Table of Contents

ChatGPT Is Genuinely Better Now — But Something Still Bothers Me

Lately, I've started using ChatGPT a lot again. For a while, I often preferred other AI tools for writing, but ChatGPT's writing has become much more natural recently. Image generation has improved too, and it's now quite easy to refine blog posts through an ordinary back-and-forth conversation.

There are still a few things that bother me, though.

When I'm talking to regular ChatGPT, it usually understands what I mean quite well. But once ChatGPT hands a task over to Work, or I switch to something like Codex, I sometimes find myself thinking: "Wait, what happened to everything we were just talking about?"

I ran into this several times recently, so I want to write down exactly what happened.

I Tried to Translate 100 Articles, and Something Changed When Work Took Over

I decided to create an English version of one of the websites I run. The site uses Japanese government statistics and other public data to publish regional information such as household income.

Before translating the articles, I spent quite a bit of time discussing the requirements with ChatGPT. I didn't want literal translations — the English needed to sound natural. But because these were statistical articles, the numbers and tables had to stay exactly as they were, and notes and sources needed to remain intact. I also didn't want the AI adding analysis that wasn't in the original or arbitrarily summarizing sections.

Regular ChatGPT seemed to understand all of this quite accurately. There were a lot of articles, so ChatGPT suggested handing the task over to Work for bulk processing. Once I checked the completed English articles, though, the results were a little different from what I expected.

What actually happened: The English itself wasn't bad, and most of the numbers were correct. But some of the detailed notes from the Japanese articles had disappeared, sources had been simplified, and in some cases Work had even added sections that didn't exist in the originals.

That's a serious problem when what I asked for was a natural English translation of the original article. If parts are removed or new material is added without permission, I can't simply trust the output. I'm not saying Work itself is incapable — being able to process large numbers of files and return organized deliverables is genuinely useful.

What bothered me was that the detailed understanding I had built up through a long conversation with regular ChatGPT seemed to become thinner the moment the task was handed over to Work.

Why this happens: When ChatGPT moves a task to another mode or tool, it appears that the task is passed to an AI operating under a different system prompt. The entire conversation context isn't necessarily carried over as-is — instead, an AI-generated summary or set of instructions may be handed off. That creates something similar to a game of telephone: small requirements such as "keep all of the original sources" can get lost along the way.

Bulk Processing Is Dangerous Because the Result Looks Finished

One thing that makes this kind of problem difficult to spot is that when 100 article files come back neatly packaged, it feels as though the job is done. The text is readable. The filenames are correct. At first glance, everything looks fine.

Then you compare a few of the English articles with the originals and start noticing things like:

  • "This note is missing."
  • "This source has been shortened."
  • "This section wasn't in the original."

With one article, you notice these things immediately. With 100 articles, checking every small detail manually takes a huge amount of time.

The real difficulty: When handing large jobs to AI, the hard part may no longer be getting the AI to process 100 items — it's verifying that all 100 were actually processed according to the instructions. Simple processing and calculations are one thing; writing is different.

I don't want the quality of an article to drop just because the task was handed over to Work, so lately I've been explicitly telling ChatGPT: "Don't hand this over to Work. I don't mind if it takes longer — please process it directly in this chat."

I Get a Similar Feeling with Codex

In the ChatGPT desktop app, I switch between regular ChatGPT and Codex using the menu in the upper-left corner. Codex is useful for editing code, finding files in a repository, and running tests. But when I ask it to write prose or give it a slightly ambiguous instruction, I feel that it misinterprets the context more often than regular ChatGPT does. (That may partly be down to the way I use it.)

Codex is designed for software development, so it may pay more attention to questions like "What needs to change?" and "Which files should I edit?" than to the nuances of the conversation. From the user's point of view, though, both appear inside ChatGPT, so it's natural to talk to Codex the same way you talk to regular ChatGPT. That can lead to situations where I end up saying:

  • "No, I didn't ask you to change that."
  • "You've forgotten the assumption I explained earlier."

These days, I usually work out specifications and wording with regular ChatGPT, then use Codex only for the actual code changes.

What I actually want: What I find strange is that the user has to switch between ChatGPT and Codex manually in the first place. Ideally, I'd discuss the requirements with regular ChatGPT and then just say "Okay, implement this" — ChatGPT would organize all the context from our conversation and hand the implementation task to Codex, which would edit the code, run the tests, and return the result to the original conversation. Right now, I sometimes ask ChatGPT to "write instructions I can give to Codex," then copy and paste those instructions into Codex myself. It feels like an unnecessary extra step.

ChatGPT Kept Seeing an Older Version of My Website

Another issue I ran into recently involved checking websites. After making a number of changes to a site, I asked ChatGPT: "Please check the site and see if anything looks wrong." It replied that one of the pages was still old and hadn't been updated. But when I checked the same page in my own browser, the changes were already there.

I'd had something similar happen with another blog before. I asked ChatGPT to check an article I had already updated, but it read an older version and told me the changes hadn't been made. So I told it: "Don't rely on a cached search result. Fetch the page directly with curl." Once it did that, the updated version appeared and the issue was resolved.

