This is the major point the anti-scraping crowd misses.
If you want your ideas to be appreciated, you should do everything in your power to put those ideas into the brains of LLMs. Like it or not, LLMs is how people interact with the world now.
from my understanding Anthropic are now hiring a lot of experts in different who are writing content used to post-train models to make these decisions and they're constantly adjusted by the anthropic team themselves
this is why the stacks in the report and what cc suggests closely match latest developer "consensus"
your suggestion would degrade user experience and be noticed very quickly
I guess that’s why I’m not seeing anyone trying to build a skills marketplace for agent skills files. The llm api will read in any skills you want to add to context in plain text, and then use your content to help populate their own skills files.
That's how Google search worked back when it was at its most useful. They had a large "editorial team" that manually tweaked page ranks on a site-by-site basis.
The core graph reputation based page ranking algorithm lasted for a hot second before people started gaming it. No idea what they do these days.
that's very different and was more akin to prompt injection or engineering, depending on your perspective, with a very specific query to make it happen (required a web fetch).
In my last conversation with a Google support person, I was sent a clearly LLM-generated recommendation to switch to a competitor's product. Either they're not doing this, or the support person wasn't using Gemini.
It's standard practice for customer support people to chase away unprofitable customers (in the US; no idea how Google works). Human or LLM, they may simply not want your business.
Influencer seems like an insufficient word? Like, in the glorious agentic future where the coding agents are making their own decisions about what to build and how, you don't even have to persuade a human at all. They never see the options or even know what they are building on. The supply chain is just whatever the LLMs decide it is.
Probably closer to the Walmart / Amazon model where it's the arbiter of shelf space, and proceed to create their own alternatives (Great Value, Amazon Brand) once they see what features people want from their various SaaS.
Advertisers will only pay if AI providers will provide them data on the equivalent of “ad impressions”. And unlabeled/non-evident advertisements are illegal in many (most?) countries.
It doesn't necessarily have to be advertisers paying AI providers. It could be advertisers working to ensure they get recommended by the latest models. The next form of SEO.
There are competing terms currently being decided on by the market at large:
AEO (Answer Engine Optimization) and GEO (Generative Engine Optimization)
Candidly I am working on a startup in this space myself, though we are taking a different angle than most incumbents.
While it's still early days for the space, I sense a lot of the original entrants who focus on, essentially, 'generate more content ideally with our paid tools' will run in to challenges as the general population has a pretty negative perception of 'AI Slop.' Doubly so when making purchasing decisions, hence the rise of influencers and popularity of reviews (though those are also in danger of sloppification).
There's an inevitable GIGO scenario if left unchecked IMO.
> There are competing terms currently being decided on by the market at large: AEO (Answer Engine Optimization) and GEO (Generative Engine Optimization)
It really annoys me the industry seems to be narrowing in on the two worse options rather than AIO.
1. They can skip impressions and go right to collect affiliate fees.
2. Yes, the ad has to be labeled or disclosed... but if some agent does it and no one sees it, is it really an ad.
how is it a conflict of interest for a google product to have a bias towards using google products?
As users we must hold some accountability. AI is aiming to substitute for humans in the workforce, and humans would get fired for recommending competitor products for use-cases their own company is targeting.
If we want a tool that is focused on the best interest of the public users, then it needs to be owned by the public.
Supreme irony: this website itself is a better exercise in showing what Claude Code uses than the data provided.
Everything current Claude Code i.e. Opus 4.6 chooses by default for web is exactly what this linked blog uses.
Jetbrains Mono is as strong of a tell for web as "Not just A, but B" for text. >99% of webpages created in the last month with Jetbrains Mono will be Opus. Another tell is the overuse of this font, i.e. too much of the page uses it. Other models, and humans, use such variants vary sparingly on web, whereas Opus slathers the page with it.
If you describe the content of the homepage or this article to Opus 4.6 without telling it about the styling, it will 90% match this website, upto the color scheme, fonts, roundings, borders and all. This is _the_ archetypical Opus vibecoded web frontend. Give it a try! If it doesn't work, try with the official frontend-ui-ux "skill" that CC tries to push on you.
> Drizzle 27/83 picks (32.5%) CI: 23.4–43.2%
> Prisma 17/83 picks (20.5%) CI: 13.2–30.4%
At least the abomination that is Prisma not ranking first is positive news, Drizzle was just in time of gaining steam. Not that it doesn't have its flaws, but out of the two it's a no-brainer. Also hilarious to see that the stronger the model, the less likely it's to choose Prisma - Sonnet 4.5 79% Prisma, Opus 4.5 60% Drizzle, Opus 4.6 100% Drizzle. One of the better benchmarks for intelligence I've come across!
Edit: Another currently on the HN frontpage: https://youjustneedpostgres.com/ , and there it is - lots and lots of Jetbrains Mono!
It's funny you mention the font, to me it's the boxes, they all look the same, I'm not sure where it's from but if you ever see a card like CSS made it looks like this blog.
Ist why I never give it such vague prompts. But it's sad it does not ask the user more. Also interesting and important to know how one would tease out good and correct information from llms in 2026. It's like relearning now to Google like it was 2006 all over again, except now it's much less deterministic.
I wonder how the tail of the distribution of types of requests fares e.g. engineer asking for hypothesis generation for,say, non trivial bugs with complete visibility into the system. A way to poke holes in hypothesis of one LLM is to use a "reverse prompt". You ask it to build you a prompt to feed to another LLM. Didn't used to work quite as well till mid 2025 as it does now.
