> What I found interesting is how it forced me to think differently about the development process itself. Instead of jumping straight into code, I found myself spending more time articulating what I actually wanted to build and high level software architectural choices.
This is what I already do with Claude Code. Case in point, I spent 2.5 hours yesterday planning a new feature - first working with an agent to build out the plan, then 4 cycles of having that agent spit out a prompt for another agent to critique the plan and integrate the feedback.
In the end, once I got a clean bill of health on the plan from the “crusty-senior-architect” agent, I had Claude build it - took 12 minutes.
Two passes of the senior-architect and crusty-senior-architect debating how good the code quality was / fixing a few minor issues and the exercise was complete. The new feature worked flawlessly. It took a shade over 3 hours to implement what would have taken me 2 days by myself.
I have been doing this workflow a while, but Claude Code released Agents yesterday (/agents) and I highly recommend them. You can define an agent on the basis of another agent, so crusty-architect is a clone of my senior-architect but it’s never happy unless code was super simple, maintainable, and uses well established patterns. The debates between the two remind me of sitting in conf rooms hashing an issue out with a good team.
I've been attempting to do this kind of thing manually w/ mcp - took a look at "claude swarm" https://github.com/parruda/claude-swarm - but in the short time I spent on it I wasn't having much success - admittedly I probably went a little too far into the "build an entire org chart of agents" territory
[EDIT]: looks like I should be paying attention to the changelog on the gh repo instead of the release notes
[EDIT 2]: so far this seems to suffer from the same problem I had in my own attempts which is that I need to specifically tell it to use an agent when I would really like it to just figure that out on its own
like if I created an agent called "code-reviewer" and then I say - "review this code" ... use the agent!
Roo Code has had Orchestrator mode doing this for a while with your models of choice. And you can tweak the modes or add new ones.
What I have noticed is the forcing function of needing to think through technical and business considerations of ones work up front, which can be tedious if you are the type that likes to jump in and hack at it.
For many types of coding needs, that is likely the smarter and ultimately more efficient approach. Measure twice, cut once.
What I have not yet figured out is how to reduce the friction in the UX of that process to make it more enjoyable. Perhaps sprinkling in some dopamine triggering gamification to answering questions.
You planned and wrote a feature yesterday that would have taken yourself 2 whole days? And you already got it reviewed and deployed it and know that 'it works flawlessly'?
....
That reminds me of when my manager (a very smart, very AI-bullish ex-IC) told us about how he used AI to implement a feature over the weekend and all it took him was 20 mins. It sounds absolutely magical to me and I make a note to use AI more. I then go to review the PR, and of course there are multiple bugs and unintended side-effects in the code. Oh and there are like 8 commits spread over a 60 hour window... I manually spin up a PR which accomplishes the same thing properly... takes me 30mins.
This sounds like a positive outcome? A manager built a proof-of-concept of a feature that clearly laid out and fulfilled the basic requirements, and an engineer took 30 mins to rewrite it once it's been specified.
How long does it typically take to spec something out? I'd say more than 20 mins, and typical artifacts to define requirements are much lossier than actual code - even if that code is buggy and sloppy.
What was claimed was that a complete feature was built in record time with AI. What was actually built was a useless and buggy piece of junk that wasted reviewer time and was ultimately thrown out, and it took far longer than claimed.
There were no useful insights or speed up coming out of this code. I implemented the feature from scratch in 30 mins - because it was actually quite easy to do manually (<100 loc).
This seems more of a process problem than a tooling problem. Without specs on what the feature was, I would be inclined to say you manager had a lapse in his "smartness", there was a lot of miscommunication on what was happening, or you are being overly critical over something that "wasted 30 minutes of your time". Additionally, this seems like a crapshoot work environment...there seems to be resentment for the manager using AI to build a feature that had bugs/didn't work...whereas ideally you two sit down and talk it out and see how it could be managed better next time?
Not at all, there is no resentment - that's your imagination. There is nothing about what I described that indicates that it's a bad work environment - I quite like it.
You're bringing up various completely unrelated factors seemingly as a way of avoiding the obvious point of the anecdotal story - that AI for coding just isn't that great (yet).
> This long (hopefully substantial) blog post was written collaboratively, with Kiro handling the bulk of the content generation while I focused on steering the direction.
So Kiro wrote whatever Kiro "decided", better said, guessed, what to write about and did most of the "content generation" - a weird but fitting term to use by a machine in writing a fake human blog. And the human kind of "directed it", but we dont really know for sure, because language is our primary interface and an author should be able to express their thoughts without using a machine?
I'd happier if the author shared their actual experience in writing the software with this tool.
Why does it matter, as long as the output is of high quality? E.g., a Spielberg directed movie indicates a level of quality, even if Spielberg didn't do everything himself.
The words-to-thoughts ratio is way too high, it reads like a elementary school book report, it’s way too long for how dry it is, I could go on but these are just some of my initial thoughts while reading the article. Also, knowing it is mostly written with AI, how do I know if details are real or made up? There’s a reason you are reading my comment: it expressed thoughts or an image that you found captivating. Being able to write well is a privileged skill that improves your communication, ability to express ideas, your humor; the things that make you an interesting person. You should not be outsourcing your voice to AI. Also Spielberg wasn’t writing an article - he was directing a movie.
So we can debate the quality of this particular article, but in general if an author closely instructs the ai what to write, edits it, and uses the ai to express his thoughts faster and better, I have no problem with the use of ai as a tool to write better.
You’re making assumptions on the quality of the article just because it’s written with the help of ai, I think that not justified in general.
This article has the tell-tale signs of AI slop. It’s perfectly fine to use AI to write, but if you don’t know how to write well (which you learn by writing a lot and getting feedback), then you are just going to produce slop. And no I’m not making assumptions I’m making observations based on my experience reading AI generated text.
