Currently IMO you have to be a programmer to use Gemini to write programs effectively.
I've noticed beginners make mistakes like using singular terms when they should have used plural ("find the bug" vs "find the bugs"), or they fail to specify their preferred platform, language, or approach.
You mentioned your wife is using Excel, which is primarily used on Windows desktops and/or with the Microsoft ecosystem of products such as Power BI, PowerShell, Azure, SQL Server, etc...
Yet you mention she got a bash script using sed, both of which are from the Linux / GNU ecosystem. That implies that your wife didn't specify that she wanted a Microsoft-centric solution to her problem!
The correct answer here would have likely to have been to use Microsoft Fabric, which is an entire bag of data analysis and reporting tools that has data pipelines, automation, publishing, etc...
Or... just use the MashUp engine that's built-in to both Excel and PowerBI, which allows a surprisingly complex set of text, semi-structured, and tabular data processing. It can re-run the import and update graphs and charts with the new data.
PS: This is similar to going up to a Node.js programmer with a request. It doesn't matter what it is, they will recommend writing JavaScript to solve the problem. Similarly, a C++ developer will reach for C++ to solve everything they're asked to do. Right now, the AIs strongly prefer Linux, JavaScript, and especially Python for problem solving, because that's the bulk of the open-source code they were trained with.
Are you specifically talking about 2.5 Flash? It only came out an hour ago, I dont know how you would have enough experience with it already to come to your conclusion.
(I am very impressed with 2.5 Pro, but that is a different model that's been available for several weeks now)
I suspect that if Gemini is giving you bash scripts it's because you're note giving it enough direction. As you pointed out, telling it to use Python, or giving it more expectations about how to go about the work or how the output should be, will give better results.
When I am prompting for technical or data-driven work, I tend to almost walk through what I imagine the process would be, including steps, tools, etc...
Good news: Prompting is a skill you can develop.
This is one case where I've found writing code with LLMs to be effective.
With some unfamiliar tool I don't care about too much (e.g. GitHub Actions YAML or some build script), I just want it to work, & then focus on other things.
I can spend time to try and come up with something that works; something that's robust & idiomatic.. but, likely I won't be able to re-use that knowledge before I forget it.
With an LLM, I'll likely get just as good a result; or if not, will have a good starting point to go from.
I want to be able to just tell chat GPT or whatever to create a full project for me, but I know the moment it can do that without any human intervention, I won't be able to find a job.
I have a few programs now that are written in Python (2 by 3.7, one by 2.5) used for business daily, and I can tell you I didn't, and frankly couldn't, check a single line of code. One of them is ~500 LOC, the other two are 2200-2700 LOC.
Case in point: there was a post here recently about implementing a JS algorithm that highlighted headings as you scrolled (side note: can anyone remember what the title was? I can't find it again), but I wanted to test the LLM for that kind of task.
Pretty much no matter what I did, I couldn't get it to give me a solution that would highlight all of the titles down to the very last one.
I knew what the problem was, but even guiding the AI, it couldn't fix the code. I tried multiple AIs, different strategies. The best I could come up with was to guide it step by step on how to fix the code. Even telling it exactly what the problem was, it couldn't fix it.
So this goes out to the "you're prompting it wrong" crowd... Can you show me a prompt or a conversation that will get an AI to spit out working code for this task: JavaScript that will highlighting headings as you scroll, to the very last one. The challenge is to prompt it to do this without telling it how to implement it.
I figure this should be easy for the AI because this kind of thing is very standard, but maybe I'm just holding it wrong?
What do you mean by "highlight as you scroll"? I guess you want a single heading highlighted at a time, and it should be somehow depending on the viewport. But even that is ambiguous. Do you want the topmost heading in the viewport? The bottom most? Depending on scroll direction?
This is what I got one-shot from Gemini 2.5 Pro, with my best guess at what you meant: https://gemini.google.com/share/d81c90ab0b9f
It seems pretty good. Handles scrolling via all possible ways, does the highlighting at load too so that the highlighting is in effect for the initial viewport too.
The prompt was "write me some javascript that higlights the topmost heading (h1, h2, etc) in the viewport as the document is scrolled in any way".