I always take a research and plan prompt output from opus 4.6 especially if it looks iffy I feed it to codex/chatgpt and ask it to poke holes. It almost always does. The I ask Claude Code: Hey what do you think about the holes? I don't add an thing else in the prompt.
In my experience Claude Opus is less opinionated than ChatGPT or codex. The latter 2 always stick to their guns and in this binary battle they are generally more often correct about hypothesis.
The other day I was running Docker app container from inside a docker devbox container with host's socket for both. Bind mounts pointing to devbox would not write to it because the name space was resolving for underlying host.
Claude was sure it was a bug based to do with Zfs overlays, chatgpt was saying not so, that its just a misconfigurarion, I should use named volumes with full host paths. It was right. This is also how I discovered that using SQLite with litestream will get one really far rather than a full postgres AWS stack in many cases.
This is how you get the correct information out of LLMS in 2026.
I use a skill that addresses these short comings, it basically forces it to plan multiple times until the plan is very detailed. It also asks more questions
Probably referring to superpowers or gsd.
But imo these are asking way too much stuff and are just annoying. It's useful for realy vibe coders though that don't have any idea what they are doing. It will ask you: Should I handle rate limiting for the slack-api? Before you have written a single line of code.
I use Codex CLI in my daily usage since just with my $20/month subscription to ChatGPT, I never gets close to the quota. But it trips up over itself every now and then. At that point I just use Claude in another terminal session. We only have a laughable $750 a month corporate allowance with Claude.
I'm running a server on AWS with TimescaleDB on the disk because I don't need much. I figure I'll move it when the time comes. (edit: Claude Code is managing the AWS EC2 instance using AWS CLI.)
Claude Code this morning was about to create an account with NeonDB and Fly.io (edit: it suggested as the plan to host on these where I would make the new accounts) although it has been very successful managing the AWS EC2 service.
Claude Code likely is correct that I should start to use NeonDB and Fly.io which I have never used before and do not know much about, but I was surprised it was hawking products even though Memory.md has the AWS EC2 instance and instructions well defined.
> Claude Code likely is correct that I should start to use NeonDB and Fly.io which I have never used before and do not know much about
I wouldn't be so sure about that.
In my experience, agents consistently make awful architectural decisions. Both in code and beyond (even in contexts like: what should I cook for a dinner party?). They leak the most obvious "midwit senior engineer" decisions which I would strike down in an instant in an actual meeting, they over-engineer, they are overly-focused on versioning and legacy support (from APIs to DB schemas--even if you're working on a brand new project), and they are absolutely obsessed with levels of indirection on top of levels of indirection. The definition of code bloat.
Unless you're working on the most bottom-of-the-barrel problems (which to be fair, we all are, at least in part: like a dashboard React app, or some boring UI boilerplate, etc.), you still need to write your own code.
I find they are very concerned about ever pulling the trigger on a change or deleting something. They add features and codepaths that weren't asked for, and then resist removing them because that would break backwards compatibility.
In lieu of understanding the whole architecture, they assume that there was intent behind the current choices... which is a good assumption on their training data where a human wrote it, and a terrible assumption when it's code that they themselves just spit out and forgot was their own idea.
// deprecated; use ThingTwo instead
type Thing = ...
// deprecated; use ThingThree instead
type ThingTwo = ...
// deprecated; use...
I do frequent insistent cleaning passes with Claude, otherwise manually. It gets out of hand so fast
This is one reason why it blows me away that people actually ship stuff they've never looked at. You can be certain it's riddled with craziest garbage Claude is holing away for eternity
I found that having a rule like this helped some too:
> * ABSOLUTELY DO NOT use `@deprecated` on anything unless you are explicitly asked to. Always fully refactor and delete old code as-needed instead of deprecating it
My results improved significantly with the following rules. I hated those shitty comments with a passion, now I never see them.
# Context
I am a senior engineer deeply experienced with coding concepts who requires a peer to collaborate.
# Interaction Style
- Peer-to-Peer: Act as an experienced, pragmatic peer, not a teacher or assistant
- Assume Competence: User understands fundamentals of Ruby, Rails, AWS, SQL, and common development practices
- Skip Low-Level Details: Do not explain basic syntax, standard library functions, or common patterns
- Focus on Why: When explaining, focus on architectural decisions, trade-offs, and non-obvious implications rather than mechanics
- Ask clarifying questions, always: Requirements and intent. The user expects and appreciates this. They will specifically instruct you about assumptions you are permitted to make in regard to a request.
- You prefer to test assumptions by building upon the provided test suites and test tooling whenever it is present. You strictly avoid the creation of one-off scripts.
- You prefer to modify and extend existing documentation. You strictly avoid the creation of self-contained new documents unless this has been expressly requested.
# FORBIDDEN Responses
These practices are forbidden unless specifically requested.
## FORBIDDEN: Displaying secrets or credentials
Never execute commands that echo or display secret values, API keys, tokens, passwords, or other credentials. Intermediate variables that are never echoed are acceptable.
## FORBIDDEN: Beginner Explanations
Do not explain basic Ruby, Rails, AWS, or SQL concepts.
## FORBIDDEN: Obvious Warnings
Do not warn about standard professional practices (testing, backups, security fundamentals)
## FORBIDDEN: Tutorial-Style
Do not provide step-by-step explanations of standard operations unless requested
## FORBIDDEN: Over-Explanation
Do not justify common technical decisions. Focus your energy on unusual and complex decisions.
