavaScript.
Facebook solutions for Facebook problems. Most of us can be deeply happy our 99 problems don’t include theirs, and live a simpler easier life.
Not sure why you lumped React in there. Hack is loopy, and GraphQL was overhyped but conditionally useful, but React was legitimately useful and a real improvement over other ways of doing things at the time. Compare React to contemporary stuff like jQuery, Backbone, Knockout, Angular 1.x, etc.
I agree with you very much, if what you are building actually benefits from that much client side interactivity. I think the counterpoint is that most products could be server rendered html templates with a tiny amount of plain js rather than complex frontend applications.
First of all, I read the documentation for the tools I'm trying to configure.
I know this is very 20th century, but it helps a lot to understand how everything fits together and to remember what each tool does in a complex stack.
Documentation is not always perfect or complete, but it makes it much easier to find parameters in config files and know which ones to tweak.
And when the documentation falls short, the old adage applies: "Use the source, Luke."
> As someone who loves coding pet projects but is not a software engineer by profession, I find the paradigm of maintaining all these config files and environment variables exhausting
Then don’t.
> How do you deal with this better, my fellow professionals?
By not doing it.
Look, it’s your project. Why are you frustrating yourself? What you do is you set up your environment, your configuration, what you need/understand/prefer and that’s it. You’ll find out what those are as you go along. If you need, document each line as you add it. Don’t complicate it.
Don't fall for the "JS ecosystem" trap and use sane tools. If a floobergloob requires you to add a floobergloob.config.js to your project root that's a very good indicator floobergloob is not worth your time.
The only boilerplate files you need in a JS repo root are gitignore, package.json, package-lock.json and optionally tsconfig if you're using TS.
A node.js project shouldn't require a build step, and most websites can get away with a single build.js that calls your bundler (esbuild) and copies some static files dist/.
You start with the cleanest most minimal config you can get away with, but over the years you keep adding small additions and tweaks until it becomes a massive behemoth that only you will ever understand the reasoning behind.
Part of doing it well is adding comments as you add options. When I used vim, every line or block in the config had an accompanying comment explaining what it did, except if the config’s name was so obvious that a comment would just repeat it.
That's a good call. It's a big problem for JSON configs given pure JSON's strict no-comments policy. I like tools that let you use .js or better yet .ts files for config.
I like this idea a lot, and pushed for json5 at a previous job, but I think there are a few snags:
- it's weird and unfamiliar, most people prefer plain JSON
- there are too many competing standards to choose from
- most existing tools just use plain JSON (sometimes with support for non-standard features, like tsconfig allowing trailing commas, but usually poorly documented and unreliable)
Much easier just to make the leap to .ts files, which are ergonomically better in almost every way anyway.
A lot of json parsers will permit comments even though it isn't meant to be valid. Worth trying it, see if a comment breaks the config, and if not then use comments and don't worry about it.
I think the concept has value, but I think targeting today's LLMs like this is short sighted.
It's making what is likely to be a permanent change to fix a temporary problem.
I think the thing that would have value in the long term is an option to be concise, accurate, and unambiguous.
This isn't something that should be considered to be only for LLMs. Sometimes humans want readability to understand something quickly adding context helps a great deal here, but sometimes accuracy and unambiguity are paramount (like when doing an audit) if dealing with a batch of similar things, the same repeated context adds nothing and limits how much you can see at once.
So there can be a benefit when a human can request output like this for them to read directly. On top of this is the broad range of of output processing tools that we have (some people still awk).
So yes, this is needed, but LLMs will probably not need this in a few years. The other uses will remain
Rather than an LLM=true, this is better handled with standardizing quiet/verbose settings, as this is a question of verbosity, where an LLM is one instance where you usually want it to be quieter, but not always.
Secondly, a helper to capture output and cache it, and frankly a tool or just options to the regular shell/bash tools to cache output and allow filtered retrieval of the cached output, as more so than context and tokens the frustration I have with the patterns shown is that often the agent will re-execute time-consuming tasks to retrieve a different set of lines from the output.
A lot of the time it might even be best to run the tool with verbose output, but it'd be nice if tools had a more uniform way of giving output that was easier to systematically filter to essentials on first run (while caching the rest).
Yes! After seeing a lot of discussions like this, I came up with a rule of thumb:
Any special accommodations you make for LLMs are either a) also good for humans, or b) more trouble than they're worth.
It would be nice for both LLMs and humans to have a tool that hides verbose tool output, but still lets you go back and inspect it if there's a problem. Although in practice as a human I just minimise the terminal and ignore the spam until it finishes. Maybe LLMs just need their own equivalent of that, rather than always being hooked up directly to the stdout firehose.
Yes, what's preventing the LLM from running myCommand > /tmp/out_someHash.txt ; tail out_someHash.txt and then greping or tailing around /tmp/out_someHash.txt on failure?
There isn't really anything other than training, but they generally don't. You probably can get them to do that with some extra instructions, but part of the problem - at least with Claude - is that it's really trigger-happy about re-running the commands if it doesn't get the results it likes, assuming the results reflects stale results. Even with very expensive (in time) scripts I often see it start a run, pipe it to a file, put it in the background, then loop on sleep statements, occasionally get "frustrated" and check, only to throw the results away 30 seconds after they are done because it's made an unrelated change.
A lot of the time this behaviour is probably right. But it's annoyingly hard to steer it to handle this correctly. I've had it do this even with make targets where the makefile itself makes clear the dependencies means it could trust the cached (in a file) results if it just runs make <target>. Instead I regularly find it reading the Makefile and running the commands manually to work around the dependency management.
It feels wild to have to keep reminding people, but AI changes very little. Tools have always had a variety of output, and ways to control this, and bad tools output a lot by default, whilst good tools hide it behind version of "-v" or easy greps. Don't add a --LLM or whatever, do add cleaner and consistent verbosity controls.
The UNIX philosophy of tools that handle text streams, staying "quiet" unless something goes wrong, doing one thing well, etc. are all still so well suited to the modern age of AI coding agents.
Also an acceptable solution - create a "runner" subagent on a cheap model, that's tasked with running a command and relaying the important parts to the main agent.
Why can't the agent harness dynamically decide whether outputs should be put into the context or not? It could check with an LLM to determine if the verbatim output seems important, and if not, store the full output locally but replace it in the prompt with a brief summary and unique ID. Then make a tool available so the full output can be retrieved later if necessary. That's roughly how humans do it, you scroll through your terminal and make quick decisions about what parts you can ignore, and then maybe come back later when you realize "oh I should probably read that whole stack trace".