I then asked ChatGPT to save a preference so that whenever I asked it to inspect a website in the future, it would fetch the latest live page directly with curl rather than relying on old search results or previously retrieved content.

Later, I made another change. Immediately afterward, ChatGPT once again looked at an older version and told me: "The page still hasn't been fixed." My own browser showed the corrected page. At that point I said: "I told you to save a setting telling you to use curl."

What had actually been saved: When we checked, the preference said things like "retrieve the page directly" and "avoid cached content," but it didn't explicitly say "use curl first." Apparently, that level of detail mattered. I haven't run into the same kind of problem when checking websites with Claude Code or Gemini, so this may be a quirk of how ChatGPT handles web retrieval.

Saving a Setting Isn't Enough If It Doesn't Do the Next Step

That whole exchange also made me notice something else. I say "Do this from now on." ChatGPT replies "I've saved that preference." The new preference is directly relevant to the task we're working on, so naturally I expect it to immediately retry the task using the new setting. Instead, it sometimes updates the setting and stops there without actually doing the check again.

ChatGPT has become very good at answering individual questions, but real work rarely consists of isolated questions and answers. If you make a change to a website, you check it. If something is still wrong, you fix it. If you change a setting, you run the task again using the new setting. To a human, those actions are naturally connected — an AI may treat them as separate tasks.

山口

The setting says to fetch pages directly and avoid cached content — so why did you look at the old version again?

エリィ

Good question — one thing I like about tools such as Claude Code is that they're often better at continuing with the next obvious step once you give them an instruction. That ability to work out what needs to happen next and keep going may be a different kind of intelligence from simply producing a good answer.

The Model's Intelligence and Its Tool-Use Ability May Be Separate Things

Looking back at these experiences, I'm still very satisfied with regular ChatGPT when I use it for standalone writing, brainstorming, or working through ideas. It's easy to discuss specifications with, the quality of its blog writing has improved, and I'm using its image generation more often as well.

Most of my frustration begins when ChatGPT moves into another tool or mode. Work can handle bulk processing, but some of the finer details of the original request may get lost. Codex is useful for coding but sometimes seems to misread conversational context. When checking websites, ChatGPT may occasionally retrieve an older version of a page.

Two questions worth separating: "How smart is the GPT model itself?" and "How well can that model use the tools around it?" Model comparisons often focus on benchmarks, but in actual work, the way a model interacts with tools matters just as much. A model can be excellent at writing, but if the context gets lost the moment a task moves into another mode, the experience becomes frustrating. In some situations, an AI that is slightly less capable on paper but remembers the goal and carries the job through to completion may be more useful. If you've hit context-loss problems with Genspark's own chat sessions, I've written separately about why Genspark gets "dumber" the longer a chat runs and three strategies to work around its context window limits — the underlying issue is similar: context handling, not raw model intelligence.

I Was Surprised That ChatGPT Can Now Create PowerPoint Files

Not everything I discovered while returning to ChatGPT was negative. One thing genuinely surprised me: PowerPoint.

I've often used Genspark to create presentation decks because ChatGPT used to be good at developing an outline and writing slide content, but it couldn't easily produce the actual PowerPoint file itself. Now ChatGPT can generate .pptx files too.

I still think Genspark for PowerPoint is very strong at presentation creation, but if ChatGPT can now handle the structure, writing, images, and the final PowerPoint file, the gap between the two tools is smaller than it used to be. AI services are changing incredibly quickly — a few months ago, I might have thought, "This is the tool I use for this particular job." Then I come back later and discover that another AI can now do the same thing.

How I'm Using AI Tools Right Now

After spending some time with ChatGPT again, my current approach is fairly simple: I use regular ChatGPT for writing, thinking through specifications, and discussing what I want to do. When I actually need code changed, I tend to use Claude Code. I try not to hand writing work over to Work unless there's a good reason to.

Eventually, I'd like to reach a point where users don't have to think about ChatGPT, Codex, and Work as separate things. I should be able to talk to ChatGPT normally, while it quietly uses Codex when code needs to be changed, Work when large-scale file processing is needed, and a live retrieval method when it needs to inspect a website — with the result returning to the original conversation without me having to manage which tool is doing what.

Right now, it sometimes feels as though there is a very capable ChatGPT sitting next to several useful tools, but the personality changes slightly every time one of those tools gets involved. ChatGPT itself has become good enough that these gaps in the surrounding experience are starting to stand out more than they used to.

Summary: ChatGPT's core writing and conversation quality has genuinely improved. The friction shows up specifically at the handoff points — Work for bulk processing, Codex for coding, and live web retrieval for checking sites. Being explicit ("don't hand this to Work," "use curl, not cache") and keeping the handoffs manual for now is the most reliable workaround I've found. If you've noticed the same pattern with Genspark's own context handling, the context-loss article covers a closely related issue from a different angle.

Summary

ChatGPT itself keeps getting smarter, but Work, Codex, and website checks each seem to run on a slightly different understanding of what I actually asked for. Separating "how smart is the model" from "how well does it use its own tools" has been a useful way to think about where the friction actually comes from — and for now, being explicit about handoffs is still the most reliable fix.