Nolen began their writing career in 2019, with three years dedicated to editing the Creative section at MakeUseOf. Their expertise lies at the crossroads of technology and creativity, covering areas like photography, video editing, and graphic design.
Outside of work, you'll often find Nolen diving into a good book, writing their own stories, or playing video games.
Giving a messy spreadsheet to a chatbot seems like a job that should be quick and easy, and for the technical parts it mostly is, since spotting a formula that throws errors is what they're tuned for. But what if the numbers are fixable, and the chatbot still just doesn't seem to understand what you meant? My article tracker sheet was a good recent example of this.
I use this sheet to track my workload, such as how long I spend on an article, how many words, sections, screenshots, and so on. The gist is to give myself an overview of my working pace, and that's a context that AI struggles with more than hard numbers. Sure, fixing the technicalities will improve my overview - for example, some draft times were logged in minutes, which made one article look like it took 40 hours. This will completely skew what my pace looks like.
I gave ChatGPT, Claude, and Gemini the same file and the same prompts. Fixing the formulas is the easy part of a job like this, but what I actually wanted to see was whether any of them would notice that the numbers make no sense for how much I actually write every week.
Gemini had a strange first request
But it eventually spotted some of the problems
I went with the more advanced models for all three on medium effort level, and I currently also have the Pro plan for each one. So in Gemini, that meant 3.1 Pro, which is the heavier option compared to 3.8 Flash. Before even reading anything on the sheet, it told me it couldn't continue because my Gmail smart features were switched off and asked me to turn them on, which is a strange first request for a spreadsheet question. I declined and told it to just look at the file (there wasn't even a decline button, I had to manually prompt it, which tells me Google really wanted to make this integration happen).
It found most of the obvious mistakes, including the letter-O typo and the draft times logged in minutes. Thing is, it didn't fix any of them and instead gave me a to-do list of edits for me to make myself, with the zero-sections row getting a suggestion to change it to "at least 1" and no real thought about what the number should be.
And it also didn't get the context of what I was going for in this sheet. It counted Published and Live articles as complete, then later described everything marked Filed as a past-due backlog, which is pretty much the opposite of what Filed means to me. And every Note-taking row got lumped under Obsidian even though two of them have nothing to do with Obsidian. It did spot one of the shuffled rows, a short Excalidraw piece that suddenly had 1,426 words written in two hours, although its theory was that I leaned on automated assistance, which is a very bold and very wrong assumption. I write my own work and AI gets tasked with recalling what I've already covered so I don't repeat myself, finding sources to back up my claims, tracking analytics, and those kinds of things that would be easier for a human to miss.
Overall, Gemini was just really bad at this. It opened with wanting my email, didn't help me fix the spreadsheet, and made completely wrong assumptions about my work.
ChatGPT rebuilt everything
It was the complete opposite experience of Gemini
I went with GPT-5.6 Sol on medium effort in ChatGPT. It did run into a little trouble straight away, timing out on the first try. But it at least gave me an explanation of what went wrong and had another go with a more careful setup. It also really took its time, spending almost four minutes on it, which is about as long as it took to build a portfolio site I recently tasked it with, so I knew that it really dug into it.
It responded with a completely rebuilt workbook, with every formula error sorted and it also converted the text dates properly so my monthly and category totals finally added up properly. It actually did this on every prompt without me having to ask, which means there was always a working file for it to check its claims against.
Its first answer wasn't entirely correct; it said only 50 articles were complete. But after I nudged it, it went through every status and pointed out that if Filed means I'd finished and submitted the piece, the real number would be 57, which was exactly what I meant. Then it rebuilt the summary to show both numbers side by side, and that's a pretty neat way to handle the ambiguity of my sheet.
But when it came to the shuffled word counts, that was another story. It flagged my load-shedding piece, 1,812 words across only three sections, as an outlier and then left it alone because nothing looked corrupted, and by the last prompt it was calling that same row the biggest structural oddity in the sheet while still deciding my word counts mostly made sense. This was definitely a bit of an oversight.
Overall, ChatGPT really delivered. It regenerated the sheets with the technicalities fixed and managed to understand the nuance of the sheet to some extent.
Claude asked before I had to push
And it considered how I filled in the sheet, not just what was in it
In Claude, I used Opus 5.5 on medium effort, which is one of the newer models and also happens to be my favorite for research and analytics. One of its steps failed partway through, but it recovered without any fuss and came back with a fixed spreadsheet plus a short breakdown of what went wrong.
Instead of guessing, it noticed that Sections matched Related Links on every other row and used that to set the missing value to 4, leaving a comment on the cell telling me to double-check it. It raised Filed in its very first reply as well, asking whether I wanted those counted as complete alongside Published and Live, and on the follow-up it pointed out that Live only shows up until early September, so it was probably an older label I just stopped using.
Something the other two never did, but Claude did, was holding up "the last six weeks" against the calendar, and it noticed the data stops on 25 September, then asked me where the rest of my articles went. These are conclusions you only reach when you actually think something through and approach it from multiple different angles.
It did trip over one of my own rules, though. Related Links matching Sections on all 60 rows got flagged as data that couldn't be real, when one related link per section is just how I write most of the time. So that did end up undercutting the clever fix it made earlier. To be fair, it did admit as much, which I appreciate.
Only one of them asked what I meant before I had to push
None of these three managed to figure out that a whole block of my word counts was just shuffled, which was a little disappointing. ChatGPT came closest to giving me a fixed spreadsheet that I could use as-is, and it nailed the Filed question. But Claude understood what I meant because it asked the right questions before I had to push it. So when it comes to fixing the numbers, I'd say ChatGPT and Claude are head-to-head, but for understanding ambiguity like how my working pace is going, that goes to Claude. And Gemini wasn't even in the same race from the start.
ChatGPT