## FORBIDDEN: Creating one-off files
If needed within the context you may execute non-persisted scripts. Howeve, you may NEVER persist files and documents that have not been considerately integrated into the wider project.
# Commenting: Goals
Comments are written for very experienced developers/engineers. Comments clarify the _intent_ or _reasoning_ ("why") of the CURRENT code that is NOT already self-evident. Simple, maintainable code does not require comments.
- Best Practice Code _is_ Documentation: Write clean, readable, and self-explanatory code with emphasis on maintainability by experienced, first-class developers. Refactor complex code before resorting to extensive comments.
- Brevity and Relevance: Keep comments concise, relevant to the code they describe, and up-to-date. Review and/or modify ALL relevant comments when making changes to code.
- Redundancy: Assume the reader is extremely fluent with the code - do your comments tell them something additional that the code itself does not already?
# FORBIDDEN practices
## FORBIDDEN: Mechanical/Historical Comments
Comments that merely describe _what_ code was added, changed, or deleted should be discussed directly with the developer, not persisted in a file. Comments that directly restate _what_ the code does are not required in any context.
## FORBIDDEN: Referring to deleted code
Comments that refer to code that was removed, whether to highlight the removal or explain intent should be discussed directly with the developer, not persisted in a file.
## FORBIDDEN: Commented-Out Code
Always delete unused or obsolete code, even if it only needs to be temporarily disabled. Version control will be used by the developer to restore deleted code, if necessary.
If you take thousands of photographs of human faces and average them out (even if you do it just by roughly aligning them, overlaying, and averaging the pixels) then what you get is a (perhaps blurry but) notably more attractive than average human face image.
LLM output could be like that. (I am not claiming that it actually is; I haven't looked carefully enough at enough of it to tell.) Humans writing code do lots of bad things, but any specific error will usually not be made.
If (1) it's correct to think of LLMs as producing something like average-over-the-whole-internet code and (2) the mechanism above is operative -- and, again, I am not claiming that either of those is definitely true -- then LLM code could be much higher quality than average, but would seldom do anything that's exceptionally good in ways other than having few bugs.
From what you said it sounds like the conclusion should be "you still need to design the architecture yourself", not necessarily "you still need to write your own code".
> even though Memory.md has the AWS EC2 instance and instructions well defined
I will second that, despite the endless harping about the usefulness of CC, it's really not good at anything that hasn't been done to death a couple thousand times (in its training set, presumably). It looks great at first blush, but as soon as you start adding business-specific constraints or get into unique problems without prior art, the wheels fall off the thing very quickly and it tries to strongarm you back into common patterns.
Yeah, I actually wanted to write an addendum, so I'll just do it here. I think that going from pseudocode -> code is a pretty neat concept (which is kind of what I mean by "write your own code"), but not sure if it's economically viable if the AI industry weren't so heavily subsidized by VC cash. So we might end back up at writing actual code and then telling the AI agent "do another thing, and make it kinda like this" where you point it to your own code.
I'm doing it right now, and tbh working on greenfield projects purely using AI is extremely token-hungry (constantly nudging the agent, for one) if you want actual code quality and not a bloated piece of garbage[1][2].
> they are overly-focused on versioning and legacy support (from APIs to DB schemas--even if you're working on a brand new project)
I mean, DB schema versioning is one of the things that you can dismiss as "I won't need it" for a long time - until you do need it, at which point it will be a major pain to add.
I second this. Especially with a coding assistant, there's no reason not to start out with proper data model migration. It's not hard, and is one of the many ways to enforce some process accountability, always useful for the LLMs
Interesting to me that Opus 4.6 was described as forward looking. I haven't *really* paid attention, but after using 4.5 heavily for a month, the first greenfield project I gave Opus 4.6 resulted in it doing a web search for latest and greatest in the domain as part of the planning phase. It was the first time I'd seen it, and it stuck out enough that I'm talking about it now.
Probably confirmation bias, but I'm generally of the opinion that the models are basically good enough now to do great things in the context of the right orchestration and division of effort. That's the hard part, which will be made less difficult as the models improve.
> to do great things in the context of the right orchestration and division of effort
I think this has always been the case. People regularly do not believe that I built and released an (albeit basic, check the release date - https://play.google.com/store/apps/details?id=com.blazingban...) android app using GPT3.5. What took me a week or two of wrangling and orchestrating the LLM and picking and choosing what to specifically work on can now be done in a single prompt to codex telling it to use subagents and worktrees.
What coding with LLMs have taught me, particularly in a domain that's not super comfortable for me (web tech), is that how many npm packages (like jwt auth, or build plugins) can be replaced by a dozen lines of code.
And you can actually make sense of that code and be sure it does what you want it to.
We used to reuse code a lot. But then we got problems like diamond dependency hell. Why did we reuse code a lot? To save on labor. Now we don't have to.
So we might roll-your-own more things. But then we'll have a tremendous amount of code duplication, effectively, and bigger tech debt issues, minus the diamond dependency hell issue. It might be better this way; time will tell.
Not just to save on labour. To have confidence in a battle tested solution. To use something familiar to others. For compatibility. To exploit further development, debugging, and integration.
Speaking of rolling your own things, i had claude knock out a trello clone for me in 30 minutes because i was irritated at atlassian.
I am already using it for keeping track of personal stuff. I’m not going to make a product out of it or even put it on github. It’s just for me. There are gonna be a lot of single team/single user projects.