It wouldn't even need to send the full output to make a decision, it could just send "npm run build output 500 lines and succeeded, do we need to read the output?" and based on the rest of the conversation the LLM can respond yes or no.
Sort of, but you also want to keep the sub-agent context small for as long as possible, and if you're paying per token there's no reason to be sending thousands of tokens that are probably useless.
We’ve got a long way to go in optimising our environments for these models. Our perception of a terminal is much closer to feeding a video into Gemini than reading a textbook of logs. But we don’t make that ax affordance at the moment.
I wrote a small game for my dev team to experience what it’s like interacting through these painful interfaces over the summer www.youareanagent.app
Jump to the agentic coding level or the mcp level to experience true frustration (call it empathy). I also wrote up a lot more thinking here www.robkopel.me/field-notes/ax-agent-experience/
So frequently beginners in linux command lines complain about the irregularity or redundance in command line tool conventions (sometimes actual command parameters -h --help or /h ? other times: man vs info; etc...)
When the first transformers that did more than poetry or rough translation appeared everybody noticed their flaws, but I observed that a dumb enough (or smart enough to be dangerous?) LLM could be useful in regularizing parameter conventions. I would ask an LLM how to do this or that, and it would "helpfully" generate non-functional command invocations that otherwise appeared very 'conformant' to the point that sometimes my opinion was that -even though the invocation was wrong given the current calling convention for a specific tool- it would actually improve the tool if it accepted that human-machine ABI or calling convention.
Now let us take the example of man vs info, I am not proposing to let AI decide we should all settle on man; nor do I propose to let AI decide we should all use info instead, but with AI we could have the documentation made whole in the missing half, and then it's up to the user if they prefer man or info to fetch the documentation of that tool.
Similarily for calling conventions, we could ask LLM's to assemble parameter styles and analyze command calling conventions / parameters and then find one or more canonical ways to communicate this, perhaps consulting an environment variable to figure out what calling convention the user declares to use.
> Indeed hallucinated cases are "better law." Drawing on Ronald Dworkin's theory of law as integrity, which posits that ideal legal decisions must "fit" existing precedents while advancing principled justice, this article argues that these hallucinations represent emergent normative ideals. AI models, trained on vast corpora of real case law, synthesize patterns to produce rulings that optimally align with underlying legal principles, filling gaps in the doctrinal landscape. Rather than errors, they embody the "cases that should exist," reflecting a Hercules-like judge's holistic interpretation.
Seems naive. You can get an LLM to agree with almost anything if you say the right things to it, and it will hallucinate citations to back you up without skipping a beat. You can probably get it to hallucinate case law to legalize murder on Mondays.
It has long been a pet peeve of mine that the *nix world has no standard reliable convention for how to interrogate a program for it's available flags. Instead there are at least a dozen ways it can be done and you can't rely on any one of them.
I didn't say it was, but I simply don't care what problems any other kind of system has because they aren't my problems. Last time I had windows on any of my computers it was windows 98.
I've been thinking about using an OpenAPI schema to describe cli tools. It would probably need extensions to describe stdin/stout and other things that don't happen in REST.
Ngl.. I can see the merit and simultaneously recoil in horror as I am starting to understand what linux greybeards hate about windofication of linux ( and now proposed llm-ification of it :D).
> The environment wins (less tokens burned = less energy consumed)
This is understandable logic, but at a systemic level it's not how things always go. Increasing efficiency can lead to increased consumption overall. You might save 50% in energy for your workload, but maybe now you can run it 3 times as much, or maybe 3 times more people will use it, because it's cheaper. The result might be a 50% INCREASE in energy consumed.
This is the standing reason that is always given for why we must all sit in freeway traffic clogs, and I think it's B.S., because it assumes that there are viable alternatives available in near-medium term, but that isn't always the case. The alternative to freeways that are supposed to compensate is a joint combination of denser housing and mass transit, which in California, is not happening at all...zoning laws and the slow pace of building mass transit due to regulation slow-down and the need to service urban sprawl, prevent that solution from relieving traffic pressure. Don't speak of busses, because taking two hours to get to work is not better than one hour. So..the freeways stay the same number of lanes and my commute time continues to grow, and I am tired of hearing it is for the best.
So yes, lower LLM costs would probably lead even more LLM usage and greater energy expenditures, but then again, so does having a moving economy, and all that comes with that.
Yeah, probably. I wonder where speed-running fixing all the low-hanging fruit for AI-related efficiency improvements will leave us? It still seems worth doing. Maybe combined with a carbon tax.
Surprisingly often people refuse to document their architecture or workflow for new hires. However, when it's for an LLM some of these same people are suddenly willing to spend a lot of time and effort detailing architecture, process, workflows.
I've seen projects with an empty README and a very extensive CLAUDE.md (or equivalent).
I like the gist of this, however LLM may not be the best name for this: what if a new tech (e.g., SLM) takes over? AGENT may be a more faithful name until something better is standardized.
Looks like the blog could use a HN=True. Hope the author won't get banned...
> Error: API rate limit exceeded for app ID 7cc6c241b6e6762bf384. If you reach out to GitHub Support for help, please include the request ID E9FC:7BEBA:6CDB3B4:6485458:699EE247 and timestamp 2026-02-25 11:51:35 UTC. For more on scraping GitHub and how it may affect your rights, please review our Terms of Service (https://docs.github.com/en/site-policy/github-terms/github-t...).
I would use this as a human. That npm output is crazy. Maybe a better variable would be "CONCISE=1". For LLMs, there are a few easier solutions, like outputing in a file (and then tail)., or running a subagent
On a lot of linux distros there is the `moreutils` package, which contains a command called `chronic`. Originally intended to be used in crontabs, it executes a command and only outputs its output if it fails.
I think this could find another use case here.
This is great, I like this. Wrote a 'chronic-file' variant that just dumps everything into a tmpfile and outputs the filepath for the agent in case of error and otherwise nothing
Many unix tools already print less logging when used im a script, ie. non-interactively. (I don't know how they detect that.) For example, `ls` has formatting/coloring and `ls | cat` does not. This solution seems like it would fit the problem from the article?
> This solution seems like it would fit the problem from the article?
Might not be a great idea. The world is probably already full of build tools pipelines that expect to process the normal terminal output (maybe with colours stripped). Environment variables like `CI` are a thing for a reason.
For Claude the most pollution usually comes from Claude itself.
It's worth noting thet just by setting the right tone of voice, choosing the right words, and instructing it to be concise, surgical in what it says and writes, things change drastically - like night and day.
It then starts obeying, CRITICALs are barely needed anymore and the docs it produces are tidy and pretty.