Author here, and I mostly agree with this sentiment. With that said, part of my Kiro spec here, had Kiro read my other blog posts and attempt to emulate my style.
I too found the write-up to be a little dry and long-winded, even though the same critique might be leveled at myself [hopefully to a lesser extent]. I did feel like it was a good step up from the usual AI slop out there, and served to illustrate Kiro's capabilities.
I suppose in this context, I wanted to see whether I could take my reader on a journey long enough for them to be surprised by the AI reveal at the end. I'm not sure whether it achieved that but that was the artistic intent of this specific piece.
It matters as its not about a vague notion of a "level of quality". It's about reading about a personal experience written by an actual person. It's about not insulting the intelligence of one's readers by throwing up a wall of LLM-text and signing oneself, it's about not being intellectually and morally dishonest by kinda mentioning it, but only half-way through the text. The comparison with Spielberg is almost there, but not there yet, as the director does whatever it is the directors do - not outsourcing it to some "Kiro". The right comparison would have been if the AI created the next sequel of E.T., Gremlins or whatever it was that Spielberg became famous for. Who cares? I want new and genuine insights that only another human can create. Not the results of a machine minimising the statistical regression error in order to produce the 100th similar piece of "content" that I have already seen the first 99 times. I have a feeling that none of the ghouls pushing for the AI-generated "content" have ever tryly enjoyed the arts, whether 'popular' or 'noble'. Its about learning something about yourself and the world, trying to graps the author's struggles during the creation. Not about mindless consumption. That's why it matters.
> It's about reading about a personal experience written by an actual person
This seems to be the dividing line in the AI writing debate.
If one cares about the interpersonal connection formed with the author, generally they seem to strongly dislike machine-generated content.
If one cares about the content in isolation, then generally the perceived quality is more important than authorship. "The author is dead" and all that.
Both points are valid IMO. It's okay to dislike things, and it's okay to enjoy things.
> I want new and genuine insights that only another human can create.
This is a good illustration of what I mean: you personally value the connection with the author, and you can't get a human connection when there was never a human to begin with.
If you take a look at the others in the thread who had a positive view of the work, they generally focused on the content instead.
It's more that a human author has had an experience and writes to communicate it to other people. They have objectives that make sense in human contexts. They have learned things, run into issues, solved problems, developed arguments, evolved their understanding, felt emotions, considered and rejected actions, and integrated those recent experiences into a lifetime of experience. Then they order and condense all of that into a form that communicates it effectively to other experiencing beings.
LLMs, in contrast, experience nothing. When they "write", they are not even vaguely approximating what a human writer does.
> I want new and genuine insights that only another human can create.
I’m not sure that’s true in an iron-clad, first-principles way. I think that many of the insights created by humans are combinations/integrations of existing concepts. That category of insight does not seem to require carbon-based reasoning.
I don’t claim that it can be achieved by statistical text generation, but I doubt the typical blog author is creating something that forever will be human-only.
Is it not high quality output when a ghost writer writes someone’s work? Can I use a thesaurus? What about a translator?
As long as the person who is putting their name on the article has provided the ideas and ensures that the output matches their intent, I see no reason to draw arbitrary lines.
Thesaurus: Sure, since you're looking up words but using your own brain to decide whether a word is appropriate or better. Caveat: some authors just throw in words to look smart. Concordedly.
Translator: Only if it's not creative writing and only if the output doesn't need to be accurate / there is margin for error.
Granted, the line is hard to place - see also the other post about where the line for adult content should be. But the thing is, when writing is AI generated, how can you know for sure that all the output was actually read and verified by the author? And text is not that important, but what about code?
The other argument: if an author didn't take the time and effort to write something, why should I take the time and effort to read and understand it? This applies to poorly written content as well. Ironically, a lot of people will already feed this article (and this comment) into an LLM to summarize it, if not to help them form an opinion about it. Summarizing isn't new though, there were tools some years ago to summarize webpages already.
I’m pretty sure Mr Spielberg DOES something. If he does absolutely nothing, I really doubt the phrase “directed movie indicates a level of quality” can have any level of truth in a general case.
Ai is just another tool, it doesn’t do anything by itself, in the hands of a person with a great story it can be a great tool to help write it. In the hands of a bad writer it will probably just create bad content. Ai is a tool controlled by a person.
Do you really care if Spielberg’s team manually edits the movie or uses an AI powered video editing tool? In the end Spielberg is responsible for the end quality of the movie.
Well, you just answered the question yourself, by inadvertently using a good example.
You see, a movie is fictional work, but a blog article most likely isn't (or shouldn't). In this case, I am reading the article because I want to know an objective, fair assessment of Kiro from a human, not random texts generated from an LLM.
I would actually rather read his incomprehendable notes, and then the actual prompt he tried to send to the LLM, then skip reading the resultant generated part. This is actually what is worth reading to me
The output is not of high quality. It is extremely verbose for what it is trying to say, and I found myself skimming it while dealing with
1. The constant whiplash of paragraphs which describe an amazing feature followed by paragraphs which walk it back ("The shift is subtle but significant / But I want to be clear", "Kiro would implement it correctly / That said, it wasn't completely hands-off", "The onboarding assistance was genuinely helpful / However, this is also where I encountered", "It's particularly strong at understanding / However, there are times when");
2. Bland analogies that detract from, rather than enhance, understanding ("It's the difference between being a hands-on manager who needs to check every detail versus setting clear expectations and trusting the process.", "It's like having a very knowledgeable colleague who..."); and
3. literal content-free filler ("Here's where it got interesting", "Don't let perfect be the enemy of good", "Most importantly / More importantly"), etc etc.
Kiro is a new agentic IDE which puts much more of a focus on detailed, upfront specification than competitors like Cursor. That's great. Just write about that.
