commented: “AI skills” in this post means the skills of humans who use AI. I had initially thought it meant skills for AI agents, the Markdown-based format for reusable prompts. commented: ah! I was confused about it. Clear writing skills still matter. commented: I wasn't. It was obvious to me from the title alone that it was about how proficient a human is at using a generative AI. Option 1: I'm a genius and you're not. Option 2: It's the first time I'm hearing about those SKILL.md files. Here's a wild guess: just like me, the author didn't know of the Markdown format for prompts, or at least was hardly ever exposed to them. Clear writing is harder when jargon hijacked common words in the mind of your audience. commented: I agree that clear writing is harder for the uninformed! commented: For me it was the quotes. I assumed correctly based on the current title; but would have assumed it was about SKILL.md if the title started with AI "Skills" are [...]. commented: the ways of working with models that get us to a useful result faster, if they work, are going to be folded back into the product, packaged, and become the default experience a few months later Feel free to flag as unhelpful etc. but it seems like the default experience a few months from today will be a billing screen saying you're out of tokens. commented: I don’t think that’ll be the default experience a few months from now for a few reasons: Token prices are dropping. Here’s a blog post from yesterday. From Anthropic’s S1 filing, their inference margins are 85% on rented hardware. Anthropic also said in their S1 that most of their capacity comes from Amazon, Google, and Microsoft and that it wants more than what the cloud providers are willing/able to provide. Regarding Accenture, I think they’re trying to cut down on “convert this slide deck to PDF” style usage, but it appears that software developers will fall under the $250/week budget cap. commented: wonderful irony that you linked to an article no one can read without paying. a billing screen is the future of everything. commented: This is a huge societal problem and I have no clue how to even approach it. Good journalism costs a lot of money. Journalistic outfits need a source of revenue, and the only way to get that without a conflict of interest is to make your readers pay for articles. But there are lots of funding options if you're willing to do the bidding of billionaires, politicians and industry. As a result, good journalism is inaccessible, while propaganda floods the market. commented: There is the model radio and television used: https://en.wikipedia.org/wiki/Television_licence In Germany this was extended somewhat to internet-based publishing. Essentially a flat fee payed by every household. People hate paying that. commented: Government-funded independent media is a model which can work, depending on the country. There's obviously huge risks of conflicts of interest there too, not all governments will happily fund media companies which focus on harsh criticism of the government. I don't think the US is in a place where that works though. Also, the TV license thing is somewhat orthogonal. The government can fund independent media irrespective of how the government obtains the money to fund independent media. commented: i think this is true, i have coworkers spending $20k+ a month in AI. I do wonder what will happen when the price of tokens normalise to a break even point commented: I do wonder what will happen when the price of tokens normalise to a break even point Given the developments over the past few years and the trends they imply, I’d say what will happen is a reduction in price for the same output quality. But I’m still unsure what that implies for the spending ability of people like your coworkers. commented: It is always a mistake to assume that this will be a linear progression. commented: Most of the genuinely competent developers I know have fairly low AI spends, maybe $1-2k per month max, and often a lot lower. The only way to spend $20k is to flail around with agent swarms and fully commit to no human ever seeing the code. Maybe Anthropic can make this work. Most people can't, and they're just making a mess. Or they're writing giant "skill" files with "programs" written in English prose, and relying on the LLM as a very expensive interpreter. So what happens to the people burning cash? Most likely, somebody in accounting will make them switch to one of Luna 5.6, Haiku 5.5, or a commodity inference provider serving a DeepSeek or GLM "Flash" model. These models are all relatively smart and close to free at any kind of scale, with open weight options around $0.10/0.20 or less per million tokens. Yeah, they're also less competent, but the latest round of mid-sized "flash" models compete with Sonnet at less than Haiku prices. Or just wait for some hyperscalers to go bankrupt and try to snag one the GPUs sitting in the warehouses at a liquidation sale. commented: Most of the genuinely competent developers I know have fairly low AI spends, maybe $1-2k per month max, and often a lot lower. Err, after paying for health insurance, unemployment insurance, and retirement plan, I have 2.4k euros left per month. Call it a $4-5k package per month, tops. That "low-spend" is like up to 50% my entire salary. No way I'm spending anywhere near that much. And I'm not sure my employer is willing to either. In fact, I have yet to be given meaningful access, and don't expect that to change for at least several months. I understand you gave a ceiling, but even as a ceiling I have to say as someone who only prompted a free chatbot from time to time, this feel absolutely ludicrous. commented: I've been spending a lot of time looking at this chart recently because it's very counter intuitive for me. => https://artificialanalysis.ai/?models=claude-haiku-5-5-low%2Cclaude-fable-5-1#intelligence-index-vs-cost-per-task-chart What it shows is that Fable 5.1 (on max effort) is 380 times more expensive than Haiku 5.5 (on low). Anthropic and OpenAI are selling, mostly successfully, their "frontier" models. But Haiku 5.5 is better than frontier models released a year ago, and there are a lot of tasks that it does very well on. It raises questions about the nature and value of intelligence (is it worth it to spend 100x times as much for a more "intelligent" model?) that I'm not really prepared to answer right now. But I don't think anyone is answering those questions, I think the vibe coders are slapping the "Fable" button and throwing money at Anthropic. commented: I’ve been unwell lately (well, I have a physical ailment that’s progressively worsened over nine years or so), so I’ve