I think about what I do in these verbose situations; I learn to ignore most of the output and only take forward the important piece. That may be a success message or error. I've removed most of the output from my context window / memory.
I see some good research being done on how to allow LLMs to manage their own context. Most importantly, to remove things from their context but still allow subsequent search/retrieval.
LLMs (Claude Code in particular) will explicitly create token intensive steps, plans and responses - "just to be sure" - "need to check" - "verify no leftovers", will do git diff even tho not asked for, create python scripts for simple tasks, etc.
Absolutely no cache (except the memory which is meh) nor indexing whatsoever.
Pro plan for 20 bucks per month is essentially worthless and, because of this and we are entering new era - the era of $100+ monthly single subscription being something normal and natural.
I'm on the Pro plan. If I run out of tokens, which has only happened 2 or 3 times in months of use, I just work on something else that Claude can't do, or ...write the code myself.
You do have to keep a close eye on it, but I would be doing that anyway given that if it goes haywire it's wasting my time as well as tokens. I'd rather spend an extra minute writing a clearer prompt telling it exactly what I want it to do, than waste time on a slot machine.
>LLMs (Claude Code in particular) will explicitly create token intensive steps, plans and responses - "just to be sure" - "need to check" - "verify no leftovers", will do git diff even tho not asked for, create python scripts for simple tasks, etc. Absolutely no cache (except the memory which is meh) nor indexing whatsoever.
Most of these things can be avoided with a customized CLAUDE.md.
Not in my projects it seems. Perhaps you can share your best practices?
Moreover, avoiding these should be the default behaviour. Currently the default is to drain your pockets.
P.S
CLAUDE.md is sometimes useful but, it's a yet another token drain. Especially that it can grow exponentially.
Huh. I've noticed CC running build or test steps piped into greps, to cull useless chatter. It did this all by itself, without my explicit instructions.
Also, I just restart when the context window starts filling up. Small focused changes work better anyway IMO than single god-prompts that try do do everything but eventually exceed context and capability...
Surely any distinguished connoisseur of terminology gatekeeping such as yourself is able to distinguish between 'cc' and 'CC'. My terminal is able to spot the difference, you should be able to as well.
CC is an environment variable / internal variable used by most build tools to identify the current C compiler. cc is the standardized name of the executable, now usually an alias for gcc.
Both CC and cc refer to the C compiler, in slightly different ways.
Make (or whatever) targets that direct output to file and returns a subset have helped me quite a bit. Then wrap that in an agent that also knows how and when to return cached and filtered data from the output vs. rerunning. Fewer tokens spent reading output details that usually won't matter, coupled with less context pollution in the main agent from figuring out what to do.
If you use a smaller model for the sub agent you get all three
Of course you can combine both approaches for even greater gains. But Claude Code and like five alternatives gaining an efficient tool-calling paradigm where console output is interpreted by Haiku instead of Opus seems like a much quicker win than adding an LLM env flag to every cli tool under the sun
Dunno about that. Having used the $20 claude plan, I ran out of tokens within 30 minutes if running 3-4 agents at the same time. Often times, all 3-4 will run a build command at the end to confirm that the changes are successful. Thus the loss of tokens quickly gets out of hand.
Edit: Just remembered that sometimes, I see claude running the build step in two terminals, side-by-side at nearly the same time :D
So you can debug it without having to do a second build with extra flags, and in order to have a sense of what the build is doing at any particular time.
Given that most of the utility of Typescript is to make VSCode play nice for its human operator, _should_ we be using Typescript for systems that are written by machines?
Seeing a JSON configuration file that stores environment variables makes me want to cry. Just to think that somewhere, somehow, it's going to launch an entire JavaScript VM (tens of megabytes) just to parse a file with 12 lines in it, then extract from a JavaScript the fields, munge it, eventually turn into more or less an array of VAR=val C strings which get passed to a forked shell....
I noticed this on my spring boot side project. Successful test runs produce thousands of log lines in default mode because I like to e.g. log every executed SQL statement during development. It gives me visibility into what my orm is actually doing (yeh yeh I know I should just write SQL myself). For me it's just a bit of scrolling and cmd+f if I need to find something specific but Claude actually struggles a lot with this massive output.
Especially when it then tries to debug things finding the actual error message in the haystack of logs is suddenly very hard for the LLM. So I spent some time cleaning up my logs locally to improve the "agentic ergonomics" so to say.
In general I think good DevEx needs to be dialed to 11 for successful agentic coding. Clean software architecture and interfaces, good docs, etc. are all extremely valuable for LLMs because any bit of confusion, weird patterns or inconsistency can be learned by a human over time as a "quirk" of the code base. But for LLMs that don't have memory they are utterly confusing and will lead the agent down the wrong path eventually.
This seems like a really solid idea: using an environment variable in command line tools and small apps to control output for AI vs. human digestion. Even given efficient attention mechanisms, slop tokens in the context window are bad.
I also like a discussion in this thread: using custom tools to reduce the frequency of tool calls in general, that is, write tool wrappers specific for your applications or agents.
I wonder whether attention-free architectures like Mamba or Gated DeltaNet are distracted less by irrelevant context, because they don't recall every detail inside their context window in the first place. Theoretically it should be fairly easy to test this via a dedicated "context rot benchmark" (standard benchmarks but with/without irrelevant context).
great idea. thought about the waste of tokens dozens of times when I saw claude code increase the token count in the CLI after a build. I was wondering if there's a way to stop that, but not enough to actually look into it. I'd love for popular build tools to implement something along those lines!
Actually what we have is an entire stack, starting with Von Neumann arch, the kernel, the browser, auth —- it is optimized for the intuition of neither humans nor agents. All the legacy cruft that we glibly told people to RTFM on is now choking your agent and burning your tokens.
I have a solution to all this of course but why should I tell anyone.
Speaking of obvious questions. Why are you counting pennies instead of getting the LLM to do it? (Unless the idea was from an LLM and the executive decision was left to the operator, as well as posting the article)
So much content about furnishing the Markdown and the whatnot for your bots. But content is content?
This all seems like a lot of effort so that an agent can run `npm run build` for you.
I get the article's overall point, but if we're looking to optimise processing and reduce costs, then 'only using agents for things that benefit from using agents' seems like an immediate win.
You don't need an agent for simple, well-understood commands. Use them for things where the complexity/cost is worth it.
Feedback loops are important to agents. In the article, the agent runs this build command and notices an error. With that feedback loop, it can iterate a solution without requiring human intervention. But the fact that the build command pollutes the context in this case is a double-edge sword.