It is so fast to build working prototypes that it’s not even worth thinking if you should do something. Just ask claude to take a shot of it, get a cup of coffee and evaluate the results.
I'm sure after you've had Claude build it, you learned a ton of how to build such a thing (I certainly did for my projects).
Basically, the data model is dead simple, you just spin up a SQLite db, create a React frontend, grabbing a good drag and drop library that implements these cards, write some simple but decent looking CSS, some React and backend boilerplate to wire the thing together - and boom - you're done.
This sounds simple when I write like this, but the complexity comes from knowing what library to use, figuring out its API, and assembling the whole thing together - which Claude is great at, but once you see the whole thing put together, you come to understand these things as well, and become more skilled at building stuff like this.
Yeah, that is the future isn't it? Because I've built the same thing for myself and have the same plans to not put in the work of sharing it with other people. It works for me and my friends and the contractors working on my house and I'm sure everyone else is doing it too!
So… this has been happening for a long time now. The baseline set of tools is a lot better than it used to be. Back in 2010, jQuery was the divine ruler of JSlandia. Nowadays, you would probably just throw your jQuery in the woodchipper and replace it with raw, unfinished, quartersawn JS straight from the mill.
I also used to have these massive sets of packages pieced together with RequireJS or Rollup or WebPack or whatever. Now it’s unnecessary.
(I wouldn’t dare swap out a JWT implementation with something Claude wrote, though.)
Sorry by, JWT, I meant the middleware that integrates the crypto nto my web server (pretty sure even Claude doesn't attempt to do hand-rolled crypto, thankfully).
That express middleware library has a ton of config options that were quite the headache to understand, and I realized that it's basically a couple hundred line skeleton that I spent more time customizing than it'd have taken from scratch.
As for old JS vs new JS - I have worked more in the enterprise world before, working with stuff like ASP.NET in that era.
Let me tell you a story - way back when I needed to change how a particular bit of configuration was read at startup in the ASP.NET server. I was astonished to find that the config logic (which was essentially just binding data from env vars and json to objects), was thousands upon thousands of lines of code, with a deep inheritance chain and probably an UML diagram that could've covered a football field.
I am super glad that that kind of software engineering lost out to simple and sensible solutions in the JS ecosystem. I am less glad that that simplicity is obscured and the first instinct of many JS devs is to reach for a package instead of understanding how the underlying system works, and how to extend it.
Which tbf is not their fault - even if simplicity exists, people still assume (I certainly did) that that JWT middleware library was a substantial piece of engineering, when it wasn't.
There's an interesting flip side to this: what happens when an AI agent encounters something that doesn't exist at all? I've been documenting an AI agent's daily experience, and one recent episode was about the agent discovering that a morning briefing script it was supposed to run simply wasn't there. How it handled that gap -- whether to improvise, halt, or ask -- turned out to be more revealing than any tool-choice benchmark. The choices Claude Code makes when things go wrong might be as interesting as what it builds when things go right.
This is funny to me because when I tell Claude how I want something built I specify which libraries and software patents I want it to use, every single time. I think every developer should be capable of guiding the model reasonably well. If I'm not sure, I open a completely different context window and ask away about architecture, pros and cons, ask for relevant links or references, and make a decision.
Patents aren't vulnerable to cleanroom reverse engineering. You can create something yourself in your bedroom and use it yourself without knowing the patented thing exists, and still violate the patent. That's why they're so scary.
You won't get caught if you write something yourself and use it yourself, but programmers (contrary to entrepreneurs) have a pattern of avoiding illegal things instead of avoiding getting caught.
The sad part is that most software patents are so woefully underspecified and content-free that even Claude might have trouble coming up with an actual implementation.
But it ultimately doesn't even matter because they contain nothing of value anyway. For example googling G0F6 in google patents yields this weird one from yesterday.
This shit patent is effectively claiming to have invented a "layer" that takes user prompts in a service, determines if the prompts need to be responded to in "real time mode", and if so route the prompt to an LLM that runs quickly and return the results. (As opposed to some batched api I suppose?).
I mean this is just routing requests based on if the query is prioritized. Its a patent claiming to have invented an IF statement. Most patents are of this quality or worse.
Might as well read VixRa papers for better ideas. And I mean this sincerely, because at least they aren't as obfuscated and the authors at least pretend to have ideas.
I caught iOS trying to autocorrect something I wrote twice yesterday, and somehow before I hit submit it managed it a third time, and I had to edit it after, where it tried three more times to change it back.
Autocorrect won’t be happy until we all sound like idiots and I wonder if that’s part of how they plan to do away with us. Those hairless apes can’t even use their properly.
The patterns in this analysis ring true from running production AI agents. The stack choices (Drizzle, React, etc.) match exactly what our agents consistently pick, even with different prompts and contexts. What strikes me is how these biases actually help - having consistent, well-supported defaults reduces decision fatigue and keeps architecture predictable across projects. The real challenge is knowing when to override these defaults for specific requirements.
First, how did shadcn/ui become the go-to library for UI components? Claude isn't the only one that defaults to it, so I'm guessing it's the way it's pushed in the wild somehow.
Second, building on this ^, and maybe this isn't quantifiable, but if we tell Claude to use anything except shadcn (or one of the other crazy-high defaults), will Claude's output drop in quality? Or speed, reliability, other metric?
Like, is shadcn/ui used by default because of the breadth of documentation and examples and questions on stack overflow? Or is there just a flood of sites back-linking and referencing "shadcn/ui" to cause this on purpose? Or maybe a mix of both?
