This is Lesson #02 of the ANTIghostwriter course — a free, complete system for creating authentic content with AI assistance.
New here? Start from the full course overview.
Previous lesson: #01: The Complete AI Content Creation Tool Stack
What You’ll Learn
Not all AI models are equal. In this lesson, you’ll learn which AI to use for which task: Claude 4.5 for writing (it maintains your voice better), ChatGPT for research (Deep Research feature), and options for image generation. We’ll also cover budget alternatives and why paid versions are worth the investment for serious content creators.
Time to complete: ~15 minutes to choose and subscribe
Choosing AI Models
Different artificial intelligence models have distinct characteristics. Some are well-suited for specific purposes, while others are trained on different datasets, making them more effective for particular tasks. Understanding which model works best for each purpose is crucial for getting optimal results.
In our case, this is verified purely through empirical testing. You can experiment with various tools yourself. While they are all interchangeable, the quality of output will vary between different models because they are trained on different datasets and have different parameter settings that affect the results they produce. This is why careful selection matters.
You can go through this entire evaluation process independently and compare different models to see which ones work better for specific tasks. However, I’ve already done this research for you, so you can simply use my recommendations. If my approach doesn’t suit your needs for some reason, feel free to conduct your own testing.
Here are my findings:
For Writing Tasks
We need a tool that excels specifically at writing. Currently, the Claude Sonnet 4.5 model from Anthropic is the best choice for this purpose. This is the LLM we’ll be using for writing throughout this course. It produces superior results, understands context exceptionally well, and works effectively with long documents—which we’ll definitely need given our context requirements.
Important note: Anthropic updated Claude Sonnet to version 4, and it writes much worse than version 3.7. The text has become dry, more concise, and doesn’t consider many aspects of the given context. Based on my observations, I’ve written articles multiple times, started fresh chats several times, and requested revisions, but consistently ended up with dry, robotic text. As soon as I switched back to version 3.7, I got text that sounds like my own writing—my phrases, my words, and my thoughts, complete with all the nuances, and it considers all the prompt requirements. I’m sticking with version 3.7 for now.
New update: Claude Sonnet 4.5 writes good-quality content, and it seems to follow prompt instructions pretty well. So for now, I’m sticking with 4.5.
The context window essentially functions as the AI’s memory. Within a single chat session, we’ll be interacting with the AI multiple times. Claude works by reading the entire context anew each time, which makes it a relatively expensive model to use. However, if you handle everything carefully (which I’ll explain), it will be more than sufficient for our needs.
Claude remains highly consistent in maintaining its voice. When you provide reference materials that demonstrate your writing style, Claude manages to maintain that tone effectively. If the reference materials remain the same and don’t change, the consistency of the output text also remains constant—which is exactly what we need.
For Research Tasks
To write comprehensive articles for newsletters, YouTube scripts, or videos, you need to conduct thorough research using various materials. I prefer basing my research on studies and statistics that either support or challenge the ideas I’m expressing.
Research also allows you to find relevant quotes from various people that fit your context. At minimum, research performs the key function of confirming or refuting the chain of thoughts you express in your articles.
Unfortunately, Claude doesn’t yet have a Deep Research function, but this feature exists in ChatGPT, which I recommend for research purposes. Perplexity also works well for this in its latest versions. However, ChatGPT remains my personal favorite for research tasks—it has consistently produced the best results for me in practice.
For Image Generation
MidJourney remains the top choice today, working exceptionally well with various types of images. If you need image generation capabilities, this is an excellent option.
Purchasing LLM Models
I mentioned on the landing page for this course that we’ll need to purchase these tools and upgrade to their paid versions. Free versions don’t provide the necessary flexibility and complete feature set. Additionally, they have severely limited context windows and reduced output volumes.
To create content consistently on a daily basis, free versions definitely won’t be sufficient. If you want to try this entire system as a learning exercise first, you can certainly start with free versions to get familiar with the process.
Alternative Options
If you have an X subscription, you can try using Grok as an alternative to ChatGPT—it’s included with paid X subscriptions. There are other free alternatives that can work with similar tools if you can’t use paid versions for some reason. For example, you could try DeepSeek as an alternative to ChatGPT, which is free.
