As a professional writer, I was immediately enthusiastic about ChatGPT, but over time I also learned the limitations of these tools. In this post, I'll show you how I use them, from brainstorming to researching and distributing copy. I'll give you tips on good prompts and explain how you can gradually achieve better results with your AI texts.

I have often tried AI writing tools in recent years and have always found them disappointing. What they produced was vaguely reminiscent of text, but the results were hardly usable. They rarely had a meaningful and comprehensible structure.

This changed abruptly with ChatGPT. Its responses were amazingly good. This "chatbot" finally responded the way I had long hoped it would. In everyday use, however, there are still weaknesses. For example, the writing style is boring and interchangeable by default. This is no wonder, as it represents an average of the training material. However, this can be improved with the right prompt. More on this in a moment.

Another problem is "hallucinations": tools like ChatGPT are trained to give human-sounding answers and not to check facts. You need to be aware of this. The providers are also aware of this weakness and have now made improvements. The right prompt can also help here.

In the following, I will give you tips and suggestions on how you can use ChatGPT, Claude and similar services to come up with ideas, research topics, create outlines, write articles, improve them and disseminate them. As an introduction, please also read our other posts on the topic:

• Content & AI, the basics
• Hacks & prompts for ChatGPT & Co.
• Use of AI in digital agencies and B2B web design

Find ideas for AI texts

As already mentioned in the previous part of the series: tools like ChatGPT are ideal for getting you out of your creative bubble or a mental dead end. They offer you a different perspective, are always available and always ready to help.

They are therefore no substitute for exchanging ideas with colleagues and other people. However, they can complement this very well and help you find ideas. Just make sure that the tool of your choice knows exactly what you want. You should definitely mention which target group the topic ideas are aimed at.

Or export an overview of previous posts from your CMS and ask for further suggestions. This helps to tailor the answers more specifically to your needs. As a general rule, if you ask in general terms, you will only get general answers. I'll explain what makes a good prompt in more detail in a moment.

Find examples and sources

As mentioned at the beginning, ChatGPT and others are not particularly good at checking facts. The situation is different, at least in part, with Bing Chat. Microsoft's offering uses a variant of ChatGPT and combines it with its own search engine.

When you ask Bing Chat a question, the assistant searches the web for suitable information. In the next step, it summarizes it for its answer and links each individual piece of information to the corresponding source. If it can't find anything, Bing Chat will let you know.

As with ChatGPT, you can then continue the conversation as you wish and ask for further details, ask a comprehension question or modify your search. The answers from Bing Chat can be quite detailed. I've had searches that resulted in more than a hundred sources being cited in the footnotes. You can then manually ensure that the information is actually correct, or you can use these links as a starting point for your own research.

Another example is Google Bard. There is now a button to check the AI's statements in a second step via web search. If no suitable evidence is found, the corresponding text passages are highlighted in red.

Google's approach is commendable, but unwieldy in its current form. After all, I first get a text that reads interestingly and then, in the second step, I may realize that there are no sources for a lot of information. Microsoft's approach with Bing Chat is much more logical and sensible: first search for the information, then summarize it in a text.

Create an outline for a topic

Another task for which I use tools like ChatGPT: creating outlines for planned content. What the tool gives me is not always perfect straight away. But over the course of a short conversation, I can improve it together with the AI or simply adjust it manually.

Sometimes I give the AI the material I found during the research. A service like Claude is ideal here, as it has a very large context window of 100,000 tokens. This means that even extensive documents can be read in and used as a starting point. GPT-4 has an overview of up to 32,000 tokens at once.

Such outlines are not only suitable for articles or blog posts. You can also have a concept created for a presentation or a video. This can be particularly useful for content formats that you are not yet very familiar with: first have best practices explained to you and then develop the specific outline for your topic in the same chat window.

AI texts: Tips for good prompts

An important basic rule for good prompts for AI offers such as ChatGPT or Claude: Be as precise and detailed as possible. This includes, for example:

• Give a signal about the capacity in which you need the AI. For example: "Act as a professional blogger."
• Explain the content format (blog posts, specialist articles ...)
• Name the target group (existing customers, interested parties, followers ...)
• If necessary, also specify the language level. A text for non-professionals should be formulated very differently from one for professionals.
• The tonality is also part of it: factual, informative, entertaining ...
• It may be useful to say something about the objective of the content.
• And for more complex tasks, a tried and tested tip is to provide the AI with a successful example.