Bullshit is what we call it when a speaker wants to look good saying something but doesn't care if what they say is true. It's a focus on form over content.
You’re making assumptions that aren’t grounded in facts, you do not know if they didn’t care, and if the blog contains fiction. He may have checked and corrected every single line created by the ai.
I was only responding to a philosophical question about what's problematic about putting someone else's work under your own name, not making an accusation.
To respond to you though, yes that is possible, but anyone who put in that level is work is highly unlikely to credit the AI as a writer.
And anyway, I had (and have) a choice. Plenty of people are still using their own words and have (or are building) confidence in their own voice, so those are the people I seek out because by giving them support I help make the world I want to live in
I think a big reason why Claude Code is winning is because it’s such a thin wrapper over a very strong base model which is why people are afraid of comparing it directly.
All these IDE integrations and GUIs and super complex system prompts etc are only bloating all these other solutions with extra complexity, so comparing something inherently different becomes also harder.
Agree. I stopped reading after the blurb below because it tells me this person has not actually even used Copilot or Cursor to a serious degree. This is an AI-written sentence that seems fine, but is actually complete nonsense.
> Each tool has carved out its own niche in the development workflow: Copilot excels at enhancing your typing speed with intelligent code completion, Cursor at debugging and helping you implement discrete tasks well, and recently pushing more into agentic territory.
Cursor's autocomplete blows Copilot's out of the water. And both Copilot and Cursor have pretty capable agents. Plus, Claude Code isn't even mentioned here.
This blog post is a Kiro advertisement, not a serious comparative analysis.
Outside of its excellent capabilities, the thing I most love about Claude Code is that I can run it in my containers. I don’t want Cursor or other bloated, dependency-ridden, high-profile security targets on my desktop.
I tried out Kiro last week on quite a gnarly date time parsing problem. I had started with a two hundred-ish word prompt and a few bits of code examples for context to describe the problem. Similar to OP it forced me to stop and think more clearly about the problem I was trying to solve and in the end left my jaw on the floor as I saw it work through the task list.
I think only early bit of feedback I had was in that my tasks were also writing a lot of tests, and if the feedback loop to getting test results was neater this would be insanely powerful. Something like a sandboxed terminal, I am less keen on a YOLO mode and had to keep authorising the terminal to run.
This sort of comment always fascinates me. Having a machine do the last leaps for you is a time saver I guess, but I often wonder whether the real thing people are discovering again is that sitting down and really thinking about the problem you're trying to solve before writing some code results in better solutions when you get to it.
its not that i havent spent time thinking about it - i at least still my thinking first mostly on paper.
the LLM however asks me clarifying questions that i wouldnt have thought about myself. the thinking is a step or two deeper than it was before, if the LLM comes up with good questions
I kind of hate the implications of it, but if HN (or someone else) wanted to add value, they could show one-line sentiment analyses of the comments in the HN articles so you can decide what's what without even clicking.
I wanted a tiny helper tool to display my global keyboard shortcuts for me on macOS. I gave Kiro a short spec and some TypeScript describing the schema of the input data.
It wrote around 5000 LOC including tests and they... worked. It didn't look as nice as I would have liked, but I wasn't able to break it. However, 5000 lines was way too much code for such a simple task, the solution was over-engineered along every possible axis. I was able to (manually) get it down to ~800LOC without losing any important functionality.
> I was able to (manually) get it down to ~800LOC without losing any important functionality.
This is funny. Why would you a) care how many LOC it generated and b) bother injecting tedious, manual process into something otherwise fully automated?
What year is it? Back in 2000 or before, the same arguments were made about webpages made in Dreamweaver and Frontpage. Shortly after there was a big push towards making the web faster and more efficient, which included stepping away from web page builders and building tools that optimized and minified all aspects of a webpage.
Then, we ended up bundling 50MB of minified frameworks in every page load. Maybe the next part of the cycle will be for optimizing this aspect of LLMs, maybe even to be able to fit more meaningful code in the context windows of these very LLMs.
I care about the complexity because I want/need to maintain the code down the line. I find it much easier to maintain shorter, simpler code than long, complex code.
Also because it was an experiment. I wanted to see how it would do and how reasonable the code it wrote was.
I'm sure somebody is going to point out that it was written by AI and is a toy, therefore it can be maintained by AI. I share your desire to have human maintainable code, but I imagine one of the goals of AI written code is to allow the AI to manage it, end-to-end.
I am just gonna say it. This is not something Kiro came up with. People were already using this workflow. Perhaps they should’ve added more features instead of spending time making promo videos of themselves. I fail to see any add value here especially considering it’s Amazon. Sonnet 4 is effectively unlimited for many MAX users so giving that away to work out their list of bugs is a non-starter.
at this point i think we ought to start having a tag in the submission title for when the submission is (primarily) llm-generated, like we do for video/pdf/nsfw
I use https://sketch.dev, which seems to be approximately the same kind of thing. I find it is very organized, it does not need nearly as much babysitting as this guy is talking about. Perhaps a difference is the "todo read" and "todo write" tools, which it uses to break the main goal into achievable tasks, and then judge itself on if it has finished correctly or not. Also, Sketch is easy-going with corrections: if you tell it that it went the wrong direction, it quickly adjusts and follows your direction.
Anyway, two big thumbs up for agentic dev workflows from me, but only so far with Sketch.
Kiro’s main advantage is Amazon is paying for my LLM usage instead of me.
For the most part it’s unlimited right now. Vs Code’s Copilot Agent mode is basically the same thing , tell it to write a list of tasks , but I have to pay for it.
I’m much happier with both of these options, both are much cheaper than Claude Code.
IMO the real race is to get LLM cost down. Someone smarter than me is going to figure out how to run a top LLM model for next to nothing.
