Today I want to explain why I’m moving from ChatGPT to Claude for more of my online work. This is a decision I have been thinking about for some time, and it comes after repeated problems with straightforward instructions, inconsistent results and incorrect information.
I have used ChatGPT extensively since late 2022, so I am approaching four years of using it. In many respects, it has been brilliant. However, when a tool regularly makes a simple job more difficult, there comes a point when you have to try something else.
This article accompanies my video, “Why I’m Moving From ChatGPT To Claude”. I want to share the actual problems I have experienced, explain what I hope to gain by trying Claude, and be honest about what I have yet to find out.
ChatGPT Has Been A Big Part Of My Online Work
I use AI for a wide variety of tasks connected with my websites, YouTube videos and social media. That includes planning content, working with video transcripts, drafting descriptions, creating images and discussing ideas for my online business.
When ChatGPT follows the instructions and gives a useful answer, it can save a considerable amount of time. I have produced a great deal of content with its help, and I do not want to overlook that simply because I am frustrated with it now.
I have also tried to organise my work properly. For example, I have written about how I use ChatGPT Projects to manage real work. Keeping related work together makes sense when you have different websites, audiences and branding requirements.
The difficulty is consistency. A task may work well for a long period, and then the same instruction starts being ignored. That makes it hard to build a dependable daily routine around the tool.
Ten Attempts To Get Plain Text
One recent example sums up the problem very clearly. I wanted a YouTube description, tags and hashtags in plain text, ready to copy into the appropriate fields. I made that requirement clear in my original request and repeated it when the response was unsuitable.
In the conversation I showed in my video, I counted ten attempts before I received the text in the form I needed. Earlier, I had described the problem as asking eight times, but when I went through that example on screen, I counted ten.
This was a routine content task. I already knew what I wanted, and I had explained how I needed to receive it. Yet I ended up spending time repeating the same instruction instead of getting on with publishing my video.
For my workflow, plain text means copyable words without unwanted formatting or extra material that I have to remove. I want to paste the description into YouTube, or a post into Facebook or LinkedIn, without having to repair it first.
The frustrating part is that ChatGPT has managed this successfully before. When something has worked for months, it is reasonable to expect that the same clear instruction will continue to produce a usable result.
The Cost Is More Than A Few Extra Messages
A single correction may not sound particularly serious. However, I use these tools frequently, and small interruptions soon add up. If I have to check the formatting, repeat the request and then check again, the task takes far longer than it should.
It also breaks my concentration. I might be preparing a video, uploading it, adding the description and then creating the related social media posts. Having to argue with the tool in the middle of that process interrupts the whole sequence.
I pay for software because I want it to help me work more effectively. The subscription cost matters, but so does the time I spend correcting its output. A tool that looks inexpensive can become costly if it continually creates more work.
Repeated Images And Unusable Collages
Image creation has been another source of frustration. I have been creating motivational pictures for Facebook, as well as scripture images for my Christian content. These are intended to be individual images that I can publish separately.
I deliberately keep related requests in the same conversation and ask for fresh backgrounds, different scenes and new wording. Despite that, I have seen similar images and repeated material appear quite quickly.
With the scripture images, there is a huge amount of material available to choose from. It is therefore disappointing when familiar verses or very similar scenes start appearing after only a small number of requests, even though I have asked for variety.
I have also asked for more individual motivational pictures and received several designs combined into one image. That does not fit the way I want to use them. I need separate pictures for separate posts, and a collage creates another job for me to sort out.
I have shown some of the useful possibilities in my article on creating a YouTube thumbnail using ChatGPT. The potential is there, which makes the repeated failures to follow the requested format especially frustrating.
Branding Should Be Part Of The Finished Result
For Roy Bretton Online, I want the correct logo and website address included when I request a branded image. I also have other content with different branding, so the details need to match the website or channel the image is for.
When a logo is omitted or the wrong details appear, the image needs another round of work. The same applies if a traditional YouTube thumbnail is produced in the wrong shape or an unwanted person appears in the design.
I appreciate that supplying the original logo may be necessary for a particular image task. My concern is that the requirements I provide should be followed. I should be able to review the finished design rather than repeatedly discover that a requested element is missing.
Confident Answers Can Still Be Wrong
Incorrect information is a more serious problem than unwanted formatting. One example involved Camtasia, which I have used for many years and which I was using to record this video.
I asked about a teleprompter, and ChatGPT told me that Camtasia had a built-in feature for that purpose. It gave convincing instructions about where to look and which settings to use, even though I had never encountered the feature in my own use of the software.
When I could not find it, the conversation continued with more instructions and explanations. Eventually, it acknowledged that the feature it had described was not there. The menus and steps had sounded plausible, but they did not help me accomplish the task.
I have covered that experience separately in I Use AI Every Day, But Here’s The Problem. It is an example of why an answer’s confident tone is not enough to establish that it is accurate.
