1. Introduction

I wrote the previous article using AI’s help. In this article, which I write the conventional way, I describe

what I did
why I did, and
what I think about the result.

2. How I wrote the LLM article

As a first step, I created some handwritten notes and bullet points about what I wanted to discuss in the article. As a very first attempt, I tried to do the article without the structure at hand, but the result was terrible. I was deviating from the topic, so I created a long, blossoming text that would be acceptable in a live conference speech but not in a written article.

After creating the notes, I dictated the article, walking up and down in my room and feeding the text into MacWhisperer. I was essentially dictating the text of the article.

I have used English for 45 years, but it is not my first language. The transcription is fine; my accent is good enough for dictation. I can speak English in front of an audience; dictating an article’s text is a different trade. I was not content with the result.

I repeated the process in my mother tongue, Hungarian. The result was better, but still not perfect, but I had no more options for this phase.

I copied the transcribed text into the article and asked calude.ai to clean it up. Since it was Hungarian, I used Hungarian for the prompting, but the result is language-independent.

Claude.ai first made a short version of the article, like a one-pager, though I asked it not to shorten it. I amended the prompt and could get as far as 70% of the original text.

Making the text shorter is acceptable to some level you can expect. Spoken text is more verbose than one in an article. However, 30% was too much, and I looked at some specific content that claude.ai skipped. I asked why it skipped that part, and strangely, it fixed the text instead of answering the question. I had to command it to answer the question and not treat it as scolding.

The explanation was reasonable, but again, it apologized and explained that it was its fault for assuming those parts were unnecessary. And it provided a corrected full-length text, which was finally acceptable.

I could use it, but as a programmer, I wanted a workflow I could repeat without haggling with an AI. So, I asked calue.ai to provide me with the prompt I should have asked in the first place. It did, I tried, and it worked.

Now, I had a Hungarian text that was almost perfect. I was proofreading it and fixing some of the grammar, which was correct but smelled like a machine. Then, I asked calude.ai to translate it into English.

I fed the translated text to Grammarly, and I fixed it.

Then I asked calude.ai to translate the English into German, Spanish, and Russian. I do not know these languages well enough to check, though. I understand them to various levels, but the level is far from my English enhanced with Grammarly or my Hungarian.

The Russian translation was a bit tricky. Claude.ai has already translated the text to German from English, but it has started to complain that the text is too long for one session. How about summarizing it in Russian or translating it into smaller etaps?

I told it that it had already translated the English version without question to German. And then claude.ai said that okay, then, and it translated the text.

My Russian-speaking colleague told me it reeks a bit of AI, but that should be okay.

It is okay because… and here comes the next part:

3. What I wanted to achieve

Writing an article is delivering information. It takes a lot of time and effort to write one. With the recent development of AI technologies, I started to use AI to proofread my English. I thought I would also use it to transcribe it from spoken text. After all, I can speak faster than I can write. If there is something that a machine can do instead of a human, then let the machine do it.

To answer the question of how to do it is engineering. And I did what I did; here you are.

4. Conclusion, Feelings

Is the result good? Please read the article; I linked it at the top, and you can tell. Dictating and speaking are different from writing. It is faster, but the style will be different. Even in the English version, you may feel it was produced using a different method.

At least it is a good experience playing with the bleeding edge technology.

Comments

Please leave your comments using Disqus, or just press one of the happy faces. If for any reason you do not want to leave a comment here, you can still create a Github ticket.