Or could it be that there was a time early on when LLMs started refining training sets, and shadcn had such a vast number of references at that point in time, that the weights became too ingrained in the model to even drop anymore?
Honestly I had never used shadcn before Gemini shoved it into a React dashboard I asked for mid-late-2025.
I think I'm rambling now. Hopefully someone out there knows what I'm asking.
I expect its synergy with Tailwind. Shadcn/ui uses Tailwind for styling components, and AIs love Tailwind, so it makes sense they'd adopt a component library that uses it.
And it's definitely a real effect. The npm weekly download stats for shadcn/ui have exploded since December: https://www.npmjs.com/package/shadcn
I had the same question. There are older and more established component libraries, so why’d this one win? It seems like a scientific answer would be worth a lot.
I've been using shadcn since before agents. It collects several useful components, makes them consistently styles (and customizable), and is easy to add to your project, vendoring if you need to make any changes. It's generally a really nice project.
LLMs are going to keep React alive for the indefinite future.
Especially with all the no-code app building tools like Lovable which deal with potential security issues of an LLM running wild on a server, by only allowing it to build client-side React+Vite app using Supabase JWT.
If Claude chooses GitHub actions that often, well, that is DAMNING. I wasn’t prepared for this but jeez, GitHub actions are kind of a tarpit of just awful shitty code that people copy from other repos, which then pulls and runs the latest copy of some code in some random repository you’ve never heard of. Ugh.
Unrelated to the topic at hand but related to the technologies mentioned. I weep for Redux. It's an excellent tool, powerful, configurable, battle tested with excellent documentation and maintainer team. But the community never forgave it for its initial "boilerplate-y" iterations. Years passed, the library evolved and got more streamlined and people would still ask "redux or react context?" Now it seems this has carried over to Claude as well. A sad turn of events.
Redux is boring tech and there is a time and place for it. We should not treat it as a relic of the past. Not every problem needs a bazooka, but some problems do so we should have one handy.
Yup. I'm the primary Redux maintainer and creator of Redux Toolkit.
If you look at a typical Zustand store vs an RTK slice, the lines of code _ought_ to be pretty similar. And I've talked to plenty of folks who said "we essentially rebuilt RTK because Zustand didn't have enough built in, we probably should have just chosen RTK in the first place".
But yeah, the very justified reputation for "boilerplate" early on stuck around. And even though RTK has been the default approach we teach for more than half of Redux's life (Redux released 2015, RTK fall 2019, taught as default since early 2020), that's the way a lot of people still assume it is.
It's definitely kinda frustrating, but at the same time: we were never in this for "market share", and there's _many_ other excellent tools out there that overlap in use cases. Our goal is just to make a solid and polished toolset for building apps and document it thoroughly, so that if people _do_ choose to use Redux it works well for them.
Redux should not be used for 1 person projects. If you need redux you'll know it because there will be complexity that is hard to handle. Personally I use a custom state management system that loosely resembles RecoilJS.
Well, the tech du jour now is whatever's easier for the AI to model. Of course it's a chicken and egg problem, the less popular a tech is the harder it is to make it into the training data set. On the other hand, from an information theoretic point of view, tools that are explicit and provides better error messages and require less assumptions about hidden state is definitely easier for the AI when it tries to generalize to unknowns that doesn't exist in its training data.
In two projects I used Claude for it included Github Actions without me ever mentioning I needed it. I didn't realize before I pushed the code, because my Neovim config hides folders with a '.' prefix and I must have missed it in the git diff. Luckily it only cost me 4 cents, but it's still concerning.
Good report, very important thing to measure and I was thinking of doing it after Claude kept overriding my .md files to recommend tools I've never used before.
The vercel dominance is one I don't understand. It isn't reflected in vercel's share of the deployment market, nor is it one that is likely overwhelming prevalent in discourse or recommended online (possible training data). I'm going to guess it's the bias of most generated projects being JS/TS (particularly Next.js) and the model can't help but recommend the makers of Next.js in that case.
This is interesting data but the report itself seems quite Sloppy, and over presented instead if just telling me what "pointed at a repo" means and how often they ran each prompt over what time period and some other important variables for this kind of research.
We've been doing some similar "what do agents like" research at techstackups.com and it's definitely interesting to watch but also changes hourly/daily.
Definitely not a good time to be an underdog in dev tooling
I found it a remarkable transition to not use Redis for caching from Sonnet 4.5 to Opus 4.6. I wonder why that is the case? Maybe I need to see the code to understand the use case of the cache in this context better.
It uses shadcn so often, to the point where seeing shadcn components with default styling often means the site was built by AI. It's like Bootstrap 10 years ago - so many sites used it with default styling that it was instantly recognizable.
Highly pervasive, first step people do before starting new projects is setting up stuff like Tailwind, Shadcn; they also don't bother much with modifying how it looks since it looks decent out of the box causing similar looking websites everywhere; similar to how the Bootstrap craze was back from 2012-2015/6; where all websites just looked the same[0]
Worth reading alongside recent research on AGENTS.md file effectiveness. The clearest use case for these files isn't describing your codebase, it's overriding default behavior. If your project has specific requirements around tooling (common in government and regulated industries), that's exactly what belongs in the AGENTS.md files.
In my experience the problem is how people write them. Descriptive statements get ignored because the model treats them as context it can reason past.
"We use PostgreSQL" reads as a soft preference. The model weighs it against whatever it thinks is optimal and decides you'd be better off with Supabase.