This person will be a billionaire. Nvidia and AMD are probably already working on it. I want Deepseek running on a 100$ computer that uses a nominal amount of power.
My thoughts exactly. Inference should be dirt cheap for LLMs to truly become powerful.
It's similar to how computing used to be restricted to mega corps 100 years ago, but today, a smartphone has more computing power than any old age mainframe. Today we need Elon Musk to buy 5 million GPUs to train a model. Tomorrow, we should be able to train a top of the line model using a budget RTX card.
Tbh, if the model is small enough you can train locally.
I don't need my code assistant to be an expert on Greek myths. The future is probably highly specialized mini llms. I might train a model to code my way.
I'm not that smart enough to figure this out, but the solution can't be to just brute force training with more gpus.
Spec based development is a game changer for people with non-coding background to work on side-projects. I’m using this to fork a design flow I use for analog/RFIC design, and I can finally mend together open-source CAD tools to attune to the design flow, that I’d otherwise use in segregation.
> What I found interesting is how it forced me to think differently about the development process itself.
> Instead of jumping straight into code, I found myself spending more time articulating what I actually wanted to build and high level software architectural choices.
I don't want to sound rude, but isn't this is something that happens to you with experience after some years?
How can you even be a senior developer without "spending more time articulating what I actually wanted to build and high level software architectural choices"?
I believe that's why some software engineers dread this new future, because it's no longer software but rather project management and code review, pretty much all day
I personally find code review more exhausting than code writing, and that goes 50x for when I'm reviewing code from an intern because they are trying to cheat me all the time. And I have always hated PM stuff, on both sides of that relationship
Been using it for a while, now all other tools feel lacking. The multistep process with manual refinement options after every step just works so much better for keeping LLMs in line. The only times I’ve used anything but kiro since, has been when the Sonnet 4.0 was once again overloaded (tried it with 3.7 for a bit, but it went way off the rails of kiro’s tasks structure every time).
I bogged me that it says that he spend loads of time upfront planning but not with cursor. You can do the same with cursor, and on my experience, they are about the same.
When they reach certain level of complexity of a medium sized project they start messing up.
I agree. It’s just not a fair comparison. But then the article is written by Kino with the author “sprinkling in” sentences, so I guess I can’t really expect much.
Super interesting that no one here has mentioned replit.
I found the experience of using Kiro and replit really similar with one important difference: replit mostly worked.
Kiro tore off and wrote tonnes of code and tests. It asked me to peck at approval requests (one thing I liked was the regexp type trust this tool to do this requests) it spent half a day creating an app to do what I asked... and it was incomprehensible bullshit. I couldn't do anything with the project and I have not touched it since.
Replit was a bit more interactive, but pretty autonomous, and it got me to 90% of the solution and then stalled out - wouldn't correct some of the problems I identified to it. About 2 hrs with Cursor sorted that out though.
I did use Cursor to do it "AI assisted" and that took about the same amount of time. The advantage is that I really do know whats going on in the code base, but the Replit + Cursor solution is actually better in the sense that it looks better, and works better because the agent did some bits a bit more nicely than I did with Cursor - so I got those ideas for free.
Anyway :
Hand coding = a walk through the wilderness
Cursor = motor cross scrambler bike up the mountain
Replit = a helicopter ride to somewhere higher up the mountain selected at random that you didn't know about but now you have to get to the peak by yourself buddy, good luck
Kiro = you are blindfolded in a container of some sort and it's moving.
It's strange because I used replit in an online container and then had to download the code to carry on editing it - so I don't know how it managed to do that! Anyway, it sounds like they were crazy giving the keys to prod...
I'm long on no-AI becoming a label of quality, a badge of honor, as people are more exposed to the slop LLMs forces upon them.
I want prose, music, photography and movies made by humans, for humans.
I want to use software built by developers. The last part in particular because I've seen my share of bug riddled code spewed by the LLM du jour, I'm convinced our industry is vibecoding itself into the abyss.
Very good write up. This is indeed a qualitative change in the work of software engineering.
What they’re describing can also be done with Claude Code, and it’s way too broad in scope to get any benefit at all from approving code before it’s written. These tools are the way for now.
Cursor is in my opinion not geared for this level of hands-off.
From the article:
> What I found interesting is how it forced me to think differently about the development process itself. Instead of jumping straight into code, I found myself spending more time articulating what I actually wanted to build and high level software architectural choices.
This is what I already do with Claude Code. Case in point, I spent 2.5 hours yesterday planning a new feature - first working with an agent to build out the plan, then 4 cycles of having that agent spit out a prompt for another agent to critique the plan and integrate the feedback.
In the end, once I got a clean bill of health on the plan from the “crusty-senior-architect” agent, I had Claude build it - took 12 minutes.
Two passes of the senior-architect and crusty-senior-architect debating how good the code quality was / fixing a few minor issues and the exercise was complete. The new feature worked flawlessly. It took a shade over 3 hours to implement what would have taken me 2 days by myself.
I have been doing this workflow a while, but Claude Code released Agents yesterday (/agents) and I highly recommend them. You can define an agent on the basis of another agent, so crusty-architect is a clone of my senior-architect but it’s never happy unless code was super simple, maintainable, and uses well established patterns. The debates between the two remind me of sitting in conf rooms hashing an issue out with a good team.
I love how I am learning about a new claude code feature in a comment on HN - nowhere to be found on their release notes https://docs.anthropic.com/en/release-notes/claude-code
Thanks for the tip!
I've been attempting to do this kind of thing manually w/ mcp - took a look at "claude swarm" https://github.com/parruda/claude-swarm - but in the short time I spent on it I wasn't having much success - admittedly I probably went a little too far into the "build an entire org chart of agents" territory
[EDIT]: looks like I should be paying attention to the changelog on the gh repo instead of the release notes
https://github.com/anthropics/claude-code/blob/main/CHANGELO...