If you really need that, the easy solution here is to get a list of errors using an LSP (or any other way of getting a list of errors, even grep "Error:"), and only giving that list of errors to the LLM if the build fails. Otherwise just tell the LLM "build succeeded".
That's an extremely simple solution. I don't see the point in this LLM=true bullshit.
> Then a brick hits you in the face when it dawns on you that all of our tools are dumping crazy amounts of non-relevant context into stdout thereby polluting your context windows.
I've found that letting the agent write its own optimized script for dealing with some things can really help with this. Claude is now forbidden from using `gradlew` directly, and can only use a helper script we made. It clears, recompiles, publishes locally, tests, ... all with a few extra flags. And when a test fails, the stack trace is printed.
Before this, Claude had to do A TON of different calls, all messing up the context. And when tests failed, it started to read gradle's generated HTML/XML files, which damaged the context immensely, since they contain a bunch of inline javascript.
And I've also been implementing this "LLM=true"-like behaviour in most of my applications. When an LLM is using it, logging is less verbose, it's also deduplicated so it doesn't show the same line a hundred times, ...
> He sees something goes wrong, but now he cut off the stacktraces by using tail, so he tries again using a bigger tail. Not satisfied with what he sees HE TRIES AGAIN with a bigger tail, and … you see the problem. It’s like a dog chasing its own tail.
I've had the same issue. Claude was running the 5+ minute test suite MULTIPLE TIMES in succession, just with a different `| grep something` tacked at the end. Now, the scripts I made always logs the entire (simplified) output, and just prints the path to the temporary file. This works so much better.
> Claude is now forbidden from using `gradlew` directly, and can only use a helper script we made. It clears, recompiles, publishes locally, tests, ... all with a few extra flags. And when a test fails, the stack trace is printed.
I think my question at this point is what about this is specific to LLMs. Humans should not be forced to wade through reams of garbage output either.
Humans have the ability to ignore and generally not remember things after a short scan, prioritize what's actually important etc. But to an LLM a token is a token.
There's attempts at effectively doing something similar with analysis passes of the context - kinda what things like auto-compaction is doing - but I'm sure anyone who has used the current generation of those tools will tell you they're very much imperfect.
The “a token is a token” effect makes LLMs really bad at some things humans are great at, and really good at some things humans are terrible at.
For example, I quickly get bored looking through long logfiles for anomalies but an LLM can highlight those super quickly.
Isn’t the purpose of self attention exactly to recognize the relevance of some tokens over others?
That may help with tokens being "ignored" while still being in the context window, but not context window size costs and limitations in the first place.
In my experience, it's the old time-invested vs time-saved trade off. If you're not looking at these reams of output often enough, the incentive to figure out all the flags and configs for verbosity to write these script is lower: https://xkcd.com/1205/
And because these issues are often sporadic, doing all this would be an unwanted sidequest, so humans grit their teeth and wade through the garbage manually each time.
With LLMs, the cost is effectively 0 compared to a human, so it doesn't matter. Have them write the script. In fact, because it benefits the LLM by reducing context pollution, which increases their accuracy, such measures should be actively identified and put in place.
Lots of tools have a --quiet or --output json type option, which is usually helpful
The way I've solved this issue with a long running build script is to have a logging scripts which redirects all outputs into a file and can be included with ``` # Redirect all output to a log file (re-execs script with redirection) source "$(dirname "$0")/common/logging.sh" ``` at the start of a script.
Then when the script runs the output is put into a file, and the LLM can search that. Works like a charm.
This has been my exact experience with agents using gradle and it’s beyond frustrating to watch. I’ve been meaning to set up my own low-noise wrapper script.
This post just inspired me to tackle this once and for all today.
Wow, I'd love to do this. Any tips on how to build this (or how to help an LLM build this), specifically for ./gradlew?
How is it forbidden? I tell agents to use my wrappers in AGENTS but they ignore it half the time and use the naked tool.
If you get desperate, I've given my agent a custom $PATH that replaces the forbidden tools with shims that either call the correct tool, or at least tell it what to do differently.
~/agent-shims/mvn:
That way it is prevented from using the wrong tools, and can self-correct when it tries.Permissions scoping
Then they attempt to download the missing tool or write a substitute from scratch. Am I the only one who runs into this??
> Then a brick hits you in the face when it dawns on you that all of our tools are dumping crazy amounts of non-relevant context into stdout thereby polluting your context windows.
Not just context windows. Lots of that crap is completely useless for humans too. It's not a rare occurrence for warnings to be hidden in so much irrelevant output that they're there for years before someone notices.
The old unix philosophy of "print nothing on success" looks crazy until you start trying to build pipes and shell scripts that use multiple tools internally. Also very quickly makes it clear why stdout and stderr are separate
Also becomes rapidly apparent that most modern tooling takes an extremely liberal view of logging levels. The fact that you’ve successfully processed a file is not INFO, that’s DEBUG.
"Finished conversion of xyz.mp3 to xyz.ogg" is valuable progress information to a regular user, not just to developers, so it belongs in INFO, not DEBUG.
I suppose this is subjective, but I disagree. If I want to know the status of each item, I’d pass -v to the command. A simple summary at the end is sufficient; if I pass -q, I expect it to print nothing, only issuing a return code.
It never felt crazy to me, with the exception that are many situations where having progress and diagnostic information (usually opt-in) made sense.
I guess it comes down to a choice of printing out only relevant information. I hate noisy crap, like LaTeX.
Yeah. Maybe we only need:
for the unusual case when you do want the `route print` on a BGP router to actually dump 8 gigabytes of text throughout next 2 minutes. Maybe it's fine if a default output for anything generously applies summarization, such as "X, Y, Z ...and 9 thousand+ similar entries".Having two separate command names (one for human/llm, one for batch) sucks.
Having `-h` for human, like ls or df do, sucks slightly less, but it is still a backward-compatibility hack which leads to `alias` proliferation and makes human lifes worse.
Something related to this article, but not related to AI:
As someone who loves coding pet projects but is not a software engineer by profession, I find the paradigm of maintaining all these config files and environment variables exhausting, and there seem to be more and more of them for any non-trivial projects.
Not only do I find it hard to remember which is which or to locate any specific setting, their mechanisms often feel mysterious too: I often have to manually test them to see if they actually work or how exactly. This is not the case for actual code, where I can understand the logic just by reading it, since it has a clearer flow.
And I just can’t make myself blindly copy other people's config/env files without knowing what each switch is doing. This makes building projects, and especially copying or imitating other people's projects, a frustrating experience.
How do you deal with this better, my fellow professionals?
Software folks love over-engineering things. If you look at the web coding craze of a few years ago, people started piling up tooling on top of tooling (frameworks, build pipelines, linting, generators etc.) for something that could also be zero-config, and just a handful of files for simple projects.
