There is so much hype, but after spending a truck load of time trying to get comparatively simple things done, I’m getting increasingly disillusioned. One off tasks great, but try to get automation going is frustrating. I sent this to my friend this morning after my morning brief failed again after another attempted fix: “don’t bother, I have wasted enough time on this. I had such high hopes, and it is even identified by Anthropic as a good use for Claude, but it is so unreliable as to be useless. If I got in a car , turned the key and it didn’t start, just once, I would be looking for a new car, or at least a permanent fix. This car hasn’t started reliably for months, even after the fixes. It’s a lemon”
Second thought, does anyone else find it really meta, that content creators seem to increasingly creating content about using AI to - wait for it - create content. I hope to one day find a content creator who will create some content about doing “real” work.
Sure. It was a Claude Cowork scheduled task which we evolved together over a couple of weeks. The problem is not in being “able” to achieve the outcome, it is in doing so reliably. The task was to obtain a list of reminders, a list of received emails, a list of calendar events and populate a daily brief. Worked sometimes, but not always, which is worse than useless. Run it as a standalone task and it works “mostly”. Run it unattended and it workes sometimes, but"mostly" fails. The HTML output looks VERY pretty, but if you can’t get your schedule for the day reliably then it’s just as easy to look up Calendar, Reminders and Mail yourself while having your coffee. Maybe I’ll have more luck with baked in Siri AI when the time comes.
The hype around AI is ridiculous so it misleads people into thinking it is great at everything when it clearly is not. Like any tool, it is really good for some things, but bad at others.
For your specific example, I’d say the issue is probably that AI is incredibly bad with dates and numbers. Which is ironic because computers are traditionally good at these things. But you have to remember that at the heart of all of these AI tools is an LLM. And an LLM is really good with words and predicting the next token.
So don’t ask it for things like “give me the reminders due today”, it will get confused on what today means. You have to be specific and use a traditional tool like a Python script to calculate today’s date and feed it “give me the reminders due on 7/16/2026”
This is a very important point, but I would say combining it with one more point sums up the whole subject.
A hammer is a fantastic tool for building a house. But if you don’t know how to build a house or don’t know how to wield a hammer, then a hammer is useless for building a house.
@jcarucci’s point stands, but even for things it is good at… you need to know what you’re doing.
I’ve been using Claude for building some iOS and Mac apps. I’ve got one on the App Store already and another in (hopefully final) review. Both of these are not super complex (though well featured) and each has taken (curiously) 71 build iterations to get to a releasable state.
A handful of those iterations were of the “it doesn’t even compile” type and a signficant number were “feature building”, but an awful lot were fine tuning the behaviours, both visible and under the covers. I happen to have been writing code for 45 years so this back-and-forth makes total sense to me.
Apparently I am fortunate, because I have been able to have Claude (I have the Max plan) create a weekly brief for me. It pulls from multiple sources, including the weather, email, Apple Reminders, Google Docs, and Calendar, to produce a nicely formatted result.
Ironically, after using the weekly brief for several weeks, I concluded I do not need it. I have everything I need in Apple Reminders, email, and my calendar. It was, however, a useful and successful experiment, and I have since deleted the project from Claude.
Extensive Reminders Project
Additionally, I was able to have Claude create an extensive list of one hundred seventy-two tasks with dates directly in two Apple Reminders lists for two related projects. The tasks span a two-year time frame and include due dates and notes for each, some of which also include links that Claude provided. What was particularly useful is that Claude conducted extensive research on the project and supplied the list of tasks, notes, and web links needed, which I reviewed and revised before having Claude create them in Reminders. This project, from the first prompt to my final approval, took about two hours.
This reminds me of the saying “when all you have is a hammer, everything looks like a nail” which is also very applicable to some of the AI hype and/or content creation these days.
The only really solid use I have found for the LLMs is debugging code that is not working and I can’t work out why. My usual prompt is “are there any obvious errors in this code”, to which I get a reply of “you’ve missed the semicolon on line 12”. In other words, it’s useful as hypothesis-generation in optimisation phase, not as de novo idea generation.
The problem is ‘AI’ is not just one thing. Which makes your statement too ambiguous. For automation it can be problematic with many moving parts, permissions and unknowns. My experience has been positive and I’ve just built my first Electron app.
Indeed. We must be careful not to absolutize when editorializing about AI. To say it is “bad” or “good” is too simplistic. It depends on the AI platform, the model, the quality of the prompts, the nature of the project or task, and more. I have found AI extremely capable and helpful in some instances, and useless in others. As they say, “it’s complicated.”
This has been my main intentional use of AI. I’ve been tasked with redoing a lot of scripts into C#, which I don’t know well. So basic syntax checking, or even basic rewriting to C# has been very useful.
Converting code from one language to another has been…spotty at best.
Interesting, that is one of its best uses for me. I have a complex iOS app, many asked for an Android version. CC made a marvelous conversion that works and conforms to Android internals and UI. But its not “here is ios code, convert to android pls”, guidance and boundaries are needed and the iOS version was accompanied with all the tickets and PRs CC had access to also.
Mines often what I would think of as very simple – vbscript to C#. As mentioned, for fixing syntax errors I don’t see, or missing arguments, it’s great.
But asking for a flat rewrite of a vbscript to C# ---- not as useful. So far it would be faster for me to rewrite, and then use AI to help find errors in my rewrite.
I’m just using co-pilot, so that may be where the difference is. I might have different results with a different AI.
I have a question about your reminders task list. You put in tasks for a 2 year project? How do you know what to work on now with such a large number of tasks in Reminders?
In reality, this is one project with two main sub-projects. Not all of the tasks in each have a specific date. In most instances, I had Claude set task due dates to the first workday of the month in which they need to be completed. When those appear in my Today list, I change the due date, if needed, to a specific date that month. The two lists I refer to above are unique in that they necessarily span two years with meaningful consequences if certain deadlines are missed. Most of my projects are far simpler with much shorter timeframes. This was a unique project and one for which Claude was especially helpful.