[EDIT 2]: so far this seems to suffer from the same problem I had in my own attempts which is that I need to specifically tell it to use an agent when I would really like it to just figure that out on its own
like if I created an agent called "code-reviewer" and then I say - "review this code" ... use the agent!
Roo Code has had Orchestrator mode doing this for a while with your models of choice. And you can tweak the modes or add new ones.
What I have noticed is the forcing function of needing to think through technical and business considerations of ones work up front, which can be tedious if you are the type that likes to jump in and hack at it.
For many types of coding needs, that is likely the smarter and ultimately more efficient approach. Measure twice, cut once.
What I have not yet figured out is how to reduce the friction in the UX of that process to make it more enjoyable. Perhaps sprinkling in some dopamine triggering gamification to answering questions.
You planned and wrote a feature yesterday that would have taken yourself 2 whole days? And you already got it reviewed and deployed it and know that 'it works flawlessly'?
....
That reminds me of when my manager (a very smart, very AI-bullish ex-IC) told us about how he used AI to implement a feature over the weekend and all it took him was 20 mins. It sounds absolutely magical to me and I make a note to use AI more. I then go to review the PR, and of course there are multiple bugs and unintended side-effects in the code. Oh and there are like 8 commits spread over a 60 hour window... I manually spin up a PR which accomplishes the same thing properly... takes me 30mins.
This sounds like a positive outcome? A manager built a proof-of-concept of a feature that clearly laid out and fulfilled the basic requirements, and an engineer took 30 mins to rewrite it once it's been specified.
How long does it typically take to spec something out? I'd say more than 20 mins, and typical artifacts to define requirements are much lossier than actual code - even if that code is buggy and sloppy.
Not at all.
What was claimed was that a complete feature was built in record time with AI. What was actually built was a useless and buggy piece of junk that wasted reviewer time and was ultimately thrown out, and it took far longer than claimed.
There were no useful insights or speed up coming out of this code. I implemented the feature from scratch in 30 mins - because it was actually quite easy to do manually (<100 loc).
This seems more of a process problem than a tooling problem. Without specs on what the feature was, I would be inclined to say you manager had a lapse in his "smartness", there was a lot of miscommunication on what was happening, or you are being overly critical over something that "wasted 30 minutes of your time". Additionally, this seems like a crapshoot work environment...there seems to be resentment for the manager using AI to build a feature that had bugs/didn't work...whereas ideally you two sit down and talk it out and see how it could be managed better next time?
Not at all, there is no resentment - that's your imagination. There is nothing about what I described that indicates that it's a bad work environment - I quite like it.
You're bringing up various completely unrelated factors seemingly as a way of avoiding the obvious point of the anecdotal story - that AI for coding just isn't that great (yet).
Would you mind sharing the prompts you use for your subagents? It sounds very interesting!
How exactly to you "create an agent" with the personalities you are talking about?
/agents
The parent commenter had agents with personalities before the release of the agents feature in Claude Code, that's why I was asking.
Prompt it to behave in a certain way?
> This long (hopefully substantial) blog post was written collaboratively, with Kiro handling the bulk of the content generation while I focused on steering the direction.
So Kiro wrote whatever Kiro "decided", better said, guessed, what to write about and did most of the "content generation" - a weird but fitting term to use by a machine in writing a fake human blog. And the human kind of "directed it", but we dont really know for sure, because language is our primary interface and an author should be able to express their thoughts without using a machine?
I'd happier if the author shared their actual experience in writing the software with this tool.
Why does it matter, as long as the output is of high quality? E.g., a Spielberg directed movie indicates a level of quality, even if Spielberg didn't do everything himself.
The words-to-thoughts ratio is way too high, it reads like a elementary school book report, it’s way too long for how dry it is, I could go on but these are just some of my initial thoughts while reading the article. Also, knowing it is mostly written with AI, how do I know if details are real or made up? There’s a reason you are reading my comment: it expressed thoughts or an image that you found captivating. Being able to write well is a privileged skill that improves your communication, ability to express ideas, your humor; the things that make you an interesting person. You should not be outsourcing your voice to AI. Also Spielberg wasn’t writing an article - he was directing a movie.
So we can debate the quality of this particular article, but in general if an author closely instructs the ai what to write, edits it, and uses the ai to express his thoughts faster and better, I have no problem with the use of ai as a tool to write better.
You’re making assumptions on the quality of the article just because it’s written with the help of ai, I think that not justified in general.
This article has the tell-tale signs of AI slop. It’s perfectly fine to use AI to write, but if you don’t know how to write well (which you learn by writing a lot and getting feedback), then you are just going to produce slop. And no I’m not making assumptions I’m making observations based on my experience reading AI generated text.
I personally really enjoyed that writing style and didn't notice any "slop".
I appreciate that!
Author here, and I mostly agree with this sentiment. With that said, part of my Kiro spec here, had Kiro read my other blog posts and attempt to emulate my style.
I too found the write-up to be a little dry and long-winded, even though the same critique might be leveled at myself [hopefully to a lesser extent]. I did feel like it was a good step up from the usual AI slop out there, and served to illustrate Kiro's capabilities.
I suppose in this context, I wanted to see whether I could take my reader on a journey long enough for them to be surprised by the AI reveal at the end. I'm not sure whether it achieved that but that was the artistic intent of this specific piece.