I guess this happens when you're too deep in a topic and forget that eventually the overhead of maintaining the tooling outweights the benefits. It's a curse of our profession. We build and automate things, so we naturally want to build and automate tooling for doing the things we do.
I don’t think those web tooling piles are over-engineered per se, they address huge challenges at Google and Facebook, but the profession is way too driven by hype and fashion and the result is a lot of cargo culting of stuff from Big Dogs unquestioningly. Wrong tooling for the job creates that bubble of over complicated app development.
Inventing GraphQL and React and making your own PHP compiler are absolutely insane and obviously wrong decisions — for everyone who isn’t Facebook. With Facebook revenue and Facebooks army of resume obsessed PHP monkeys they strike me as elegant technological solutions to otherwise intractable organizational issues. Insane, but highly profitable and fast moving. Outside of that context using React should be addressing clear pain points, not a dogmatic default.
We’re seeing some active pushback on it now online, but so much damage has been done. Embracing progressive complexity of web apps/sites should leave the majority as barebones with minimal if any JavaScript.
Facebook solutions for Facebook problems. Most of us can be deeply happy our 99 problems don’t include theirs, and live a simpler easier life.
Not sure why you lumped React in there. Hack is loopy, and GraphQL was overhyped but conditionally useful, but React was legitimately useful and a real improvement over other ways of doing things at the time. Compare React to contemporary stuff like jQuery, Backbone, Knockout, Angular 1.x, etc.
I agree with you very much, if what you are building actually benefits from that much client side interactivity. I think the counterpoint is that most products could be server rendered html templates with a tiny amount of plain js rather than complex frontend applications.
First of all, I read the documentation for the tools I'm trying to configure.
I know this is very 20th century, but it helps a lot to understand how everything fits together and to remember what each tool does in a complex stack.
Documentation is not always perfect or complete, but it makes it much easier to find parameters in config files and know which ones to tweak.
And when the documentation falls short, the old adage applies: "Use the source, Luke."
> As someone who loves coding pet projects but is not a software engineer by profession, I find the paradigm of maintaining all these config files and environment variables exhausting
Then don’t.
> How do you deal with this better, my fellow professionals?
By not doing it.
Look, it’s your project. Why are you frustrating yourself? What you do is you set up your environment, your configuration, what you need/understand/prefer and that’s it. You’ll find out what those are as you go along. If you need, document each line as you add it. Don’t complicate it.
Honestly... ask an AI agent to update them for you.
They do an excellent job of reading documentation and searching to pick and choose and filter config that you might care about.
After decades of maintaining them myself, this was a huge breath of fresh air for me.
Don't fall for the "JS ecosystem" trap and use sane tools. If a floobergloob requires you to add a floobergloob.config.js to your project root that's a very good indicator floobergloob is not worth your time.
The only boilerplate files you need in a JS repo root are gitignore, package.json, package-lock.json and optionally tsconfig if you're using TS.
A node.js project shouldn't require a build step, and most websites can get away with a single build.js that calls your bundler (esbuild) and copies some static files dist/.
Simplify your tools and build processes to as few as possible, and pick tools with fewer (or no) config files.
It could depend on what you're doing, but if it's not for work the config hell is probably optional.
You start with the cleanest most minimal config you can get away with, but over the years you keep adding small additions and tweaks until it becomes a massive behemoth that only you will ever understand the reasoning behind.
Right, and then when you don't work on it for 6 or 12 months, you come back and find that now you don't understand it either.
Part of doing it well is adding comments as you add options. When I used vim, every line or block in the config had an accompanying comment explaining what it did, except if the config’s name was so obvious that a comment would just repeat it.
That's a good call. It's a big problem for JSON configs given pure JSON's strict no-comments policy. I like tools that let you use .js or better yet .ts files for config.
Or consider jsonc - json with comments - or jwcc - which is json with comments and trailing commas to make life a little easier.
https://jsonc.org/
https://nigeltao.github.io/blog/2021/json-with-commas-commen...
There are a lot of implementations of all of these, such as https://github.com/tailscale/hujson
I like this idea a lot, and pushed for json5 at a previous job, but I think there are a few snags:
- it's weird and unfamiliar, most people prefer plain JSON
- there are too many competing standards to choose from
- most existing tools just use plain JSON (sometimes with support for non-standard features, like tsconfig allowing trailing commas, but usually poorly documented and unreliable)
Much easier just to make the leap to .ts files, which are ergonomically better in almost every way anyway.
A lot of json parsers will permit comments even though it isn't meant to be valid. Worth trying it, see if a comment breaks the config, and if not then use comments and don't worry about it.
For reference, jq and python don't allow comments.
I think the concept has value, but I think targeting today's LLMs like this is short sighted.
It's making what is likely to be a permanent change to fix a temporary problem.
I think the thing that would have value in the long term is an option to be concise, accurate, and unambiguous.
This isn't something that should be considered to be only for LLMs. Sometimes humans want readability to understand something quickly adding context helps a great deal here, but sometimes accuracy and unambiguity are paramount (like when doing an audit) if dealing with a batch of similar things, the same repeated context adds nothing and limits how much you can see at once.
So there can be a benefit when a human can request output like this for them to read directly. On top of this is the broad range of of output processing tools that we have (some people still awk).
So yes, this is needed, but LLMs will probably not need this in a few years. The other uses will remain
Rather than an LLM=true, this is better handled with standardizing quiet/verbose settings, as this is a question of verbosity, where an LLM is one instance where you usually want it to be quieter, but not always.
Secondly, a helper to capture output and cache it, and frankly a tool or just options to the regular shell/bash tools to cache output and allow filtered retrieval of the cached output, as more so than context and tokens the frustration I have with the patterns shown is that often the agent will re-execute time-consuming tasks to retrieve a different set of lines from the output.
A lot of the time it might even be best to run the tool with verbose output, but it'd be nice if tools had a more uniform way of giving output that was easier to systematically filter to essentials on first run (while caching the rest).
Yes! After seeing a lot of discussions like this, I came up with a rule of thumb:
Any special accommodations you make for LLMs are either a) also good for humans, or b) more trouble than they're worth.
It would be nice for both LLMs and humans to have a tool that hides verbose tool output, but still lets you go back and inspect it if there's a problem. Although in practice as a human I just minimise the terminal and ignore the spam until it finishes. Maybe LLMs just need their own equivalent of that, rather than always being hooked up directly to the stdout firehose.
