Claude Code Unleashed
ymichael.com150 points by ymichael 3 days ago
150 points by ymichael 3 days ago
I am using Claude Code full-time for about 6 weeks* with the $20/month subscription. I am trying out building different products from ideas I have already had. It frees me a lot of time to talk about my founder journey.
I have not needed multiple agents or using CC over an SSH terminal to run overnight. The main reason is that LLMs are not correct many times. So I still need time to test. Like run the whole app, or check what broke in CI (GitHub Actions), etc. I do not go through code line by line anymore and I organize work with tickets (sometimes they are created with CC too).
Both https://github.com/pixlie/Pixlie and https://github.com/pixlie/SmartCrawler are vibe coded (barely any code that I wrote). With LLMs you can generated code 10x than writing manually. It means you can also get 10x the errors. So the manual checks take some time.
Our existing engineering practices are very helpful when generating code through LLMs and I do not have mental bandwidth to review a mountain of code. I am not sure if we scale out LLMs, it will help in building production quality software. I already see that sometimes CC makes really poor guesses. Imagine many such guesses in parallel, daily.
edit: typo - months/weeks
If there is barely any code in those repos that you wrote, how can you license them under the GPL? You don't hold the copyright for it.
This genuinely isn't an attack, I just don't think you can? The AI isn't granted copyright over what it produces.
I can only talk about the law in England & Wales, but:
For code generated by an LLM the human user would likely be considered the author if you provided sufficient creative input, direction, or modification.
The level of human involvement matters, simply prompting "write me a function" might not be enough, but providing detailed specifications, reviewing, and modifying the output would strengthen the claim.
the Copyright, Designs and Patents Act 1988 (CDPA), Section 9(3) staes, "In the case of a literary, dramatic, musical or artistic work which is computer-generated, the author shall be taken to be the person by whom the arrangements necessary for the creation of the work are undertaken". This was written before LLM's existed, but recent academic literature has supported this position, https://academic.oup.com/jiplp/article/19/1/43/7485196?login...
However, a comparable situation was tested with Thaler v Comptroller-General, where courts emphasised that legal rights require meaningful human involvement, not just ownership of the AI system. - https://www.culawreview.org/journal/unlocking-the-canvas-a-l... and https://www.whitecase.com/insight-our-thinking/uk-supreme-co...
I do acknowledge there is uncertainty, and this is highlighted here in "The Curious Case of Computer-Generated Works under the Copyright, Designs and Patents Act 1988.", with "section 9(3): the section is either unnecessary or unjustifiably extends legal protection to a class of works which belong in the public domain" - https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4072004
Today, I think it's doubtful that a functional application can be entirely vibe coded without decent direction and modification, but I don't think that will always be the case.
At least for art there is already precedent in US courts with someone trying to copyright an image generated by midjourney and it getting revoked in 22, because ai generated art cannot be copyrighted.
for code it hasn't been challenged yet, but I find it doubtful they'd decide differently there
I was reading Doe 1 v. GitHub for my paper. The case involves open source developers suing Github Copilot which were trained on, and generating open source code including code with MIT and AGPL license.
So far, the judge believe that training models on open source code is not a license violation as the code is public for anyone to read, but by "distribution or redistribution" (I assume, of the model's outputs?) it is still up for the court's decision whether that violate the terms of the license, among other laws.
The case is currently moved to Ninth Circuit without a decision in the district court, as there are other similar cases (such as Authors Guild's) and they wanted that the courts would offer a consistent rules. I believe one of the big delay in the case is in damages, which I think the plaintiff tried to ask for details of Microsoft's valuation of GitHub when it was acquired, as GitHub's biggest asset is the Git repositories and may provide a monetary value of how much each project is worth. Microsoft is trying to stall and not reveal this.
Assuming you're referring to Thaler v. Perlmutter, Thaler claimed to the copyright office that the image at issue was "autonomously created by a computer algorithm running on a machine". So the question of "if you claim the LLM did it itself" is settled (shocker, cf. Naruto v. Slater, 888 F.3d 418), but that definitely did not settle "_I_ used the LLM to do it".
Tbf, IANAL and was only repeating what journalists wrote back then. Ultimately, I have no deeper knowledge of the laws in question and thus don't have a qualified opinion on the matter.
I'm not so sure about that.