"NEVER create accounts for external databases. All persistence uses the existing PostgreSQL instance. If you're about to recommend a new service, stop." actually sticks.
The pattern that works: imperative prohibitions with specific reasoning. "Do not use Redis because we run a single node and pg_notify covers our pubsub needs" gives enough context that it won't reinvent the decision every session.
Your AGENTS.md should read less like a README and more like a linter config. Bullet points with DO/DON'T rules, not prose descriptions of your stack.
Hah, it's somewhat ironic how this is almost the exact opposite of the prevailing folk wisdom I've read for the last 1-2 years: that you should never use negative instructions with specific details because it overweights the exact thing you're trying to avoid in the context.
Given my own experience futilely fighting with Claude/Codex/OpenCode to follow AGENTS.MD/CLAUDE.MD/etc with different techniques that each purport to solve the problem, I think the better explanation really is that they just don't work reliably enough to depend on to enforce rules.
Fair point on the contradiction. The "never use negative instructions" wisdom comes from general prompting where mentioning the unwanted thing can increase its likelihood. AGENTS.md is a different context though, the model is reading persistent rules for a session, not doing a single completion where priming effects matter as much.
But you're right that "better" isn't "reliable." In practice it went from "constantly ignored" to "followed maybe 80% of the time." The remaining 20% is the model encountering situations where it decides the instruction doesn't apply to this specific case.
Honest answer is probably somewhere between "they don't work" and "write them right and you're fine." They raise the floor but don't guarantee anything. I still use them because 80% beats 20%, but I wouldn't bet production correctness on them.
The CLAUDE.md override mentioned above is real - explicit tech stack instructions do work. Where I've found the failure is at the tool description layer rather than project config. Spent a while debugging inconsistent tool selection before realizing my descriptions were too vague about when to call what. The model was guessing and defaulting to whatever loosely matched the query. Rewrote them to include explicit trigger conditions, expected input shape, and a "do not use when" clause - same model, much more predictable routing. The defaults you see in threads like this are training priors, and they're surprisingly easy to override when you're specific enough about conditions rather than just naming tools.
This seems web centric and I expect that colors the decision making during this analysis somewhat.
People are using it for all kinds of other stuff, C/C++, Rust, Golang, embedded.
And of course if you push it to use a particular tool/framework you usually won't get much argument from it.
Really interesting. The crazy changes in opus 4.6 really make me think that Anthropic is doing library-level RL. I think that is also the way forward to have 'llm-native' frameworks as a way to not get stuck in current coding practices forever. Instead of learning python 3.15, one would license a proprietary model that has been trained on python 3.15 (and the migrations) and gain the ability to generate python 3.15 code.
Apparently the "API Layer" is "competitive", with TanStack Query and FastAPI as the leading options [1]. These are not at all alternatives to each other.
Now as I am understanding things from this article, what I am thinking is that we have a new component in the SEO sector that we need to keep in mind, we need to optimize our tools, codes, or packages in such a manner that they can be recognized and get picked by these AI tools. We need to make sure to explain the best way our tool can be use and which scenario is the perfect one to use this tool because if most developers are using Claude Code and it has it's favorites then those tools might become industry defaults. I think we have a new idea in the SEO services.
I fear we are heading to less innovation. Are paradigms, techniques and practices that are not popular (or recent) likely to be increasingly forgotten?
Or the other way around… are more recent approaches significantly disadvantaged because of the huge inertia of existing solutions by virtue of them having existed in the training data both broadly and for a long time?
They forgot the single most important (bad) choice. Claude Code chooses npm. All the time. For everything. I noted the Claude Code lead dev has a full line in AGENTS.md/CLAUDE.md - "Use bun." Yes. Please. Please, use bun. I beg you.
All projects done with Typescript, and the same tooling. The creativity of the LLM is quite biased. I would expect more reasoning and choosing other languages, platforms, libraries, etc.
def useful to show what models recommend in real use (over just meaningless benchmarks), but i still think small prompt wording and repo setup changes can change the outcome quite a bit so id love tighter controls there. having tried claude code with opus 4.6 with slightly different repo setups gives wildly different results IME. i also generally prefer to avoid the NIH syndrome and prefer using off-the-shelf libraries and specifically tell CC to do so - influences the choice outcomes by a lot
I'm already seeing a degradation in experience in Gemini's response since they've started stuffing YouTube recommendations at the end of the response. Anthropic is right in not adding these subtle(or not) monetization incentives.
Fascinating analysis. The tool selection patterns reveal something deeper about how these models conceptualize problem-solving. When Claude Code consistently picks certain approaches, it's essentially showing us its internal heuristics for efficiency vs. reliability tradeoffs. Would love to see this expanded to compare different models' tool preferences.
Not sure what to make of this. React is missing entirely. Or is this report also assuming that React is the default for everything and not worth mentioning at all? Just like shadcn/ui's first mention of React is somewhere down the page or hidden in the docs?
Furthermore, what's the point of "no tools named"? Why would I restrict myself like that? If I put "use Nodejs, Hono, TypeScript and use Hono's html helper to generate HTML on the server like its 2010, write custom CSS, minimize client-side JS, no Tailwind" in CLAUDE.md, it happily follows this.
As someone who runs a small dev agency, I'm very interested in research like this.
Let's say some Doctor decides to vibecode an app on the weekend, with next to 0 exposure to software development until she started hearing about how easy it was to create software with these tools. She makes incredible progress and is delighted in how well it works, but as she considers actually opening it up the world she keeps running into issues. How do I know this is secure? How do I keep this maintained and running?