Yes, what's preventing the LLM from running myCommand > /tmp/out_someHash.txt ; tail out_someHash.txt and then greping or tailing around /tmp/out_someHash.txt on failure?
There isn't really anything other than training, but they generally don't. You probably can get them to do that with some extra instructions, but part of the problem - at least with Claude - is that it's really trigger-happy about re-running the commands if it doesn't get the results it likes, assuming the results reflects stale results. Even with very expensive (in time) scripts I often see it start a run, pipe it to a file, put it in the background, then loop on sleep statements, occasionally get "frustrated" and check, only to throw the results away 30 seconds after they are done because it's made an unrelated change.
A lot of the time this behaviour is probably right. But it's annoyingly hard to steer it to handle this correctly. I've had it do this even with make targets where the makefile itself makes clear the dependencies means it could trust the cached (in a file) results if it just runs make <target>. Instead I regularly find it reading the Makefile and running the commands manually to work around the dependency management.
It feels wild to have to keep reminding people, but AI changes very little. Tools have always had a variety of output, and ways to control this, and bad tools output a lot by default, whilst good tools hide it behind version of "-v" or easy greps. Don't add a --LLM or whatever, do add cleaner and consistent verbosity controls.
The UNIX philosophy of tools that handle text streams, staying "quiet" unless something goes wrong, doing one thing well, etc. are all still so well suited to the modern age of AI coding agents.
Also an acceptable solution - create a "runner" subagent on a cheap model, that's tasked with running a command and relaying the important parts to the main agent.
Yes, this is the solution. An agent that can clean up the output of irrelevant stuff
Why can't the agent harness dynamically decide whether outputs should be put into the context or not? It could check with an LLM to determine if the verbatim output seems important, and if not, store the full output locally but replace it in the prompt with a brief summary and unique ID. Then make a tool available so the full output can be retrieved later if necessary. That's roughly how humans do it, you scroll through your terminal and make quick decisions about what parts you can ignore, and then maybe come back later when you realize "oh I should probably read that whole stack trace".
It wouldn't even need to send the full output to make a decision, it could just send "npm run build output 500 lines and succeeded, do we need to read the output?" and based on the rest of the conversation the LLM can respond yes or no.
Isn't that what subagents do to a certain degree?
Sort of, but you also want to keep the sub-agent context small for as long as possible, and if you're paying per token there's no reason to be sending thousands of tokens that are probably useless.
We’ve got a long way to go in optimising our environments for these models. Our perception of a terminal is much closer to feeding a video into Gemini than reading a textbook of logs. But we don’t make that ax affordance at the moment.
I wrote a small game for my dev team to experience what it’s like interacting through these painful interfaces over the summer www.youareanagent.app
Jump to the agentic coding level or the mcp level to experience true frustration (call it empathy). I also wrote up a lot more thinking here www.robkopel.me/field-notes/ax-agent-experience/
Beautiful simulation.
So frequently beginners in linux command lines complain about the irregularity or redundance in command line tool conventions (sometimes actual command parameters -h --help or /h ? other times: man vs info; etc...)
When the first transformers that did more than poetry or rough translation appeared everybody noticed their flaws, but I observed that a dumb enough (or smart enough to be dangerous?) LLM could be useful in regularizing parameter conventions. I would ask an LLM how to do this or that, and it would "helpfully" generate non-functional command invocations that otherwise appeared very 'conformant' to the point that sometimes my opinion was that -even though the invocation was wrong given the current calling convention for a specific tool- it would actually improve the tool if it accepted that human-machine ABI or calling convention.
Now let us take the example of man vs info, I am not proposing to let AI decide we should all settle on man; nor do I propose to let AI decide we should all use info instead, but with AI we could have the documentation made whole in the missing half, and then it's up to the user if they prefer man or info to fetch the documentation of that tool.
Similarily for calling conventions, we could ask LLM's to assemble parameter styles and analyze command calling conventions / parameters and then find one or more canonical ways to communicate this, perhaps consulting an environment variable to figure out what calling convention the user declares to use.
Similarly law professor Rob Anderson joked on X that llm hallucinated cases are good law:
https://x.com/ProfRobAnderson/status/2019078989348774129
> Indeed hallucinated cases are "better law." Drawing on Ronald Dworkin's theory of law as integrity, which posits that ideal legal decisions must "fit" existing precedents while advancing principled justice, this article argues that these hallucinations represent emergent normative ideals. AI models, trained on vast corpora of real case law, synthesize patterns to produce rulings that optimally align with underlying legal principles, filling gaps in the doctrinal landscape. Rather than errors, they embody the "cases that should exist," reflecting a Hercules-like judge's holistic interpretation.
Seems naive. You can get an LLM to agree with almost anything if you say the right things to it, and it will hallucinate citations to back you up without skipping a beat. You can probably get it to hallucinate case law to legalize murder on Mondays.
It has long been a pet peeve of mine that the *nix world has no standard reliable convention for how to interrogate a program for it's available flags. Instead there are at least a dozen ways it can be done and you can't rely on any one of them.
That's not specific to unix though.
I didn't say it was, but I simply don't care what problems any other kind of system has because they aren't my problems. Last time I had windows on any of my computers it was windows 98.
I've been thinking about using an OpenAPI schema to describe cli tools. It would probably need extensions to describe stdin/stout and other things that don't happen in REST.
Have you seen Stainless's CLI generator?
https://www.stainless.com/blog/stainless-cli-generator-your-...
Ah yes, the vaunted ffmpeg-llm --"take these jpegs and turn them into an mp4 and use music.mp3 as the soundtrack" command.
Ngl.. I can see the merit and simultaneously recoil in horror as I am starting to understand what linux greybeards hate about windofication of linux ( and now proposed llm-ification of it :D).
> The environment wins (less tokens burned = less energy consumed)
This is understandable logic, but at a systemic level it's not how things always go. Increasing efficiency can lead to increased consumption overall. You might save 50% in energy for your workload, but maybe now you can run it 3 times as much, or maybe 3 times more people will use it, because it's cheaper. The result might be a 50% INCREASE in energy consumed.
https://en.wikipedia.org/wiki/Jevons_paradox
This is the standing reason that is always given for why we must all sit in freeway traffic clogs, and I think it's B.S., because it assumes that there are viable alternatives available in near-medium term, but that isn't always the case. The alternative to freeways that are supposed to compensate is a joint combination of denser housing and mass transit, which in California, is not happening at all...zoning laws and the slow pace of building mass transit due to regulation slow-down and the need to service urban sprawl, prevent that solution from relieving traffic pressure. Don't speak of busses, because taking two hours to get to work is not better than one hour. So..the freeways stay the same number of lanes and my commute time continues to grow, and I am tired of hearing it is for the best.