However, there are important considerations. DeepSeek is made in China and is available for download. If you use the online version, working with it is similar to working with ChatGPT, but it can also be installed locally. We won’t cover these local installation approaches in this course, but you can research them yourself if needed. You can actually ask these LLMs themselves (like ChatGPT or DeepSeek) how to install them locally on your computer for these purposes.
Claude Alternatives
For Claude, I don’t currently see a viable free alternative. You can use free models with the same prompts we’ll be using, but I can’t guarantee their quality. Claude is the model that delivers the needed results with proper prompting and setup.
When used correctly, Claude will produce results where the presence of artificial intelligence in the text becomes undetectable—signs of AI-generated text will be absent. This comes from using appropriate techniques, which I’ll cover in this course.
Therefore, purchasing Claude is absolutely mandatory for this course, and there’s no alternative for me personally at the moment. This is the key requirement for achieving the core results of this course, so I wouldn’t recommend trying to save money here by looking for alternatives that will clearly be weaker at this time.
The market is changing rapidly and everything could change, but this is the current state of affairs as of summer 2025.
MidJourney Alternatives
MidJourney is a relatively expensive tool and quite limited in terms of request quantities, but there is a free alternative: Stable Diffusion, which is either free or much cheaper to use.
Stable Diffusion is also an AI model for generating images, and it can be installed locally on your computer for completely free use. There are numerous online tools that allow you to make requests and generate images using Stable Diffusion. Some are paid, others are free.
You can search for and use these tools for image generation depending on your specific needs. Personally, I use Stable Diffusion installed locally, which uses the Flux model—currently one of the most comprehensive and well-trained models that produces excellent results with great stylistic consistency. I generate images for my articles using this setup.
Image generation offers perhaps the most flexibility among all our tool categories. Some people might not create images at all, others might create them manually, and still others might purchase them from stock platforms. There are many more options here than for research or text generation.
Paid Wrappers
A wrapper is essentially a program that contains paid versions of all these LLM tools under the hood. Since they use a different mechanism through APIs (developer channels), they allow you to use these models more cost-effectively.
Essentially, these services purchase capabilities in bulk and resell them wrapped in their own interface. Many wrappers exist, and you can use them if you want to save money significantly, but there are several important considerations.
For the cost of one model (or even half of one model), you’ll get access to several models at once, but you’ll be limited to only the functions provided within that wrapper. For example, a wrapper might not have the Deep Research function, so definitely check the feature set before purchasing—this could become a significant limitation.
Since ChatGPT is needed for Deep Research, some wrappers simply may not have this functionality, even though it exists in ChatGPT itself when you use the original version or download the official app.
Another example: the context window might be smaller than in the original Claude.
System Prompt Limitations
Another important limitation I consider necessary to highlight: wrappers don’t have the ability to set a system prompt that applies to all chats. You can set your own requests and try to include a system prompt within them, but almost every original tool (ChatGPT and Claude, for example) has a separate setting that allows you to set a preset for any dialogue with the LLM, which improves results.
This is one of the techniques I use, which I’ll cover later in this course. If you choose to save money by using wrappers, you’ll need to manually transfer all the settings I use in the system prompt into the individual prompts you’ll be using in that wrapper.
For the reasons described above, I can’t recommend specific wrappers because I don’t use them myself—I use the LLMs directly. Based on my observations (I did test one wrapper and compared the results), it seemed to me that the original Claude produces better results, though I acknowledge I might have some bias.
Additional Features
There are functions like canvas or document artifacts that allow you to work very conveniently with documents that Claude produces as output. This creates a separate text document you can work with later, or a code document if you’re writing code. Claude can refine these documents by adding text or making changes according to your requests.
Instead of rewriting the entire text from scratch (as happens in flat chat mode), it can modify individual parts within the document. This is very convenient, clear, and works faster. At least some wrappers definitely don’t have this functionality—I don’t know about all of them since I haven’t tested them comprehensively.
Update Speed
The final argument in favor of original versions is that they naturally update faster than APIs. As soon as a new function or improvement appears, it first appears in the tool itself, then in the API, because API updates are always a secondary task—companies primarily focus on selling their main product.