These points increase the likelihood that the results will be close to what you need. However, you shouldn't expect a perfect text, but at least a good first pitch.

If you limit the topic, target group, language level and tonality as described, this should also improve the AI texts overall: they are then tailored more precisely to your needs and are no longer as boring and interchangeable as the standard results.

If the output text does not fit properly, you basically have three options:

1. If it is halfway usable, copy the result and then generate a completely new variant. This makes sense if you want to have several options to choose from. Perhaps the next result will be better or you can take set pieces from several and combine them.
2. Adjust your prompt and then try again. This is a good idea if you can see exactly why the result doesn't fit.
3. Ask ChatGPT for improvements and adjustments as the conversation progresses. This will allow you to get closer to the result and learn which instructions the tool understands and can implement and which it cannot.

You can also tell the tools which sections you are not yet satisfied with and why. Over time, you will realize that certain problems can only be solved afterwards through manual editing. Ultimately, the text quality and flexibility of the assistant still have their limits. They are just much wider than in the past.

Incidentally, this also depends on the tool you choose. Claude, for example, understands language hints much better than GPT 3.5. You should therefore always try out whether a new tool can immediately cope with your tasks and not assume from the outset that they all work the same way.

Writing AI text step by step

If you want to write an entire post with a tool like ChatGPT, it has proven to be a good idea to do this step by step. Because while these helpers already do a good job with short AI texts, long blog posts or even white papers are otherwise too much of a challenge for them.

For example, you can provide a previously created outline and, if necessary, research material and then proceed section by section with a precise prompt. In my experience, this improves the quality of the text. It is also much easier to ask for improvements and changes to a short section than to a long article.

The disadvantage of this method can be that your assistant repeats itself, as it does not pay as much attention to previously written sections as you would like. ChatGPT also has the peculiarity of almost always wanting to write a conclusion. Either you try to prevent this via the prompt or you simply always delete the last paragraph.

Claude is much better here and, in my experience, writes more naturally. It is to be expected that these tools will become noticeably better in the coming months and years - especially if they specialize more in certain tasks. Copy.ai is already available for advertising copy and NovelAI for novelists. We can expect to see much more variety here in the future.

Provide feedback and improve AI text

Once you have written an article - with or without AI help - you can ask your assistant for feedback. However, despite all the progress made, I personally don't trust these services to make improvements to the choice of words, wording or sentence structure. It's a different story when I write something in English: I'm always happy to receive feedback.

But I find it useful to submit my finished article for proofreading. I may include my research material again. In the prompt, I ask whether I have covered the topic comprehensively or what is still missing. Of course, the rules for good prompts also apply here: I tell them who the article is aimed at, who it is intended for, etc.

The answer may well provide helpful insights. The AI may find a point that I haven't thought of yet or that I could explain in more detail. I see it as a warning signal if the assistant suggests statements and topics for the article that are actually already contained in the text. This can indicate that they are too brief or not formulated precisely enough.

Last but not least, feedback can lead to new ideas, because not every AI text has to explain every topic from A to Z.

Help with distribution and content seeding

And last but not least, ChatGPT, Claude and others help me to distribute my content. For example, they can quickly create a teaser for the website and, of course, posts for the social web. They can adapt the style of the texts to the use case, because you certainly use a different approach for LinkedIn than for Facebook.

These are only small tasks, but they do take up a lot of time. And not everyone is comfortable writing such summaries and announcements.

My conclusion: As a professional writer, I was as shocked as I was fascinated by ChatGPT's answers. From one generation to the next, this has gone from a nice toy to a serious tool. In recent months, many others have also tried these offerings and have certainly had mixed results. Sometimes the knowledge to formulate prompts correctly is lacking. Sometimes, despite all the progress, AI tools are still not well suited to certain tasks.

For me personally, I don't see ChatGPT or Claude as a threat, but as powerful helpers that are always at the ready. I don't always use them. For example, I wrote this article completely by myself. But I had Claude proofread a first draft and asked him for further content ideas. And I also created the basic outline of the article series in tandem with my AI helpers.

In other words, these tools have become an integral part of my toolbox. And I find them even more exciting and powerful for anyone who doesn't create content full-time. Because in this case, they can significantly simplify and speed up many tasks.