So yes, lower LLM costs would probably lead even more LLM usage and greater energy expenditures, but then again, so does having a moving economy, and all that comes with that.
Yeah, probably. I wonder where speed-running fixing all the low-hanging fruit for AI-related efficiency improvements will leave us? It still seems worth doing. Maybe combined with a carbon tax.
Surprisingly often people refuse to document their architecture or workflow for new hires. However, when it's for an LLM some of these same people are suddenly willing to spend a lot of time and effort detailing architecture, process, workflows.
I've seen projects with an empty README and a very extensive CLAUDE.md (or equivalent).
That could be because Claude offers a dedicated /init command to generate a CLAUDE.md if it doesn't exist.
I like the gist of this, however LLM may not be the best name for this: what if a new tech (e.g., SLM) takes over? AGENT may be a more faithful name until something better is standardized.
Looks like the blog could use a HN=True. Hope the author won't get banned...
> Error: API rate limit exceeded for app ID 7cc6c241b6e6762bf384. If you reach out to GitHub Support for help, please include the request ID E9FC:7BEBA:6CDB3B4:6485458:699EE247 and timestamp 2026-02-25 11:51:35 UTC. For more on scraping GitHub and how it may affect your rights, please review our Terms of Service (https://docs.github.com/en/site-policy/github-terms/github-t...).
Author here. Thanks for flagging. Let me look into it
I would use this as a human. That npm output is crazy. Maybe a better variable would be "CONCISE=1". For LLMs, there are a few easier solutions, like outputing in a file (and then tail)., or running a subagent
On a lot of linux distros there is the `moreutils` package, which contains a command called `chronic`. Originally intended to be used in crontabs, it executes a command and only outputs its output if it fails. I think this could find another use case here.
This is great, I like this. Wrote a 'chronic-file' variant that just dumps everything into a tmpfile and outputs the filepath for the agent in case of error and otherwise nothing
Useful enough to justify registering on HN. Thank you!
Many unix tools already print less logging when used im a script, ie. non-interactively. (I don't know how they detect that.) For example, `ls` has formatting/coloring and `ls | cat` does not. This solution seems like it would fit the problem from the article?
> I don't know how they detect that.
The OS knows (it has to because it set up the pipeline), and the process can find out through a system call, exposed in C as `isatty`: https://www.man7.org/linux/man-pages/man3/isatty.3.html
> This solution seems like it would fit the problem from the article?
Might not be a great idea. The world is probably already full of build tools pipelines that expect to process the normal terminal output (maybe with colours stripped). Environment variables like `CI` are a thing for a reason.
There’s a function isatty that detect if a file descriptor (stdout is one) is associated with a terminal
https://man.openbsd.org/man3/ttyname.3
I believe most standard libraries has a version.
I was about to comment the same thing. Usually I don't call the function directly, but via the tty command in my shell scripts:
Now I wonder how _isatty_ itself detects whether a file descriptor is associated with a terminal!In OpenBSD, with the fcntl system call
https://github.com/openbsd/src/blob/master/lib/libc/gen/isat...
https://man.openbsd.org/fcntl
https://github.com/openbsd/src/blob/master/sys/sys/fcntl.h
https://github.com/openbsd/src/blob/ba496e5267528b649ec87212...
https://github.com/openbsd/src/blob/ba496e5267528b649ec87212...
For Claude the most pollution usually comes from Claude itself.
It's worth noting thet just by setting the right tone of voice, choosing the right words, and instructing it to be concise, surgical in what it says and writes, things change drastically - like night and day.
It then starts obeying, CRITICALs are barely needed anymore and the docs it produces are tidy and pretty.
I think about what I do in these verbose situations; I learn to ignore most of the output and only take forward the important piece. That may be a success message or error. I've removed most of the output from my context window / memory.
I see some good research being done on how to allow LLMs to manage their own context. Most importantly, to remove things from their context but still allow subsequent search/retrieval.
This is merely scratching the surface.
LLMs (Claude Code in particular) will explicitly create token intensive steps, plans and responses - "just to be sure" - "need to check" - "verify no leftovers", will do git diff even tho not asked for, create python scripts for simple tasks, etc. Absolutely no cache (except the memory which is meh) nor indexing whatsoever.
Pro plan for 20 bucks per month is essentially worthless and, because of this and we are entering new era - the era of $100+ monthly single subscription being something normal and natural.
I assume you're using Opus.
I'm on the Pro plan. If I run out of tokens, which has only happened 2 or 3 times in months of use, I just work on something else that Claude can't do, or ...write the code myself.
You do have to keep a close eye on it, but I would be doing that anyway given that if it goes haywire it's wasting my time as well as tokens. I'd rather spend an extra minute writing a clearer prompt telling it exactly what I want it to do, than waste time on a slot machine.
>LLMs (Claude Code in particular) will explicitly create token intensive steps, plans and responses - "just to be sure" - "need to check" - "verify no leftovers", will do git diff even tho not asked for, create python scripts for simple tasks, etc. Absolutely no cache (except the memory which is meh) nor indexing whatsoever.
Most of these things can be avoided with a customized CLAUDE.md.
Not in my projects it seems. Perhaps you can share your best practices? Moreover, avoiding these should be the default behaviour. Currently the default is to drain your pockets.
P.S CLAUDE.md is sometimes useful but, it's a yet another token drain. Especially that it can grow exponentially.
I agree it should be the default, but it isn't. I also tend to think it's mainly to make money by default.
Another thing that helps is using plan mode first, since you can more or less see how it's going to proceed and steer it beforehand.
codex seems chill
And yet when my $100 CC Pro renewed last month my instinctive thought was wow is that all?
That can be read in two ways:
1) It's only $100, well worth the money.
2) Surprisingly little value was provide for all that money.
This is basically what RTK "Rust Token Killer" does.
Removes all the fluff around commands that agents use frequently.
https://github.com/rtk-ai/rtk
Huh. I've noticed CC running build or test steps piped into greps, to cull useless chatter. It did this all by itself, without my explicit instructions.
Also, I just restart when the context window starts filling up. Small focused changes work better anyway IMO than single god-prompts that try do do everything but eventually exceed context and capability...
cc is the C compiler.
Please don't overload that term with trendy LLM products. You can use the full name.
Surely any distinguished connoisseur of terminology gatekeeping such as yourself is able to distinguish between 'cc' and 'CC'. My terminal is able to spot the difference, you should be able to as well.
CC is an environment variable / internal variable used by most build tools to identify the current C compiler. cc is the standardized name of the executable, now usually an alias for gcc.
