A pragmatic writing workflow using Gen-AI

Combining Whisper, GPT, and your own brain to supercharge your writing without selling your soul

This post describes a loose workflow I have recently developed for quickly drafting new articles while on the go.

A while back, I published a comprehensive guide on writing technical articles. I outlined a six-step process for transforming a messy idea into a polished article in that guide. The steps can be followed loosely, guiding you from disorganized thoughts to a well-structured outline and, ultimately, a refined piece.

How to Write Technical Articles

If you don't want to read the full article, here's the gist of it.

You start by brain-dumping all your ideas onto paper (or electrons). You distill the most important concepts and extract the main talking points. Next, you rewrite these ideas, organizing them into a logical outline. Finally, you polish and refine the content. This way, your thoughts move from a jumble in your mind to a coherent structure on paper.

I intentionally did not specify a particular toolset in that article because it works even with pen and paper. However, at the end of that guide, I mentioned that you can use generative AI tools to enhance the process without losing your voice.

In this article, I will present a specific implementation of this process that diverges slightly from the original and relies heavily on generative AI, but in a good way. This approach allows you to increase productivity without losing your voice and style and makes writing much more enjoyable, at least for me.

I will also detail the tools and the exact workflow and prompts I follow. Hopefully, this will help you implement this process rather than just reading another ultimate writing productivity guide.

Let's go.

---

---

The workflow

At a high level, the process can be summarized as follows.

First, I record myself speaking about the main ideas behind the article, which can be done in multiple sessions. This talk is unscripted and doesn't require any initial structuring. It can be a spontaneous rant or a slightly more organized discussion. The main goal is to brain-dump my thoughts without overthinking them too much.

After recording, I use Whisper to transcribe the audio, which removes filler words and other extraneous sounds —like hums and ahhs. So, although the transcription is already cleaner than the original audio, it still includes repetitions and lots of unfinished, messy ideas —unless you're really good at improv talking.

Next, I skim the transcription and divide it into coherent sections based on its internal organization. If the transcription is from multiple recordings, each one often corresponds to a different main idea since I stop and restart recording whenever I change topics. However, if it is just one long recording, I may need to manually separate it into meaningful sections of around 200-300 words.

Now comes the magical part. I will use a large language model like GPT 3.5 or 4 to rewrite and improve the text. I designed a simple prompt —which I’ll show right away— that asks GPT to retain the content while making the text more understandable and improving the style. This rewrite becomes the first draft.

Next, I paste the draft into a text editor and make two editing passes.

The first pass involves rewriting anything that doesn't sound like my voice, as the AI will take poetic liberties that alter the original meaning. I also fix any errors or inaccuracies the AI introduced in the rewriting process. Finally, I make adjustments to remove fancy words chosen by the AI —like the infamous eliminate, utilize, or delve— that don't really match my style and add anything that I think is missing, such as headings and connector sentences. This is the first major rewrite.

In the second and final pass, I polish the text more finely, carefully rounding the main ideas and doing laser-focused edits to get to the most refined version of the text I can. But again, without excessively overthinking it.

After these two passes, I consider the text ready to publish.

---

Are you a technical writer looking for a community of like-minded creators to share experiences and grow together? Join the Tech Writers Stack today! It’s free.

---

Implementation details

Now, let's discuss the nitty-gritty implementation details.

For the transcription part, I already mentioned I use Whisper. There are various apps out there you can use. Still, I prefer this self-made Telegram bot that has the convenient feature of concatenating multiple audio messages into a single text note. Once the process is complete, you will receive the final transcript. You’re welcome to try the bot for free.¹

Moving on to the editing part, you can simply use ChatGPT with the following prompt (in my experience, both 3.5 and 4 give similar results for this case):

You are a proofreader. Your task is to rewrite the input text by keeping the same meaning but improving the grammar, style, and punctuation. Make sure all sentences are in active voice, crisp, and to the point. Start each paragraph with the main idea and continue with the supporting claims. Do not add any content that is not in the original text.

{{ Input text }}

However, to simplify this process, I prefer using Bearly.ai, which offers a streamlined UI that cuts all the chatty nonsense. I made a template with the previous prompt, allowing me to paste the transcribed, messy text and get a revised version back with a single click.²

Occasionally, I will run the generation process twice. Still, I often prefer to just manually adjust the result instead of regenerating until I hit the jackpot with the exact wording I was looking for.

For manual editing, I am not particularly fond of the Substack editor. It lacks many features and can be cumbersome, especially if you're a fast typer used to goodies like multiple cursors and lots of hotkeys.

Instead, I prefer using a full-featured desktop editor like Obsidian. I copy and paste the initial draft into it for the bulk of manual editing. Obsidian offers only basic spell correction on the linguistic side, which means it's just me and my writing. No more AI.