The legal standards in the United States for software copyrights are Jaslow and Altai, known to Federal courts as SSO [0] and AFC [1], respectively.
These standards consider the overall structure of code as being copyrightable. This means that you can't just rename a bunch of variables and class names. The overall organization of the code is considered an arbitrary expression. Someone would be infringing on copyright if they took your Java code and converted it to Python with different class, variable and function names but kept the same relationships between classes and the same general structure.
So what does this have to do with LLMs? Well, if the author directed the code to be structured in a certain way, directed to create specific APIs, etc, then there is a legal argument that the author has at least copyright over the arbitrary and expressive decisions that were made while building a software system.
[0] https://en.wikipedia.org/wiki/Structure,_sequence_and_organi...
[1] https://en.wikipedia.org/wiki/Abstraction-Filtration-Compari...
I have not thought about it, it is one of the things on my list. But my understanding was that developers copy code from Stack Overflow, as an example. It is not "my" code but I still am the author. Or lets say I ask my friend to add code and she/he simply passes over the code to me. I author it in my name.
The "barely" part may be important and I would like to know what others are doing.
I don't think you can just willy-nilly copy code from StackOverflow and sign it with your name. It's license forbids it. You also can't just sign your friend's code with your name unless she explicitly gives you permission. In both cases you are not the author of that code.
I get that people do it anyway but I guess it's kind of a grey-area because it's hard to tell after the fact that some snippet has been copied from SO.
I got a patch rejected (rightfully so IMO) a long-time ago from libvirt (RedHat) because I was using (and mentioning) code taken from StackOverflow.
So what would be the status of this code? Nobody holds the copyright for it, so anyone can use it in any way and nobody can sue for anything. It's not GPL, but it sounds pretty open source to me.
Yes, my understanding is that non-humans in the USA, cannot be granted copyright. This puts the work in the public domain, which means it can't be relicensed.
There was a much appealed case of a monkey taking a photo, where it was decided the photo was in the public domain.
https://en.wikipedia.org/wiki/Monkey_selfie_copyright_disput...
It boiled down to the creator not being a "legal person" and so could not hold copyright.
The real problem for software is where the line is for a "sufficient" transformation from the source material by a human to make it acquire copyright. You can write a Dickens' character derived novel and have copyright in it, but not gain control over those characters as Dickens described them.
Can you buy Jules Verns book, add comments and claim copyright on the whole book?
Claim partial copyright without specifying clearly what exactly?
Absolutely.
People sell annotated Bibles, or Shakespeare etc. You can transform it in to something that can acquire copyright, but it must have an artistic step.
This is a big thing in the fine art world as well, you can take inspiration, you can in some circumstances outright copy, but then you need to transform it sufficiently that it becomes your own art. People argue in front of judges about this stuff, of course.
Verne is a good example too, because if you print an English version, the translator acquires copyright in the translated version.
IANAL but the AI is a tool and presumably the code should be treated as any other auto-generated. Its the product of the tool user.
Unless the product includes code licensed by others, then - like any other repo - I don't see any license issue here.
If you mean there is no insight as to whether licensed code is included, that's one of the constraints of vibe-coding (which people often confuse with AI-assisted coding).
Its the job of the user to check and curate the contributions as they would any third-party human input (eg. via prs). Again though - that's not an AI coding issue, but a human process decision.
No, there is no copyright.
If you tried to sue someone for copyright infringement based on code that an LLM generated for you, you'd be laughed out of court.
But you were the one that used the LLM to generate it, so that’s your code, surely - how would unlicensed use of your code not violate copyright? Why didn’t they ‘just’ use an LLM to generate their own code?
The product is not owned by the tool user.
Use a hammer, you own the output. Use an intern, the intern does.
Of course if you're aren't a person you can't own anything.
Correct. No copyright. No legal teeth to the GPL.
Take whatever you want and relicense it cause it doesn't belong to the "author"
Lolololololol "author"
What else should they do then? GPL is a good call. Most keep the output under their own personal/company IP.
> What else should they do then?
They can say the code is in the public domain.
This is distinct from open source, yes, but in almost all cases less restricted than anything with a (open source or otherwise) license.
My trick is - ask for a plan. Revise the plan. Then ask only to work on a single step of the plan making progress incremental then ask for tests for that step and keep hitch hiking in incremental steps.