It matters as its not about a vague notion of a "level of quality". It's about reading about a personal experience written by an actual person. It's about not insulting the intelligence of one's readers by throwing up a wall of LLM-text and signing oneself, it's about not being intellectually and morally dishonest by kinda mentioning it, but only half-way through the text. The comparison with Spielberg is almost there, but not there yet, as the director does whatever it is the directors do - not outsourcing it to some "Kiro". The right comparison would have been if the AI created the next sequel of E.T., Gremlins or whatever it was that Spielberg became famous for. Who cares? I want new and genuine insights that only another human can create. Not the results of a machine minimising the statistical regression error in order to produce the 100th similar piece of "content" that I have already seen the first 99 times. I have a feeling that none of the ghouls pushing for the AI-generated "content" have ever tryly enjoyed the arts, whether 'popular' or 'noble'. Its about learning something about yourself and the world, trying to graps the author's struggles during the creation. Not about mindless consumption. That's why it matters.
> It's about reading about a personal experience written by an actual person
This seems to be the dividing line in the AI writing debate.
If one cares about the interpersonal connection formed with the author, generally they seem to strongly dislike machine-generated content.
If one cares about the content in isolation, then generally the perceived quality is more important than authorship. "The author is dead" and all that.
Both points are valid IMO. It's okay to dislike things, and it's okay to enjoy things.
> I want new and genuine insights that only another human can create.
This is a good illustration of what I mean: you personally value the connection with the author, and you can't get a human connection when there was never a human to begin with.
If you take a look at the others in the thread who had a positive view of the work, they generally focused on the content instead.
It's more that a human author has had an experience and writes to communicate it to other people. They have objectives that make sense in human contexts. They have learned things, run into issues, solved problems, developed arguments, evolved their understanding, felt emotions, considered and rejected actions, and integrated those recent experiences into a lifetime of experience. Then they order and condense all of that into a form that communicates it effectively to other experiencing beings.
LLMs, in contrast, experience nothing. When they "write", they are not even vaguely approximating what a human writer does.
> I want new and genuine insights that only another human can create.
I’m not sure that’s true in an iron-clad, first-principles way. I think that many of the insights created by humans are combinations/integrations of existing concepts. That category of insight does not seem to require carbon-based reasoning.
I don’t claim that it can be achieved by statistical text generation, but I doubt the typical blog author is creating something that forever will be human-only.
Who knows, perhaps you're right? But in the meanwhile, I don't want a Kiro sharing with me fake experiences a Kiro never had.
It's not high quality output when an opinion based article isn't written by its author. It's fiction.
Where is the line?
Is it not high quality output when a ghost writer writes someone’s work? Can I use a thesaurus? What about a translator?
As long as the person who is putting their name on the article has provided the ideas and ensures that the output matches their intent, I see no reason to draw arbitrary lines.
Thesaurus: Sure, since you're looking up words but using your own brain to decide whether a word is appropriate or better. Caveat: some authors just throw in words to look smart. Concordedly.
Translator: Only if it's not creative writing and only if the output doesn't need to be accurate / there is margin for error.
Granted, the line is hard to place - see also the other post about where the line for adult content should be. But the thing is, when writing is AI generated, how can you know for sure that all the output was actually read and verified by the author? And text is not that important, but what about code?
The other argument: if an author didn't take the time and effort to write something, why should I take the time and effort to read and understand it? This applies to poorly written content as well. Ironically, a lot of people will already feed this article (and this comment) into an LLM to summarize it, if not to help them form an opinion about it. Summarizing isn't new though, there were tools some years ago to summarize webpages already.
Thumbs up to this concise comment, but I'd not even honour the blog author by calling it fiction. Probably more of a hallucination.
It’s objectively not fiction, whatever it is.
If they can't be bothered to write it, I can't bothered to read it.
I’m pretty sure Mr Spielberg DOES something. If he does absolutely nothing, I really doubt the phrase “directed movie indicates a level of quality” can have any level of truth in a general case.
Ai is just another tool, it doesn’t do anything by itself, in the hands of a person with a great story it can be a great tool to help write it. In the hands of a bad writer it will probably just create bad content. Ai is a tool controlled by a person.
Do you really care if Spielberg’s team manually edits the movie or uses an AI powered video editing tool? In the end Spielberg is responsible for the end quality of the movie.
Well, you just answered the question yourself, by inadvertently using a good example.
You see, a movie is fictional work, but a blog article most likely isn't (or shouldn't). In this case, I am reading the article because I want to know an objective, fair assessment of Kiro from a human, not random texts generated from an LLM.
I would actually rather read his incomprehendable notes, and then the actual prompt he tried to send to the LLM, then skip reading the resultant generated part. This is actually what is worth reading to me
The output is not of high quality. It is extremely verbose for what it is trying to say, and I found myself skimming it while dealing with
1. The constant whiplash of paragraphs which describe an amazing feature followed by paragraphs which walk it back ("The shift is subtle but significant / But I want to be clear", "Kiro would implement it correctly / That said, it wasn't completely hands-off", "The onboarding assistance was genuinely helpful / However, this is also where I encountered", "It's particularly strong at understanding / However, there are times when");
2. Bland analogies that detract from, rather than enhance, understanding ("It's the difference between being a hands-on manager who needs to check every detail versus setting clear expectations and trusting the process.", "It's like having a very knowledgeable colleague who..."); and
3. literal content-free filler ("Here's where it got interesting", "Don't let perfect be the enemy of good", "Most importantly / More importantly"), etc etc.
Kiro is a new agentic IDE which puts much more of a focus on detailed, upfront specification than competitors like Cursor. That's great. Just write about that.
It's more like if Spielberg produced a documentary about himself.
Bullshit is what we call it when a speaker wants to look good saying something but doesn't care if what they say is true. It's a focus on form over content.
You’re making assumptions that aren’t grounded in facts, you do not know if they didn’t care, and if the blog contains fiction. He may have checked and corrected every single line created by the ai.
I was only responding to a philosophical question about what's problematic about putting someone else's work under your own name, not making an accusation.