done a lot more programming in the pre-AI era than the post-AI era. So maybe that’s why the idea of “AI” only skills is lost on me. From my perspective, it’s all systems engineering as I’ve had to figure out tooling for MCP servers, automating ad hoc skills development, configuring context engineering, and much more than I can describe with any meaningful technical depth in this comment. I am not sure where people are getting the “it’s only chats” impression. That’s for consumers of AI services. If you’re developing AI services, or doing research, then you’re doing a lot more than chats. The prompts part of prompt engineering is a “chat”, but you’re not going to get anything useful or productive if you haven’t done the legwork to set things up before you get chatting. The setting things up part is the software/systems engineering part that’s still the same. I am not evangelizing here, so I don’t care if you’re not convinced. I am okay because I know what’s working for me and what doesn’t. commented: I’d never seen or heard of that Stephen Bradbury ice skating Olympic race. Brilliant! commented: It reminded me that allowing your competition to make mistakes and focusing on just maintaining your own low-risk survival approach is oftentimes a valid strategy. I have personally noticed that while playing table tennis with my classmates. That game is a great example of how escalation is also risky for an aggressor. commented: I think there's a bit of nuance lost in this one. It may look like "Anyway, remember prompt engineering?" is truly a funny dig now... but that's not the whole story. You can still prompt engineer for good results and you need to learn more about how to do it. It's closer to good search skills from before Google became useless. For example: what do you expect to give you better results? "Does this app have bugs?" or "In context of this app, what approaches would an experienced software engineer use to validate it works correctly? Make a list ordered by effectiveness-vs-cost and spawn subagents to check the trivial ones. Make a plan for review to run the remaining checks. Check the project history for areas which needed most bug fixing work." The second one requires almost no familiarity with software engineering - and a similar prompt can be derived for other industries. But it takes some experience to realise this is possible and how to push the boundaries yourself. We needed the basics of prompt engineering and now we can take it to the next level. It never went away. commented: The core skill in using these tools is the same as the skill you need when writing code by hand. You have to understand the given requirement, the shape of the solution, and how to verify that the code that was produced actually matches the intent. People who don't work with LLMs seem to have this naive idea that it's sort of like a slot machine where you put a query in and then you either get a correct result or some nonsense and you have to try again. The way this process actually works is that you start by understanding the problem and figuring out how you want to solve it in terms of architecture, data structures, and algorithms that will be used. Then you get the LLM to come up with a rough shape of the solution, and iterate on it until it works the way you want. commented: By the time I've come up with the architecture, data structures and algorithms, why would I use an LLM when I can actually write the code? If I have to be explicit in what I tell the LLM (from what I understand, from several back-and-forth conversations), then closely verify the code (because we all do, right?), does that save me time? Less bugs? commented: Because there's still a ton of code to be written in any given project, and most of it is repetitive stuff you've written many times before like making a service endpoint or accessing the database. You can be explicit in a sense of defining functional requirements and providing a spec, while the LLM handles all the boring implementation details. Having worked with these tools extensively for over a year now and having built large projects this say, I can definitively say that they both save time and result in less bugs when used responsibly. commented: Isn't handing the boring implementation details the responsibility of a framework? It handles all the stuff that has to be done regardless, and the programmer just modifies the bits that handle the business logic? I mean, we had this back in the 90s with Visual Basic. Draw out the window and whatever controls you needed, then wire in the required business logic. You could even get IDEs for both Windows and Mac that would generate an empty application (main window, menu bar, etc.) saving you from having to grind all that code out. It did all the "boring implementation details" for you. We seem to have regressed in that regard. commented: A framework doesn't really solve the problem, and frameworks introduce plenty of problems of their own. They work well as long as everything you're doing is vanilla and fits neatly into the framework design, but as soon as you need to do something that's different it becomes a pain. Frameworks also introduce a ton of overhead because they have to be the lowest common denominator for a broad set of problems. And there's a reason practically nobody uses VB style tools nowadays. commented: Having turned on an off whole self configuring clusters... Twenty years ago.. like a slowish light switch with a single command and trained non programmer business analysts to use it so that I didn't have to, I implore language model users to not think so small, maybe things would actually get better, we'd need fewer models on fewer machines. Everyone having this perspective of a mountain of code as a given may not tell the whole story of what is gained and lost in designing and using systems with people. commented: The LLMs come up with the architecture, data structures, and algorithms as well now. They often do it better than I would, and when they don't I tell them what I'd do instead. That's new as-of the Fable 5 / Astra 6 / Opus 5.5 generations. Prior to that I would expect them to make enough bad choices at the higher level that I would spend a lot of time on architecture and code design myself. Then they write the code and the tests, then exercise that code (beyond just the tests), then adversarially review it, then add documentation, then sometimes they even deploy it if I let them. So yeah, they save me a ton of time and produce significantly less bugs. I get to spend my time deciding what to build next and how to improve what I've got. commented: Yup, sometimes they do neat stuff that you wouldn't