Yes and this also helps to highlight which problems I should be solving and which problems Claude code can solve. Like tbh building efficient data structures is not Claude’s thing he seems happy to just hack together some nonsense that leads to spaghetti being shot into all corners of your repo. But by iteratively building up plans and todo lists i find Claude is able to resist the temptation to hack everything all at once to solve the immediate problem in front of his face.
After getting lazy and relying on it too much and getting burned. I now try use it as a replacement for typing I have a similar workflow to you. I just go through the plan together, and review what’s happened every few steps and the correct. It’s actually made me a better PR reviewer, I’ve noticed :)
I'm actually building something to fix this.
The biggest bottleneck for background agents is code review.
I'm building a tool that can give the first pass so the result of the background agent isn't garbage most of the time.
you just rick rolled me!
I am so sorry, it was not intentional at all. I had asked Claude Code to keep a placeholder video. I did not even know what "rick rolling" means. Just searched and understood what I have done!
Apparently this is typical to LLMs: https://chatgpt.com/s/t_687a285d9f708191a884d2ad39ddcb53
I mean, it's probably the most linked YouTube video by a factor of 100x so it makes sense for it to be hardcoded in the model.
Looking at ChatInterface.tsx/handleSendMessage in https://github.com/pixlie/Pixlie/commit/3c0bd23ff16c0fcdac80..., I'd have rejected this if it came up in a PR and would not consider this production quality software.
Rewriting it to something sane would be harder and more time consuming than just writing a decent implementation upfront.
This may not be the best example, but it’s worth considering: if the code is never meant to be reviewed by humans, and it’s an ephemeral implantation detail that will only ever be read & edited by machines, do certain traditional measures of software quality even matter anymore inside the module boundary?
If no one checked it, no one is accountable. Is that how you do business? Or is that how only a few corporations are able to do business?
This sounds like a terrible situation to walk into when the pager goes off.
"Oh hey <boss>, can you update the status page to say we can't really understand the code and don't have an eta but but we're trying to talk the ai into correcting whatever the issue is?"
This is an excellent point and one that I am chasing. I do not want software (at least ones I produce) to be inferior than what I would hand code. LLMs give me huge velocity but I am still learning where to put guardrails so it keeps quality to what I would do myself.
Now the critical point: what if my own quality is inferior to many others. I think we will have this issue a lot since LLMs can generated code at 10x speeds. The quality will be what the human operator sets. Or, maybe, more and more tools will adapt best practices as guardrails and do not give humans much option to steer away from them.
'Tools with guardrails' are common. Wordpress, RAD, low-/no-code and so on. A lot of enterprise software is produced in a cycle where software interviews middle managers and writes code that then generates code that becomes part of a system.
Reinventing this space but make it slow and expensive seems like it's not a serious business idea. I believe the business idea behind coding LLM SaaS is actually about looking into other corporations and seeing what they do and how.
What would you change in that function?
It does a lot under a rather weird name. The fetch should be behind a client facade elsewhere, as should the response handler. I'd also shorten the name.
Probably more that would irk me if I looked closely.
You’re missing the point. This is not production quality software. This is mass produced software. You aren’t supposed to look at all the internals, just the really critical parts.
This is a very critical point. I think code generated by LLMs will lower code generation cost, that means mass produced software.
But: if we reduce cost, then can we not add deterministic guardrails in software that are also maintained at LLM speed and cost? This is pretty much what I am trying to understand. Choice of Rust/TypeScript in my projects in very intentional and you may see why.
So, according to you, it's like Wordpress all over again and a lot of work incoming for people that can actually build and maintain software.
Edit: And more critically, a lot of work incoming for people that can teach software development.
Yes. Someone will need to manage the 10-100x mess of LLM spaghetti codebases after they find product-market fit.
This article reads like an ad for the author’s product.
It definitely is.
But, hey; it worked on me. I'm going to try it, since I've been looking for exactly this.
I knew I was being sold to and I couldn't stop reading and signing up anyways. It's an effective ad.
Good. Advertising is valuable and important. It's how people learn about new products and services. Articles like this are part of why I come to HN. And I wanted something like Terragon for ages. It's an obvious idea, clearly where the industry needs to go, and if it's well implemented then I'd be happy to use it. OpenAI has been trying the same thing with Codex.