To respond to you though, yes that is possible, but anyone who put in that level is work is highly unlikely to credit the AI as a writer.
And anyway, I had (and have) a choice. Plenty of people are still using their own words and have (or are building) confidence in their own voice, so those are the people I seek out because by giving them support I help make the world I want to live in
Author here. Didn't realize this went a bit viral until today.
This blog post does approximate my sentiments pretty well, although its writing style diverges from my usual style.
Crtl + F -> "Claude Code" -> No Results -> Close tab
Can't really get value out reading this if you don't compare it to the leading coding agent
I think a big reason why Claude Code is winning is because it’s such a thin wrapper over a very strong base model which is why people are afraid of comparing it directly. All these IDE integrations and GUIs and super complex system prompts etc are only bloating all these other solutions with extra complexity, so comparing something inherently different becomes also harder.
Agree. I stopped reading after the blurb below because it tells me this person has not actually even used Copilot or Cursor to a serious degree. This is an AI-written sentence that seems fine, but is actually complete nonsense.
> Each tool has carved out its own niche in the development workflow: Copilot excels at enhancing your typing speed with intelligent code completion, Cursor at debugging and helping you implement discrete tasks well, and recently pushing more into agentic territory.
Cursor's autocomplete blows Copilot's out of the water. And both Copilot and Cursor have pretty capable agents. Plus, Claude Code isn't even mentioned here.
This blog post is a Kiro advertisement, not a serious comparative analysis.
Outside of its excellent capabilities, the thing I most love about Claude Code is that I can run it in my containers. I don’t want Cursor or other bloated, dependency-ridden, high-profile security targets on my desktop.
Kiro uses Claude Sonnet 4.0 if that matters.
I tried out Kiro last week on quite a gnarly date time parsing problem. I had started with a two hundred-ish word prompt and a few bits of code examples for context to describe the problem. Similar to OP it forced me to stop and think more clearly about the problem I was trying to solve and in the end left my jaw on the floor as I saw it work through the task list.
I think only early bit of feedback I had was in that my tasks were also writing a lot of tests, and if the feedback loop to getting test results was neater this would be insanely powerful. Something like a sandboxed terminal, I am less keen on a YOLO mode and had to keep authorising the terminal to run.
This sort of comment always fascinates me. Having a machine do the last leaps for you is a time saver I guess, but I often wonder whether the real thing people are discovering again is that sitting down and really thinking about the problem you're trying to solve before writing some code results in better solutions when you get to it.
its not that i havent spent time thinking about it - i at least still my thinking first mostly on paper.
the LLM however asks me clarifying questions that i wouldnt have thought about myself. the thinking is a step or two deeper than it was before, if the LLM comes up with good questions
They wouldn’t have thought more deeply about it if the model didn’t „tell” them to.
meta: if you use AI to write articles, don’t have them written so that I’m forced to use AI to summarize them
Reading the HN comments instead of the article is the best summarizing hack
I kind of hate the implications of it, but if HN (or someone else) wanted to add value, they could show one-line sentiment analyses of the comments in the HN articles so you can decide what's what without even clicking.
The reason reading comments is so useful is because it's not one summary but a variety of different, unique reactions (and reactions to reactions).
The model I want to train is ME, so a one sentence sentiment analysis offers 0 value to me while a lot of distinct human perspectives are a gold mine.
It's kinda like the difference between being able to look at a single picture of a landscape and being able to walk around it.
Ironically there was a tool just the other day that would read HN articles and summarize them.
Ai decompressor doing their best again. Seriously, though, the article feels too long imho
Better, it should be compulsory for these to lead with a summarized version
I wanted a tiny helper tool to display my global keyboard shortcuts for me on macOS. I gave Kiro a short spec and some TypeScript describing the schema of the input data.
It wrote around 5000 LOC including tests and they... worked. It didn't look as nice as I would have liked, but I wasn't able to break it. However, 5000 lines was way too much code for such a simple task, the solution was over-engineered along every possible axis. I was able to (manually) get it down to ~800LOC without losing any important functionality.
Kiro is from Amazon, so Conway's law in action?
> I was able to (manually) get it down to ~800LOC without losing any important functionality.
This is funny. Why would you a) care how many LOC it generated and b) bother injecting tedious, manual process into something otherwise fully automated?
What year is it? Back in 2000 or before, the same arguments were made about webpages made in Dreamweaver and Frontpage. Shortly after there was a big push towards making the web faster and more efficient, which included stepping away from web page builders and building tools that optimized and minified all aspects of a webpage.
Then, we ended up bundling 50MB of minified frameworks in every page load. Maybe the next part of the cycle will be for optimizing this aspect of LLMs, maybe even to be able to fit more meaningful code in the context windows of these very LLMs.
I care about the complexity because I want/need to maintain the code down the line. I find it much easier to maintain shorter, simpler code than long, complex code.
Also because it was an experiment. I wanted to see how it would do and how reasonable the code it wrote was.
I'm sure somebody is going to point out that it was written by AI and is a toy, therefore it can be maintained by AI. I share your desire to have human maintainable code, but I imagine one of the goals of AI written code is to allow the AI to manage it, end-to-end.
> one of the goals of AI written code is to allow the AI to manage
Another reason to favour shorter code.
I am just gonna say it. This is not something Kiro came up with. People were already using this workflow. Perhaps they should’ve added more features instead of spending time making promo videos of themselves. I fail to see any add value here especially considering it’s Amazon. Sonnet 4 is effectively unlimited for many MAX users so giving that away to work out their list of bugs is a non-starter.
at this point i think we ought to start having a tag in the submission title for when the submission is (primarily) llm-generated, like we do for video/pdf/nsfw
Too late. YouTube music channels already began explicitly tagging videos as "Human made music".