So I'm thinking your actual requirements are very different than what you actually wrote. That might explain why you did not have much luck with any LLMs.
Yeah, you understand what I meant. The code Gemini gave you implements the behavior, and the AI I used gave me pretty much the same thing. There's a problem with the algorithm tho -- if there's a heading too close to the bottom of the page it will never highlight. The page doesn't exhibit the bug because it provides enough padding at the bottom.
But my point wasn't that it couldn't one-shot the code; my point was that I couldn't interrogate it into giving me code that behaved as I wanted. It seemed too anchored to the solution it had provided me, where it said it was offering fixes that didn't do anything, and when I pointed that out it apologized and proceeded to lie about fixing the code again. It appeared to be an infinite loop.
I think what's happened here is the opposite of what you suggest; this is a very common tutorial problem, you can find solutions of the variety you showed me all over the internet, and that's essentially what Gemini gave you. But being tutorial code, it's very basic and tries not to implement a more robust solution that is needed in production websites. When I asked AI for that extra robustness, it didn't want to stray too far from the template, and the bug persisted.
Maybe you can coax it into getting a better result? I want to understand how.
And I still don't know what you want! Like, you want some kind of special case where the last heading is handled differently. But what kind of special case? You didn't specify. "It's wrong, fix it".
Fix it how? When the page is scrolled all the way to the bottom, should the last heading always be highlighted? That would just move the complaint to the second heading from the bottom if three headings fit on the last screen. Add padding? Can't be that, since it's exactly what this solution already did and you thought it wasn't good enough.
Sorry, I will not be playing another round of this. I don't know if you don't realize how inadequate your specifications are (in which case that's your problem with the LLMs too), or if this is some kind of a bit, but either way it doesn't feel like a good use of my time.
But if your problem is that the LLMs give a bad initial answer, get anchored on it, and can't iterate, just give all the requirements up front. If they're requirements you didn't realize existed until you saw the proposed answer, just start again with a fresh context. That tends to work well for me in Aider.
Okay, but the AI understood the requirements; It confirmed to me what I intended it to do, but it couldn't produce code that met its textual descriptions. It kept returning the tutorial code.
> You didn't specify. "It's wrong, fix it". Fix it how?
Maybe I wasn't clear here but I'm not replying as if I'm prompting you like an AI. The problem domain is described better in the link in the sibling comment. When the AI gave back the initial code, I had inquired the following:
how does this work when it gets to the bottom. Let's say I have a heading at the very bottom of the page. Will the "active" one be the bottom one or the top?
What happens at the bottom of the page?
When you're near the bottom of the page and the last section (section3, say) starts to enter the top of the viewport (within 100px), it becomes the active one — exactly as you'd want.
But here's the catch: if the last section is short, or you scroll all the way to the very bottom (past the heading), then no section's top may be ≤ 100px, especially if the last heading has already scrolled off the top. In that case, the last matched section (likely the one before the last) stays active, which feels wrong.
We can fix this either by:
Making the threshold smarter (e.g., using .bottom or checking if any part of the section is visible).
Or better: use IntersectionObserver to handle this robustly.
Why this is better:
It triggers when a section is entering the viewport (around 30% from top).
Works even when you scroll to the very bottom.
It’s less sensitive to section height or viewport size.
Show me exactly what was changed from this to the last which fixed the problem
Perfect — let's do a line-by-line diff-style breakdown of what changed between the buggy version and the fixed version so we can pinpoint exactly what fixed the issue.
Why this works:
We evaluate all visible headings, not just ones intersecting a line.
We pick the one that’s:
just above the activation line, or
just below it, if none are above
Handles edge cases like top/bottom of scroll
> But if your problem is that the LLMs give a bad initial answer, get anchored on it, and can't iterate, just give all the requirements up front. If they're requirements you didn't realize existed until you saw the proposed answer, just start again with a fresh context. That tends to work well for me in Aider.
Yeah that's what I tend to do as well. I don't tend to get good satisfying results though, to the point where coding it myself seems like the faster more reliable option. I'll keep trying to hold it better and maybe one day it'll work for me. Until then I'm a skeptic.
If that's a concern, just create another account. Doesn't even require using a separate browser profile, you can be logged into multiple accounts at once and use the account picker in the top right of most their apps to switch.