Of course, the question is price. It's "free during beta".
HN is full of ads masquerading as comments. They usually start with I used [insert tool from link] to create my [insert as for their product]. It was so great!!
There can't really be any other person recommending AI tools other than people with products to sell, can there? Anyone paying $3000 a month to a third party, coding and trying to wrangle this stuff, has got shit to sell for sure.
The one thing we dont see news stories of is how open source has finally crushed that big evil corp because ai-multiplier times many open source contributors is Unlimited Power (tm)
Am I the only person who hasn't found the time, money, permission from work, resources, etc to be able to produce software like this? I'm still at my corporate 9-5 where they're still making the decision on whether we're allowed to use copilot yet.
They gave us copilot and honestly for me it's been a bit of a double-edged sword. I often use it with Claude 4 as model. What I noticed is that the tool is able and eager to make large sweeping modifications to the code base where I would be apprehensive and look for an easier way.
For example: changing the type signatures of all functions in a module to pass along some extra state, a huge amount of work. I ended up reverting the changes and replacing the functionality with thread local storage (well, dynamically scoped variables).
So, definitely not a panacea, but still well worth the money.
No, lots of people are in that situation. Either due to IP or cost concerns. $200/month/developer is a non-trivial expensive for most companies, given that they aren't easily able to monitor developer productivity, and as people keep observing it's not clear that the $200/month price point is sustainable given how easy it is to ramp up usage what would cost $2000/month on the API.
I've paid $20 for 1 month's worth of Claude Pro. So far it seems sufficient if you want to dip your toes into this form of development.
I do stuff after work to keep up to date. Like scrolling hacker news and as my gf says “shitposting on LinkedIn”. But also I’ve been building some apps with Claude code. It’s fun actually. But I also do other stuff like gym and cycling and Russian language learning.
I am in a situation where everyone uses it but is ashamed to admit it. I often spot people secretly glancing at GH copilot chat or Chat GPT.
Its the Ozempic of programming!
Wow, that's actually a great analogy: it's hailed as the 2nd coming to cure all of society's woes, it's trying to use a quick fix instead of making lifestyle changes, it may damage your organs, and if you stop using it you go back to where you were before or worse
i'm in the same boat more or less. I sometimes ask claude for a stand alone class or webcomponent and it does that job pretty well. The attack lawyers have told me they'd burn my house down if i ever give it any of our code.
The local models I've tried can't write software of the kind and quality I require for work. When I'm programming recreationally I'm not really interested in automating it, since I do it for pleasure.
I have yet to meet someone in meatspace who is competent and isn't of the opinion that LLM SaaS are fine for querying out a bit of bash or Python here and there, but not much else. Sometimes people or corporations on the Internet are like 'wow, our entire repo is vibbed, so cool' and when I go look it's as if a bootcamped intern wrote it unsupervised but with a lot more Markdown I'm not going to look at, or a toy or a simple tool gluing some libraries.
Can someone explain to me like I'm from 2020 what these multiple agents are for? Is it something like starting five different Claude Code sessions in parallel to fix five different issues? Or are they collaborating in some capacity?
Five different issues. They don't (yet) collaborate. However, Claude Code does let the model farm out subtasks to "sub agents" which is basically just forking the chat history in order to manage context explosion. Claude often does this to find things in the codebase - it effectively instructs a sub-instance of itself to work out where something is, the sub-conversation terminates with an answer and the main context window only contains the answer. It can also do this in parallel.
So, it's kinda both. Terragon works on separate tasks in parallel, Claude Code farms out subtasks and sometimes also in parallel.
> Is it something like starting five different Claude Code sessions in parallel to fix five different issues?
yes
I tried Terragon on an open source repo I maintain and ended up with a PR: "Showing 11,643 changed files with 2,441,792 additions and 584 deletions." Yikes!
I've been using Claude Code for weeks now and I've found it to be fantastic at writing / debugging tests unit / integ tests that don't require external context. Still needs some guidance of course but it's been a huge productivity improvement.
Unfortunately, I can't say the same for other types of tests like E2E tests. It makes sense why: CC doesn't have access to all of the context to determine what's going wrong. During an E2E test suite run it can't pause to look into the console, view whats on the page, look at the backend logs, etc. I tried writing some utilities to snapshot and log the state of parts of the page when a test failed and this did help a bit, but not enough to get the same productivity boost for other types of tests.