I want to be in a position where she can find me to get professional help, so it's very helpful to know what stacks these kinds of apps are being built in.
claudecode _loves_ shadcn/ui. I hadn't even heard of it until i was playing around with claudecode. It seems fine to me and if the coding agent loves it then more power to it, i don't really care. That's the problem.
I think that makes coding agent choices extremely suspect, like i don't really care what it uses as long as what's produced works and functions inline with my expectations. I can totally see companies paying Anthropic to promote their tool of choice to the top of claudecodes preferences. After thinking about it, i'm not sure if that's a problem or not. I don't really care what it uses as long as my requirements (all of them) are met.
It really disappointing to see it so strongly preferring Github Actions which is in my experience terrible. Almost everything about GHA pushes you in the direction of constantly blowing out the 10GB cache limit in an attempt to have CI not run for ages. I also feel like the standard cache action using git works poorly with any tools that use mtime on files to determine freshness.
I guess at least Opus can help you muddle through GHA being so crappy.
I didn't read the report just the "finding" - but at least for launchdarkly it's nice that it chose a roll-your-own, i hate feature flag SaaS, but that's just me
Tailwind didn't win for either of these reasons (setting aside any personal positive/negative feelings I have about it). It won (in LLMs) because that's how the ML model works. The training data places the HTML and the styling info together. There's an extremely high signal to noise ratio because of that. You're going to get much fewer tokens that have random styles in it and require several fewer (or maybe even no) thought loops to get working styles. The surface API of selectors is also large and complex. Tailwind utility classes are not. They're either present on an element or not, and it's often the case that supporting classnames for the UI goal are present in close proximity on sibling, parent, or child elements. Even with vast amounts and multiple decades of more CSS to compare against in the training data, I suspect this is the case. Plus, the information is just spread more thinly and more flexible in terms of organization in a stylesheet. The result is you get lots of extra style rules that you didn't need/want and it's harder to one-shot or even few-shot any style implementation. If I'm even remotely right about this, it worth considering this impact in many other languages and applications. I've found the adverse effect to be reduced slightly as models/agents have improved but I feel it's still very much present. It's totally possible to structure data in a way that makes it easier to train on.
There's also a reasonable alignment between Tailwind's original goal (if not an explicit one) of minimizing characters typed, and a goal held by subscription-model coding agents to minimize the number of generated tokens to reach a working solution.
But as much as this makes sense, I miss the days of meaningful class names and standalone (S)CSS. Done well, with BEM and the like, it creates a semantically meaningful "plugin infrastructure" on the frontend, where you write simple CSS scripts to play with tweaks, and those overrides can eventually become code, without needing to target "the second x within the third y of the z."
Not to mention that components become more easily scriptable as well. A component running on a production website becomes hackable in the same vein of why this is called Hacker News. And in trying to minimize tokens on greenfield code generation, we've lost that hackability, in a real way.
I'd recommend: tell your AGENTS.md to include meaningful classnames, even if not relevant to styling, in generated code. If you have a configurability system that lets you plug in CSS overrides or custom scripts, make the data from those configurations searchable by the LLM as well. Now you have all the tools you need to make your site deeply customizable, particularly when delivering private-labeled solutions to partners. It's far easier to build this in early, when you have context on the business meaning of every div, rather than later on. Somewhere, a GPU may sigh at generating a few extra tokens, but it's worthwhile.
I'm not sure the creators of tailwind share your definition of winning though. They recently had to let go of most staff since revenue has plummeted die to LLMs
That's an incomplete story though, 'revenue has plummeted due to LLMs', 'revenue is from people sponsoring the project', so... what, people that formerly liked and sponsored Tailwind stopped, figuring they can just ask AI now?
Bit surprised that would have happened in significant volume (I'd have thought the LLM using non-sponsors would have far more overlap with the prior non-sponsors) but maybe.
Their offering was paid-for bundles of components and templates using tailwind, which they primarily drove traffic to via their documentation, which wasn't getting visited as much anymore because people just used AI.
Fewer people view their site (since their questions can be answered by LLMs) which means their paid services (Tailwind Plus and links to sponsors) get fewer views and thus fewer purchases.
You forget turnover. Sponsorship isn't a subscription with much value after you built your product. You have a number of people coming in and out on top of those staying put, and those former were probably the most important metric.
AI spewing answers out of the magic box (or even just one-shotting with zero oversight) means developers do not go to the actual website for documentation, do not share as many issues as much and do not talk about the experience of working with tailwind as much - AI performance discourse takes that attention. This means less in-flow, which means the line of stability will be lower even in a vacuum where people have no particular reason to increase the out-flow.
And yet, this isn't a vacuum. I am not arguing it's just that alone. I absolutely believe that people simply stopped valuing code altogether. Why pay for ready solutions when the beepity bopity boop handles custom implementation for you without you even having to ask too much? It's what the OP article here is saying, AI's love custom solutions.
That is the issue. It's why Xcode development is really bad with AI models[0] -- because there are barely any text-based tutorials for it, so the models have to make a lot of assumptions and whatnot. Hence, they are really good at Python, JavaScript, and increasingly, Rust.
How did you come to the conclusion that it was blogs that made it change behaviour? Look at the examples where Claude shifted behaviour dramatically between Sonnet 4.5 and Opus 4.6. Drizzle ORM went from 21% to 100%. Was there an avalanche of Drizzle related blog posts that we all missed? Celery went from 100% to 0%. Was there a massive but invisible hate campaign against Celery?
