Yeah, I saw that announcement but still can't figure out what the actual impact is - doesn't change anything for me (my non-skill slash commands still work).
3,486 karma · joined January 7, 2023
So yeah, I agree that it's all just documentation. I know there's been some evidence shown that skills work better, but my feeling is that in the long run it'll fall to the wayside, like prompt engineering, for a couple of reasons. First, many skills will just become unnecessary - models will be able to make slide decks or do frontend design without specific skills (Gemini's already excellent at design without anything beyond the base model, imho). Second, increased context windows and overall intelligence will obviate the need for the specific skills paradigm. You can just throw all the stuff you want Claude to know in your claude.md and call it a day.
On the verification front, a few examples:
1. I built an app that generates listing images and whitebox photos for my products. Results there are verifiable for obvious reasons.
2. I use Claude Code to do inventory management - it has a bunch of scripts to pull the relevant data from Amazon then a set of instructions on how to project future sales and determine when I should reorder. It prints the data that it pulls from Amazon to the terminal, so that's verifiable. In terms of following the instructions on coming up with reorder dates, if it's way off, I'm going to know because I'm very familiar with the brands that I own. This is pretty standard manager/subordinate stuff - I put some trust in Claude to get it right, but I have enough context to know if the results are clearly bad. And if they're only off by a little, then the result is I incur some small financial penalty (either I reorder too late and temporarily stock out or I reorder too early and pay extra storage fees). But that's fine - I'm choosing to make that tradeoff as one always does when one hands off work.
3. I gave Claude Code a QuickBooks API key and use it to do my books. This one gets people horrified, but again, I have enough context to know if anything's clearly wrong, and if things are only slightly off then I will potentially pay a little too much in taxes. (Though to be fair it's also possible it screws up the other way, I underpay in taxes and in that case the likeliest outcome is I just saved money because audits are so rare.)
The most fun one is this, which creates listing images for my products: https://theautomatedoperator.substack.com/p/opus-45-codes-ge...
More recently, I'm using Claude Code to handle my inventory management by having it act as an analyst while coding itself tools to access my Amazon Seller accounts to retrieve the necessary info: https://theautomatedoperator.substack.com/p/trading-my-vibe-...
Don't care about code quality; never seen the code. I care if the tools do the things I want them to do, and they verifiably do.
Could be for good reasons (e.g. they're security features that are important to the business but add friction for the user) or just because management is disconnected from the reality of their employees. Either way, not necessarily the wrong decision by the PM - sometimes you've gotta build features fast because the buyer demands them in a certain timeframe in order to get the contract signed. Even if they never get used, the revenue still pays the bills.
The design of it is based on the intention of the people creating it, not the actual outcome, and it's pretty clear from all available information, plus a general understanding of incentives, that it's designed to be as accurate as possible, even if it does make errors.
That's why an LLM is good for bookkeeping for me - I'm not actually using the books in any strategic capacity, and I know the numbers in the business well enough that I'll catch any obvious mistakes. The books are purely to give to my CPA to do my taxes at the end of the year.
But in any case, tremendous errors won't actually cause me big problems, because I'd catch them on my review of the P&L. Small errors I might miss (whether committed by human or LLM), but they're not going to be material anyway.
It's always been free, but because of a change to the way Amazon charges third party app devs, they were going to start charging next month. Since the whole app is just a couple of API calls and storing a record of which orders you've sent the request to already, Claude Code built it in 5 minutes.
In general, the Amazon Seller UI is a cluster (especially since I have one account for each brand, so I constantly have to switch between them). There are lots of subscription apps to make your Amazon data more useful and accessible, but Claude Code with access to the Amazon APIs pretty much replaces all of them. I spend very little time in the actual Amazon UI now and mostly just ask my trusty assistant for the info that I need.
You still write the code! We do lots of things that could be done more efficiently. People build their own furniture and make their own clothes and brew their own beer. If you love doing something for its own sake, keep doing it.
I suspect the author of the post would agree. This feels much more like a experiment to push the limits of LLMs than anything they're looking to seriously use as a product (or even the basis of a product).
I think the more interesting question is when the approach of completely autonomous coding will be the right way to go. LLMs are definitely progressing along a spectrum of: Can't do it -> Can do it with help -> Can do it alone but code isn't great -> Can do it alone with good code. Right now I'd say they're only in that final step for very small projects (e.g. simple Python scripts), but it seems like an inevitability that they will get there for increasingly large ones.
But these days I run a one-man business, and LLMs (currently Claude Code, previously GPT) have written me a ton of great code. To be clear, when I say "great" I don't mean up to your standards of code quality; rather, I mean that it does what I need it to do and saves me a bunch of time.
I've got a great internal dashboard that pulls in data from a few places, and right now CC is adding some functionality to a script that does my end of month financial spreadsheet update. I have a script that filters inbound leads (I buy e-commerce brands, generally from marketplaces that send me an inordinate amount of emails that I previously had to wade through myself in order to find the rare gem). On the non-code front, I have a very long prompt that basically does the first pass of analysis of prospective acquisitions, and I use Shortcut.ai to clean up some of the P&Ls I get (I buy small e-commerce brands, so the finances are frequently bad).
So while I can't speak to using LLMs to write code if you're working in any sort of real SaaS business, I can definitely say that there's real, valid code to be had from these things for other uses.
Anyway, not a great comparison because you're talking about legal regulations governing speech on the internet. This isn't a jaywalking ticket, it's a deeply complex regulatory issue involving politics, law and international relations. It's also an issue that the current administration has shown interest in, so if you're an affected American business it would be pretty foolish not to seek help there.
In this case, I'm the American taxpayer who is paying for all of this food, and, perhaps more importantly, paying for all of the medical treatment they receive because of the consequences of these choices.
When your consumption is being paid for by other people, it's perfectly reasonable for those people to put limits on your choices, especially when they're footing the bill for the consequences of any bad choices you make too. We're a wealthy country and shouldn't let people starve, but you don't need ice cream or Coke or Pringles not to starve.
If the alternative to using Ozempic is eating a healthy diet and exercising regularly, then sure, the latter is better, but the target population for it is people who have spent many years not eating healthy or exercising and who are unlikely to start in the near future.
2025: My first full year of running my own business, and things really went well. Also had my second child, a beautiful daughter who is currently not sleeping well at night. I am so tired. I lost my dog suddenly - he was my first dog and my companion for nine years. I love that guy so much.
This seems solvable by passing a law that makes the officer legally responsible for the report as if he had written it. He doesn't get to use this excuse in the courtroom and it gets stricken from the record if he tries. That honestly seems like a better solution than storing the original AI-generated version, because that can reinforce the view that AI wrote it to jurors, even if the officer reviewed it and decided it was correct at the time.
Honestly, it takes such a relatively small amount of time that it makes sense to just do it myself until there's an agent that can easily handle it; I'm really only spending time trying to automate it now as a test of AI capabilities. If I actually wanted to get it automated tomorrow, the most time-efficient way to do that would just be to involve a VA from somewhere cheap for the work I'm doing.