Both CC and cc refer to the C compiler, in slightly different ways.
Funny! I built an entire cli and ecosystem around this:
https://github.com/bodo-run/stop-nagging
could we not instruct the LLM to run build commands in a sub agents which could then just return a summary of what happened?
this avoids having to update everything to support LLM=true and keep your current context window free of noise.
Make (or whatever) targets that direct output to file and returns a subset have helped me quite a bit. Then wrap that in an agent that also knows how and when to return cached and filtered data from the output vs. rerunning. Fewer tokens spent reading output details that usually won't matter, coupled with less context pollution in the main agent from figuring out what to do.
q() { local output output=$("$@" 2>&1) local ec=$? echo "$output" | tail -5 return $ec }
There :)
That would achieve 1 of the 3 wins.
If you use a smaller model for the sub agent you get all three
Of course you can combine both approaches for even greater gains. But Claude Code and like five alternatives gaining an efficient tool-calling paradigm where console output is interpreted by Haiku instead of Opus seems like a much quicker win than adding an LLM env flag to every cli tool under the sun
Probably the main one, people mostly complain about context window management rather than token usage
Dunno about that. Having used the $20 claude plan, I ran out of tokens within 30 minutes if running 3-4 agents at the same time. Often times, all 3-4 will run a build command at the end to confirm that the changes are successful. Thus the loss of tokens quickly gets out of hand.
Edit: Just remembered that sometimes, I see claude running the build step in two terminals, side-by-side at nearly the same time :D
If the output of your build tool is too verbose for a mechanical brain to keep on top of, did the meat brain ever stand a chance?
Why was the output so verbose in the first place then?
So you can debug it without having to do a second build with extra flags, and in order to have a sense of what the build is doing at any particular time.
All of this because we only have stdout and stderr and nothing in between. I wish there was a stdlog or stddebug or something
yes, if only we had a more fine-grained log level hierarchy that we could get every piece of software to agree to...
Given that most of the utility of Typescript is to make VSCode play nice for its human operator, _should_ we be using Typescript for systems that are written by machines?
Seeing a JSON configuration file that stores environment variables makes me want to cry. Just to think that somewhere, somehow, it's going to launch an entire JavaScript VM (tens of megabytes) just to parse a file with 12 lines in it, then extract from a JavaScript the fields, munge it, eventually turn into more or less an array of VAR=val C strings which get passed to a forked shell....
Why do you presume it needs a JavaScript VM to parse JSON?
Granted, I have no idea how Claude Code operates internally, but if it’s already running in a JS VM, can’t it read the file itself?
Most of what helps LLMs here is exactly what helps humans: less noise, clearer signals, predictable output.
Can TOON format help in this, with "LLM=true" we can reduce the noise which pollutes context
I never considered the volume of output tokens from dev tools, but yeah, I like this idea a lot.
I've been using CODING_AGENT=true
I noticed this on my spring boot side project. Successful test runs produce thousands of log lines in default mode because I like to e.g. log every executed SQL statement during development. It gives me visibility into what my orm is actually doing (yeh yeh I know I should just write SQL myself). For me it's just a bit of scrolling and cmd+f if I need to find something specific but Claude actually struggles a lot with this massive output. Especially when it then tries to debug things finding the actual error message in the haystack of logs is suddenly very hard for the LLM. So I spent some time cleaning up my logs locally to improve the "agentic ergonomics" so to say.
In general I think good DevEx needs to be dialed to 11 for successful agentic coding. Clean software architecture and interfaces, good docs, etc. are all extremely valuable for LLMs because any bit of confusion, weird patterns or inconsistency can be learned by a human over time as a "quirk" of the code base. But for LLMs that don't have memory they are utterly confusing and will lead the agent down the wrong path eventually.
We’ve reinvented exit codes…
This seems like a really solid idea: using an environment variable in command line tools and small apps to control output for AI vs. human digestion. Even given efficient attention mechanisms, slop tokens in the context window are bad.
I also like a discussion in this thread: using custom tools to reduce the frequency of tool calls in general, that is, write tool wrappers specific for your applications or agents.
Interesting idea but bad tspec. A better approach would be a single env var (DEV_MODE perhaps) with “agent” and “human” as values (and maybe “ci”).
I wonder whether attention-free architectures like Mamba or Gated DeltaNet are distracted less by irrelevant context, because they don't recall every detail inside their context window in the first place. Theoretically it should be fairly easy to test this via a dedicated "context rot benchmark" (standard benchmarks but with/without irrelevant context).
great idea. thought about the waste of tokens dozens of times when I saw claude code increase the token count in the CLI after a build. I was wondering if there's a way to stop that, but not enough to actually look into it. I'd love for popular build tools to implement something along those lines!
I think I noticed LLMs doing >/dev/null on routine operations.
`Humans=True`
The best friend isn't a dog, but the family that you build. Wife/Husband/kids. Those are going to be your best friends for life.
MCP as an env-var ;)
Actually what we have is an entire stack, starting with Von Neumann arch, the kernel, the browser, auth —- it is optimized for the intuition of neither humans nor agents. All the legacy cruft that we glibly told people to RTFM on is now choking your agent and burning your tokens.
I have a solution to all this of course but why should I tell anyone.
Can we just instruct the agents to redirect output streams to files, and then use grep to retrieve the necessary lines?
Speaking of obvious questions. Why are you counting pennies instead of getting the LLM to do it? (Unless the idea was from an LLM and the executive decision was left to the operator, as well as posting the article)
So much content about furnishing the Markdown and the whatnot for your bots. But content is content?
Or, stop outputting crap and use a logger. Make an LLM-only logger for output LLMs need and use stdout for HUMAN things.
This all seems like a lot of effort so that an agent can run `npm run build` for you.
I get the article's overall point, but if we're looking to optimise processing and reduce costs, then 'only using agents for things that benefit from using agents' seems like an immediate win.
You don't need an agent for simple, well-understood commands. Use them for things where the complexity/cost is worth it.
Feedback loops are important to agents. In the article, the agent runs this build command and notices an error. With that feedback loop, it can iterate a solution without requiring human intervention. But the fact that the build command pollutes the context in this case is a double-edge sword.
If you really need that, the easy solution here is to get a list of errors using an LSP (or any other way of getting a list of errors, even grep "Error:"), and only giving that list of errors to the LLM if the build fails. Otherwise just tell the LLM "build succeeded".
That's an extremely simple solution. I don't see the point in this LLM=true bullshit.
But those simple and well understood commands can be part of a huge workflow the LLM embarks on.
Again, if your priority is to minimise costs, then not forcing every part of that huge workflow through the agent is a good start.