Has anyone had any luck with this? Any strategies to share?
Whats the benefit of this over the Claude code github action? https://github.com/anthropics/claude-code-action
It's already great at spinning up 5+ agents working on different PRs, triggered by just @mentioning claude on any github issue.
Noise? Iterating in PRs - if humans are involved.
That said, Terragon here (which is akin to Codex and Jules) are often "too autonomous" for my taste. Human in the loop is made more difficult -- commenting may not be enough: I can't edit the code from those because the workspace is so ephemeral and/or remote.
I guess this uses your subscription, when the GH action doesn't. That alone is a significant benefit, as I've turned the action off due to cost.
I think the GH action can also now use your subscription https://x.com/alexalbert__/status/1943332121814405412
Aw man, they say it's available but then the instructions refer to "getting the API key from the console". I'll play around with the installation command again, thank you!
EDIT: The command does it now, thanks! I tried if a few weeks ago and it didn't, so this is great.
Are all these Claude Code articles adverts for people building wrappers around Claude Code? I'm starting to get quite sceptical about how good it actually is
Claude code is very good, no idea about the wrapper and don’t really care. The “hidden” adverts do get on my nerves though
I disagree. It's pricy and does crap code unless precisely instructed. Maybe because I was prompting in non-english langauge
Don't do that. Quality of all models tanks in non-English unfortunately. It's fine for chat and translation tasks but their cognitive capacities does reduce a lot.
One of the next features I'm expecting wrappers to add on top is auto-translation. In many work contexts it makes more sense to translate what the user said to English, process that and translate the answer back than ask the model to speak the language natively.
If I may, LLMs have no cognitive capacities. Their corpora in English are simply bigger.
Oh I wouldn’t try anything other than English with Claude. For the quality, it depends on many things, the context given, the quality of the code given as context, style, etc.
I easily double my output with very little effect on quality. It does require constant supervision though, I don’t see AI producing professional quality code by any other mean than human supervision, for a looooong time.
This is the dream, but we are not yet there. For any reasonable complex code base, CC needs guidance, like avoiding duplicate code/use library function. Write more tests. Sometimes there is a much more elegant solution for a problem. Claude will find it, but you need to hint and ask.
If people use Claude without a critical eye, our code bases will grow immensely.
> If people use Claude without a critical eye, our code bases will grow immensely.
Sounds like the baseline for programming in teams. People are more likely to write their own helpers or install dependencies than knowing what's already available in the repository.
With most of the async Claude code systems I’ve tried you’d do iterations like this as part of code review.
And fwiw I use cc with a complex code base and an async workflow like the author describes daily.
The problem with working at the PR level is it massively sacrifices speed if you need to make small tweaks yourself interspersed with CC’s work, which seems to be the best workflow atm, at least for nontrivial changes.
Local git worktrees are way faster.
Yeah. I don’t think it’s an all or nothing thing. Sometimes a change is best done all through async iterations, sometimes you want to be interactive and sometimes agent wrangling is just not worth it.
The painful thing for me right now is setting up the agent instructions, permissions and sandboxes varies for each of those use cases and the tooling around that is very rudimentary right now.
Claudia helps regarding managment and session mental switching.
I added a web server ontop, so I can use Claudia from my phone now: https://github.com/getAsterisk/claudia/pull/216
When I work side-by-side with Claude to fix bugs or ask it to fix problems. It constantly makes mistakes, like the other day it got the concept of batchItemFailures on an sqs queue completely wrong.
How are people just firing them off to build stuff with any confidence?
https://ghuntley.com/stdlib/ applies to claude as well.
AI won't magically know your codebase unless it is pretty vanilla - but then you teach it. If it makes a mistake, teach it by adding a rule.
You have to confine the output space or else you quickly get whatever.
One might argue it should know others codebases though. The fact it wouldn’t know function signatures for AWS SDK isn’t hugely promising and reflects my own experiences using LLMs
It's just arbitrage.
You aren't a programmer. The basic nature of your work is not serving users with better software, but buying $1000 of currency for $200.
Just as a tractor operator isn't truly a farmer if he doesn't engage in heavy manual labor. Who does he think he is, traversing the fields with a machine that merely converts diesel into grain?