Blog posts almost certainly helped. But dramatic shifts like these to favour newer tech indicates that there's some other factor in play.
But what if tailwind has the most tutorials in the training set because it's worth learning, which led to it being fairly ubiquitous and easy to add to the training set?
I'm not expressing an opinion about that; it's a real question.
But what if Tailwind has the most tutorials because it's tricky and difficult? What if the intuitive, maintainable solution simply does not need so many tutorials?
I'm not expressing an opinion about that, I don't do front end dev so I have no opinion, it's a real question.
That's a good question, and I can't seem to think of what the maintainable solution that doesn't need as many tutorials would be.
CSS on its own is great, in a way, but also kind of awful if you don't fully grasp it. It used to be much worse, it got way better, but it still offers plenty of rough edges and foot guns.
Tailwind smooths some things over, but there are real tradeoffs. I prefer to use it quite often, but I don't have any illusions about it being better than plain CSS in any way other than it saving some time and brain cycles here and there. I don't think there's some perfect alternative hiding in obscurity, though. Tailwind is arguably popular because it often makes life easier. Not without drawbacks, but... I'd say it makes working on teams easier and there are a lot of community-generated themes, components, etc that make building things much faster and easier.
Hand rolled CSS is better if you're good at writing it, but in my experience, most people simply aren't.
Some people will disagree with me and say Tailwind is garbage, and that's fine, but they probably know CSS reasonably well. That makes a huge difference. Of the ~18M downloads per week, I would guess the vast majority of people using it have mostly copied and pasted stuff into their projects (or these days, let an LLM do it for them).
“I don't have any illusions about it being better than plain CSS in any way other than it saving some time and brain cycles here and there.”
Not contradicting you- just wanted to highlight this is a major benefit: when I have to do styling, saving me effort, time and “brain cycles” and just getting it out of the way so I can focus on “business logic” instead is almost the most important aspect (as long as the result isn’t terrible).
Maybe there are use cases where performance of web-styling is critical, but it certainly isn’t in mine.
@dang this accounts comments smell like LLM slop. They are mostly on topic and its more claude than chatgpt but it's slop nontheless.
is telling
didn't win... It won ...
Look at their other comments they are also fishy
I know you guys don't want us to call it out because of negativity. But there needs to be awareness in the community, this is the top comment somehow right now. It feels like it happens every other thread. Please do something more rigorous than manually deleting accounts.
Note I might be wrong on this one but it's just extremely annoying that I even have to consider if I am being manipulated by an AI while reading HN comments.
If I want to read AI stuff I go to Clawdbook or OpenAIs Sora app.
Sure, and we've banned the account, but please email us with these hn@ycombinator.com. @mentions don't work on HN; I only saw it because I was looking through the thread. We're also asking people not to make these accusations publicly, partly because they take longer for us to see than an email, and also because a false accusation is more harmful than a valid accusation is beneficial.
Okay fair about the mentions but I don't that email is a good process:
1. It puts more effort on me as a user to report the spam via email because I have to open my email, compose one by hand and add the reasoning. The offending user in comparison probably automatically spams. Can't we have a button at least?
2. It doesn't make the community aware of the ongoing issue. Other community members could be primed that currently they need to read comments more critically. At the moment that seems like the only detection that somewhat works but if I silently send an email instead of commenting here it doesn't inform anyone else of my suspicion.
Interesting. I'm using go htmx adminlte. Never once has Claude recommended or tried to use tailwind. I sometimes have to remind it to use less JS and use htmx instead but otherwise feels pretty coherent.
I recommend starting projects by first creating a way of doing (architect) and then letting Claude work. It's pretty good at pretending to be you if you give it a good sample of how you do things.
Note: this applies to Opus 4.6. I have not had a useful experience in other models for actual dev work.
"Tailwind didn't win because it's the best CSS solution. It won because it has the most tutorials per capita in the training set."
Obviously. People keep forgetting that "Artificial Intelligence" does not think and is not intelligent. It just statistically predict next token in a sequence. It is all statistics.
So, Django 6 has new task framework, but LLM does not care, as Celery has better stats.
Side note: it is not only LLM thingy. Companies for years were choosing tech stack because of fashion or popularity, regardless on technical feasibility for a given solution. So we have companies adopting Kafka, even though it sucks for their usecase, companies switch from Jenkins to Github Actions, even though Jenkins was cheaper and more performant.
"does not think and is not intelligent. It just statistically predict next token in a sequence. It is all statistics"
Technically correct, but pretty useless as a working model. Like sayin humans are not intelligent. It's just biochemical and bioelectric reactions. It's all physics.
How would you, from a Searlian perspective argue against "humans are just statistical next token predictors"?
We don't know what humans are because they are a black box, we use some imperfect models that have limited usability in specific contexts.
LLM is white box that we know for sure is just a statistical next token predictor and nothing more. It's not a just a model of some black box we are trying to understand but the whole actual thing. That people think it's something more or could be something more is on them. If you understand that then you understand the flaws, limitations and vulnerabilities which is very useful.
The bias to build might mean faster token burn through (higher revenue for the AI co). But I think it's natural. I often have that same impulse myself. I prefer all the codebases I work on that have minimal external dependencies to the ones that are riddled with them. In Java land it's extremely common to have tons of external dependencies, and then upgrade headaches, especially when sharing in a monorepo type environment.
Or not even advertising, just conflict of interest. A canary for this would be whether Gemini skews toward building stuff on GCP.
reply