necessarily think of. I find it's a bit like pair programming in that sense. You see it do something, and it gives you ideas in turn too. commented: This article is pretty narrow. I completely agree that junk like "you're an expert software developer" is a waste of time... but that's not the same as developing skill in using LLMs. The skills you need these days include: Understanding the available models and harnesses. These are constantly changing and knowing what's changed and what they can now do that didn't work before remains extremely valuable. What are the limits of the models you are working with? How likely are they to successfully complete tasks of different difficulty levels? What tools should you make available to the models, and how? Arming a model with Playwright for browser testing unlocks all sorts of new possibilities, for example Which programming languages are appropriate for a task? Agents can compile to WebAssembly now, for example - what does that let you build that you couldn't before? Cost estimates. Something might be possible with an agent at a cost of hundreds of dollars, is that worthwhile? How can you reduce costs of what you're doing? Telling models to run cheaper subagents for example, which admittedly is one of those tricks that may well become obsolete with future model releases Honestly though, quite a lot of this comes down to software engineering skills generally. Knowing what's possible on different platforms, how to think about security (what kind of projects need higher security focus), how to break down a large project into smaller segments that are easier to plan and review. commented: Yup, very much agree with all that. commented: There's so many different types of developer jobs. The thesis statement here is probably true for lots of them. One type I've had where it's less true is doing exploratory greenfield product engineering in an early stage startup. You can automate or optimise some simple and time consuming business processes with a well-considered integration to a frontier model API. This has been the case long enough that the people with a few years of experience with it have a meaningful advantage in terms of job performance over someone who doesn't. You develop an intuition for which problems are and aren't tractable to that kind of solution, which models and effort levels to try, how to formulate prompts, the likely quality level of the results and so on. Often the "AI skill" is combining that experience with an awareness of what the business is doing and spotting the next automatable thing. Smaller startups always have more work than anyone can possibly do, so rather than putting anyone out of a job you're generally just giving someone more time for other things. Some (but by no means all) employers are already beginning to filter for this kind of thing when hiring. It's a niche, but everyone's in one niche or another. Been a couple years since I've been in that niche BTW and I'm finding I miss it, my CV and email's in my bio if someone needs a (remote EU) full-stack JS kind of person like that. commented: I've actually found LLMs are great for exploratory work because you can iterate on ideas much faster. Something that would've taken weeks or months of work before can now be tried in a day. And if you're just exploring things, it's often throw away code anyways so you don't care about quality that much. Being able to try different approaches lets you land on a much better solution by seeing how each one works. commented: If life has taught me anything, it is that acquiring skills is never worthless. You embark on a journey that builds focus and discipline and lets you discover things you didn't even know exist. Even if the exact skills you've acquired later become obsolete, part of learning takes place on a generalized or meta level and can be transferred to new endeavors. For instance, you may understand why certain design decisions have won out because you experienced what it was like to work with the now obsolete tools. Such expertise can be surprisingly valuable. commented: When I was first learning the C programming language, in my early 20s, it was common knowledge that of the two ways to increment a variable, i++ was slower than ++i. The post-increment function returned the original value of i, which meant stashing a copy of that value before performing the increment. Pre-increment avoided this wasteful operation. And by “common knowledge” I really mean folklore, because I turned 20 in 1995 and even by then, the most basic of optimising compilers knew to check if you ever used the return value of i++, and if not, skip the operation to save it. The folklore survives in C++, I think, at least for class-type iterators. With built-in integers and raw pointers the two forms compile to the same thing when the result is discarded. A user-defined postfix operator++(int) has to copy the old value and return it by value, though, while the prefix one increments in place and returns a reference. For trivial iterators like std::vector<T>::iterator an optimizer will usually erase the copy, and with optimization on you probably won't measure much, as this revisit of an old benchmark found. Debug builds are another story, which is why static analyzers like PVS-Studio still flag it++ where the result is unused. ++it costs nothing to type, so it stays my default. commented: It is not clear to me why this is on topic with the new moderation regime. I thought "ai-tools" was about the development of ai-tools, not their use. commented: My “plain language” skill is the last bulky skill that should be folded into the model. Models are terrible communicators and would improve amazingly with better communication skills. Every other skill is very specific to my work, which is how this should be. Skills should generally be pretty light weight. commented: The article is about the skill of a human using an LLM, not the markdown files called skills. commented: I recognize that, but the blog is saying AI skill is mostly people write “common knowledge” into user prompt incantations, system prompts, and SKILL.md files. My argument is the final “common knowledge” ability AIs really need to fold into their weights is communicating clearly and efficiently. Everyone triggers this regularly, either through system prompts “Be terse”, skills like caveman, or user prompts like “simpler please” or “write for a new hire”. commented: An AI company can integrate "AI skills" directly into the model, because the capacity of an LLM can be almost infinite technically. So, "AI skills" will either be too small/niche to be noticed by the AI company, or integrated into the AI models. .