This narrative has been pushed hard lately that $200 a month is a “bargain” because it’s “worth” $1000. If that’s true, then these companies will be running out of venture capital money very soon. Then their only recourse will be to sell themselves or insert ads into LLM responses then this will be a total shit show.
It’s easy to see how trivial it will be to have an adwords type operation where output tokens are replaced with paid values.
For example:
this_varable_is_sponsored_by_starbucks = 42
That would be a terrible revenue model because no one's going to read code anymore, they're going to just keep piling on agents until the vibes align
How much does it seem like this will be affected by the recent headline saying that Max rate-limits are getting shadow-tightened?
It's pretty clear to me that every member Big Token is converging on practically identical models and they're now just competing on compute efficiency based on their particular capital situation. So in the short-term products like Terragon might be nerfed. But in a year or two (2027!) it's hard to imagine OpenAI, Google, Anthropic, Mistral, and so on not all having terminal modes with automatic background agent creation and separate Git branching. At that point it's a race to the bottom for consumer prices.
This sounds fantastic. I found myself nodding along the entire time as the author's experience matched mine. I'm really surprised it isn't more popular here on HN; I can only think that most people haven't really used Claude Max yet.
How are people using Claude to use this much API? I use it to write the occasional bit and the free one seems enough - I've only once been told to come back tomorrow.
write any project with a lot of math and memory shuffling and the structure will generally start eating lots of tokens.
write any project that has a lot of interactive 'dialogues', or exacting and detailed comments, eats a lot of tokens.
My record for tapping out the Claude Max API quickly was sprint-coding a poker solver and accompanying web front end w/ Opus. The backend had a lot of gpgpu stuff going on , and the front end was extremely verbose w/ a wordy ui/ux.
So you write the whole thing using Claude? Any walkthroughs/YouTube things to show how to do this? Not sure I want to, but I'm curious
Pro was not sufficient when i started to: Use git worktrees, type every command in plain english in CC.
For example “commit and push”
How good is it at writing commit messages out of interest?
Not great by default, it just lists the files that were added or modified and "signs" itself at the bottom. I see no value in its commit messages compared to just scrolling down and seeing the actual diff.
You can make it somewhat better by adding instructions in CLAUDE.md, but I did notice those instructions getting ignored from time to time unless you "remind it".
See for yourself: https://github.com/search?utf8=%E2%9C%93&q=%F0%9F%A4%96+Gene...
I have a line in my claude file that says “**Never** list files in git commit messages.’ due to Claude’s propensity to do that. So funny.
I need an open-source version of this. I’m not exactly comfortable exposing all of my code to some external service.
If you watch the following 'til the end https://youtu.be/bUBF5V6oDKw (from the AI Engineer conf, by the creator of Docker), you can I think find something satisfactory within those lines.
Did you link to a video instead of https://github.com/dagger/container-use (Apache 2)? I mean, as fun as it was to watch someone type into GhosTTY on stage ...
Anyway, lightly discussed at the time https://news.ycombinator.com/item?id=44193933
There's a lot of arguments being made here that background agents aren't effective because code review is the bottleneck.
I built a tool that understands the codebase and can give the agent the first pass. So you don't need to spend mental bandwidth to review mountains of code.
If anyone is open to trying - email is in profile
Given the number of existing companies trying that stunt, you'd need to provide more insight into your moat for anyone to be able to judge if they are "open to trying"
I still don’t get how people blow so much money per day if they were just on API usage. My tasks at work are either so trivial that it’s easier and quicker to just do them the old fashioned way, or complex enough that having to steer, correct, and edit the LLMs output is more effort than just writing it myself. What I mainly use LLMs for is “Run a `git diff master` and summarize the changes for a PR description” to save some time there.
I guess a lot of folks are vibe coding 90%+ their side hustles to the point they require 5 instances of CC, each running 8 subagents?
[dead]
[dead]
Claude Code eating up hundreds of dollars of API credits a day doesn't make $200/month "cheap," it just means that Anthropic has figured out an effective way to talk you into getting ripped off.
I tried Claude Code with my custom serverless BaaS https://saasufy.com/ - The results were excellent.
I'm hoping this will commodify the BaaS sector which is good news if you're competing on efficiency and functionality as opposed to relying on network effects.