For someone producing tutorials, that matters. If I repeat an incorrect instruction in a video or blog post, I could waste my viewers’ time as well as my own. I want my content to reflect what actually works when I test it.
I Want Constructive Advice, Not Constant Agreement
Another issue is the tendency to agree with my suggestions. Sometimes I ask whether an idea is sensible because I want help thinking it through. A useful answer should consider the practical advantages, possible problems and alternatives.
Instead, I have sometimes felt that ChatGPT accepts the idea too readily. I then have to raise the weakness myself before it offers a more balanced answer. That reduces the value of asking for advice in the first place.
I do not expect AI to make my decisions for me. I am responsible for my own business and the content I publish. However, I do want it to point out something I may have missed and explain why another approach might work better.
A polite disagreement can be far more useful than a reassuring answer. If I am overlooking something obvious, I would rather hear about it before spending time on the idea.
Why Does The Experience Keep Changing?
In my video, I wondered whether updates might be connected with the changes I have noticed. That is my suspicion, rather than an explanation I can prove. I do not have access to the information needed to identify the cause.
What I can describe is the experience at my end. Instructions that were followed previously stop producing the same result, and familiar tasks become harder to complete. Whatever the reason, I still have work to get done.
I have tried practical adjustments before. My post about one small change that made ChatGPT more reliable covers my experience of using it in a different browser. That article concerns loading and responsiveness; the problems I am discussing here concern the answers and finished content.
Why I Am Giving Claude A Try
I already have a free Claude account, but I have very little experience with it compared with ChatGPT. I intend to try the paid version and use it for the kinds of tasks I do regularly.
People on some of the platforms I use have spoken positively about Claude. That has encouraged me to investigate it, but their experiences do not tell me how it will perform with my own instructions and working routine.
At this stage, I am looking for a more dependable experience. I want usable drafts, closer attention to my requirements and answers that acknowledge uncertainty where appropriate. I will need to judge those things through actual use.
I have also heard concerns about usage limits. I have yet to establish how those will affect my work, so that will be part of the trial. A tool needs to fit both the quality of the work and the amount I need to do.
What I Will Look For In A Fair Comparison
A useful comparison will involve giving both tools similar tasks and checking the results against the same requirements. One impressive answer will not be enough. I want to see what happens over a normal period of creating and publishing content.
- Does the tool follow my requested format on the first attempt?
- Does it keep the writing consistent with my audience and instructions?
- Does it distinguish information it knows from something it is guessing?
- Does it offer helpful criticism when an idea needs improvement?
- How much editing and correction does the finished draft require?
- Do its usage limits fit the amount of work I do?
I will assess the features available in each tool separately, including images. A better experience with writing would not automatically establish that every other part of my workflow can move across in the same way.
Moving My Information Across
In the video, I show myself requesting an export of my ChatGPT data. My aim is to keep useful information from the work I have already done and use relevant material when setting up my work elsewhere.
I have not yet completed that process, so this article is not a step-by-step migration tutorial. Requesting an export is one part of the preparation; I still need to establish how best to organise and use the information with Claude.
A practical starting point will be a concise summary of my websites, audiences, writing preferences and recurring requirements. I will also keep the original branding files and relevant project material ready to supply when needed.
Will I Still Use ChatGPT?
Yes, I expect to continue using ChatGPT while I test Claude. It has been useful for many tasks, and I want to make the decision based on what works in practice rather than frustration alone.
I may eventually cancel my paid ChatGPT subscription, but I have not reached that conclusion yet. First, I want enough experience with Claude to decide whether it offers a meaningful improvement for the work I actually do.
There may also be reasons to use different tools for different tasks. My priority is getting reliable help and producing useful content, rather than remaining loyal to a particular AI platform.
AI Needs To Support Real Experience
These experiences have reinforced the importance of bringing my own knowledge to the content I create. AI can help me organise an explanation or prepare a draft, but I still need to check that the result represents what I have done and what I know.
I have discussed the wider issue in Why AI Slop Doesn’t Work Anymore. For me, useful content comes from practical experience, honest observations and a willingness to check the details.
That is also how I intend to approach Claude. I will test it, correct it where necessary and report what I find. Changing tools does not remove my responsibility for the material I publish.
What Has Your Experience Been?
I am moving more of my work towards Claude because repeated problems with ChatGPT have become too disruptive. Asking ten times for plain text is one example, but the wider concern is the amount of time spent correcting tasks that should be straightforward.
I still recognise how helpful ChatGPT can be. That is why this decision has taken time. I now want to find out whether another tool can give me more consistent support, and I plan to share a follow-up once I have properly tested Claude.
How are you finding ChatGPT? Does it follow your instructions consistently, or have you experienced similar frustrations? If you use Claude, I would also be interested to hear what works well for you and what its limitations are. Please leave your thoughts in the comments below.