I use https://sketch.dev, which seems to be approximately the same kind of thing. I find it is very organized, it does not need nearly as much babysitting as this guy is talking about. Perhaps a difference is the "todo read" and "todo write" tools, which it uses to break the main goal into achievable tasks, and then judge itself on if it has finished correctly or not. Also, Sketch is easy-going with corrections: if you tell it that it went the wrong direction, it quickly adjusts and follows your direction.
Anyway, two big thumbs up for agentic dev workflows from me, but only so far with Sketch.
Kiro’s main advantage is Amazon is paying for my LLM usage instead of me.
For the most part it’s unlimited right now. Vs Code’s Copilot Agent mode is basically the same thing , tell it to write a list of tasks , but I have to pay for it.
I’m much happier with both of these options, both are much cheaper than Claude Code.
IMO the real race is to get LLM cost down. Someone smarter than me is going to figure out how to run a top LLM model for next to nothing.
This person will be a billionaire. Nvidia and AMD are probably already working on it. I want Deepseek running on a 100$ computer that uses a nominal amount of power.
My thoughts exactly. Inference should be dirt cheap for LLMs to truly become powerful.
It's similar to how computing used to be restricted to mega corps 100 years ago, but today, a smartphone has more computing power than any old age mainframe. Today we need Elon Musk to buy 5 million GPUs to train a model. Tomorrow, we should be able to train a top of the line model using a budget RTX card.
Tbh, if the model is small enough you can train locally.
I don't need my code assistant to be an expert on Greek myths. The future is probably highly specialized mini llms. I might train a model to code my way.
I'm not that smart enough to figure this out, but the solution can't be to just brute force training with more gpus.
There is another answer.
Apart from the UI being more point and click, what does Kiro (or Cursor) add that for example Claude Code's CLI doesn't do?
> With Kiro, I spend more time upfront articulating what I want to build, but then I can step back and let it execute
This sounds like exactly the kind of exercise one does to /init a project with Claude, define tasks/spec etc.
Spec based development is a game changer for people with non-coding background to work on side-projects. I’m using this to fork a design flow I use for analog/RFIC design, and I can finally mend together open-source CAD tools to attune to the design flow, that I’d otherwise use in segregation.
> What I found interesting is how it forced me to think differently about the development process itself.
> Instead of jumping straight into code, I found myself spending more time articulating what I actually wanted to build and high level software architectural choices.
I don't want to sound rude, but isn't this is something that happens to you with experience after some years?
How can you even be a senior developer without "spending more time articulating what I actually wanted to build and high level software architectural choices"?
So with Kiro, you iterate on a spec and a task list as the key activity?
It sounds like a very PM type approach to coding.
Does that mean it fits PM types more than IC dev types?
I believe that's why some software engineers dread this new future, because it's no longer software but rather project management and code review, pretty much all day
I personally find code review more exhausting than code writing, and that goes 50x for when I'm reviewing code from an intern because they are trying to cheat me all the time. And I have always hated PM stuff, on both sides of that relationship
Been using it for a while, now all other tools feel lacking. The multistep process with manual refinement options after every step just works so much better for keeping LLMs in line. The only times I’ve used anything but kiro since, has been when the Sonnet 4.0 was once again overloaded (tried it with 3.7 for a bit, but it went way off the rails of kiro’s tasks structure every time).
I bogged me that it says that he spend loads of time upfront planning but not with cursor. You can do the same with cursor, and on my experience, they are about the same. When they reach certain level of complexity of a medium sized project they start messing up.
I agree. It’s just not a fair comparison. But then the article is written by Kino with the author “sprinkling in” sentences, so I guess I can’t really expect much.
Super interesting that no one here has mentioned replit.
I found the experience of using Kiro and replit really similar with one important difference: replit mostly worked.
Kiro tore off and wrote tonnes of code and tests. It asked me to peck at approval requests (one thing I liked was the regexp type trust this tool to do this requests) it spent half a day creating an app to do what I asked... and it was incomprehensible bullshit. I couldn't do anything with the project and I have not touched it since.
Replit was a bit more interactive, but pretty autonomous, and it got me to 90% of the solution and then stalled out - wouldn't correct some of the problems I identified to it. About 2 hrs with Cursor sorted that out though.
I did use Cursor to do it "AI assisted" and that took about the same amount of time. The advantage is that I really do know whats going on in the code base, but the Replit + Cursor solution is actually better in the sense that it looks better, and works better because the agent did some bits a bit more nicely than I did with Cursor - so I got those ideas for free.
Anyway :
Hand coding = a walk through the wilderness
Cursor = motor cross scrambler bike up the mountain
Replit = a helicopter ride to somewhere higher up the mountain selected at random that you didn't know about but now you have to get to the peak by yourself buddy, good luck
Kiro = you are blindfolded in a container of some sort and it's moving.
relevant: Replit AI deletes entire database during code freeze, then lies about it - https://news.ycombinator.com/item?id=44625119 - July, 2025 (53 comments)
It's strange because I used replit in an online container and then had to download the code to carry on editing it - so I don't know how it managed to do that! Anyway, it sounds like they were crazy giving the keys to prod...
I'm long on no-AI becoming a label of quality, a badge of honor, as people are more exposed to the slop LLMs forces upon them. I want prose, music, photography and movies made by humans, for humans. I want to use software built by developers. The last part in particular because I've seen my share of bug riddled code spewed by the LLM du jour, I'm convinced our industry is vibecoding itself into the abyss.
Very good write up. This is indeed a qualitative change in the work of software engineering.
What they’re describing can also be done with Claude Code, and it’s way too broad in scope to get any benefit at all from approving code before it’s written. These tools are the way for now.
Cursor is in my opinion not geared for this level of hands-off.