Afterward, I transfer the edited text to the Substack app for the final pass and adjustments, including adding subscribe buttons, illustrations, footnotes, links, and the typical calls to action.

In the Substack editor, I also have the Grammarly plugin turned on, which provides helpful suggestions for simplifying and polishing the text. Generally, I accept about 50% of Grammarly's recommendations, particularly eliminating passive voice and reordering sentences. However, I ignore the remaining 50% of stylistic preferences I don't necessarily embrace.

Final remarks

Once the editing process is complete, many say it is advisable to let the draft rest for a few days before revisiting it and seeking feedback from others. But, in any case, publish your work while it’s warm. Don't let it wither in your draft list, or it will fade into oblivion, never to be finished. It becomes a zombie draft, not alive but unwilling to die completely. Trust me, I've been there.

The beauty of this workflow is that it relies heavily on generative AI, but only for the most boring, uncreative part: taking your unpolished but genius ideas, transcribing them, and giving them some initial structure. The kind of work you would trust an intern to do. And it does so marvelously.

But in the end, it's you at the steering wheel. You provide all the ideas and initial content and decide which exact word survives the final cut. This is the right way to use AI as an empowering tool, not a sad substitute for lazy writers. This workflow won't remove all work, but it does remove the bulk of the boredom and leaves you all the entertainment and excitement to enjoy. It's your writing.

However, this process is not suitable for everyone or every kind of writing. It is especially useful for technical writing, in the broader sense of writing meant to inform on technical topics. For the more creative writing, you almost certainly aim for more precise wording. Likewise, for the more emotional kind of essays, I tend to write them from scratch.

That's it. Feel free to copy, modify, and share this workflow as you see fit. I hope this method makes your writing more productive, but also more enjoyable, which matters even more. For me, it certainly has.

---

At the Tech Writers Stack, we’re always looking for writers like you to share your experiences. If you want to write for us or cross-post something you’ve already written, leave us a comment.

---

Appendix

As an addendum for completists, here is an example of this very article going through the different phases. This section is just meant to show you how feasible this workflow really is. Skip it if you're already sold.

Here is the audio recording of the first section, which ended up being the introduction. Don’t listen to it fully; just notice how unscripted and spontaneous I talk. Also, don’t worry about your accent; mine is pretty thick, as you can hear, and Whisper got it perfectly.

From that audio, my Whisper-powered bot gave me the following transcription (I’ll show only the first few sentences). You can see how Whisper does a pretty good job removing filler words and weird sounds, but still, the text is really really unpolished.

This post is to explain a workflow that I've kind of been converging lately to quickly draft new articles while I'm on the go.

A few weeks ago, a few articles back I published a longish guide, technical guide, on how to write technical articles.

In that guide I highlighted a six-step process that are kind of very structured but you can take them loosely and that those processes are designed to go from a very messy idea inside your brain into a clear article.

So the way I suggest you to review that article if you want but here's the very quick summary.

The way this process works is you have some messy idea in your brain and before simply sitting down and typing that you kind of have to organize your thoughts.

So the process goes from first brain dumping everything that you have and then trying from there to distill the most important ideas and extract the main talking points and then rewrite everything that you all the ideas that you had to organize it into an outline that makes sense and finally do the editing and polishing.

…

From this unpolished text, using my proofreader prompt, I got the following first draft:

This post aims to describe the workflow I have recently developed for quickly drafting new articles while on the go.

Previously, I published a comprehensive technical guide on writing technical articles. Within that guide, I outlined a six-step process that provides a structured path for transforming a messy idea into a clear article. While the steps can be followed loosely, they effectively guide you from disorganized thoughts to a well-structured outline and, ultimately, a refined piece.

To begin, you start by brain dumping all your ideas onto paper. From there, you distill the most important concepts and extract the main talking points. Next, you rewrite these ideas, organizing them into a logical outline. Finally, you polish and refine the content. In this way, your thoughts move from a jumble in your mind to a coherent structure on paper.

You can see it’s pretty close to good enough. I then manually edited that to bring to the actual introduction of this article. All of this took me no more than three hours, 80% of which I did on my phone while waiting in a line at the supermarket. I did the final 20%, including the Substack-specific bells and whistles, sitting back home at my computer.

1. For obvious reasons, I put a limitation on the number of uses of the bot. However, if you are from the Tech Writers Stack community, you probably already know me from Substack. So, if you really enjoy using this tool, simply contact me for additional free credits.

2. Bearly.ai has a reasonable pricing model and a pretty good free tier, enough for those who write one or two articles every week. However, if you prefer, you can use ChatGPT with that exact prompt. Bearly is just a convenient wrapper.