Which AI tool do you use for documentation?

Well, your point about reflecting got to me. Indeed, my original response was significantly harsher than it should've been, although I didn't intend any of it towards you, but rather the, let's just say 'lack of pride in their work' of the people making these AI courses. My intention was getting to the point quickly and effectively for someone who'd just skim through these texts: do not use these guides, they are straight out misleading. At least by the end, we're both better off I guess: hopefully you know not to trust these oversimplified comparison tables the next time, and I learned to think a few more times about what I write, to make sure I don't come off as inconsiderate and offend someone I didn't mean to.

It's free (open source) so the diagram source should always be usable if the site disappears. But I am interested in other options always, I would prefer more design control personally (I am quite picky about positioning of each element, etc!) hence I mostly (ab)use PowerPoint, but that's not efficient use of time.

A shortcut I've used a few times for reverse documentation: run all the test cases, capture the log output, and then run that through a parser to generate such ping-pong/message sequence diagrams, and use that as documentation : ) That's only possible if the logging has enough information in it of course.

For one project that was so useful, a customer saw me using the parser tool to troubleshoot their setup, and wanted it; so from then on we packaged it with the system, and the customer could turn on logging, and run the tool, and easily see what was going on (the diagram showed external inputs arriving into the system, then the few key functions that were executed internally to ripple through to state changes or a response back out, so they could easily see if things were abnormal or not). It was quite handy to see if things were overloaded (e.g. loads of incoming events) or weird state machine changes, etc. This was for the more tech-oriented users (the more normal way was to use third party tools to monitor the entire solution at once, not just one device).

Yep, I tried it although I believe the free option is a time limited trial. Pain.

I found https://sequencediagram.org/ to be very handy and works well with AI tools like Gemini/AntiGravity. I just tell it to review my code base and create a sequencediagram script of the sequence process and it does it. Then you just copy the script into the website and out pops a diagram which you can view/download. Very useful.

Anyone tried Mermaid? It's an online diagram-drawing tool that accepts text-based input. I mention it because AI can be prompted to generate Mermaid output. Might not be correct always, but it can be useful to get an idea about a body of code that one might be new to.

Example screenshot, where the red box is my microcontroller. 

To do this was quite simple:

1. Install AntiGravity CLI (or use whatever AI you wish; AntiGravity is the consumer way of accessing the functionality in Google Gemini).

2. I used Windows PowerShell. I went to the folder where my body of code was, and ran AntiGravity (by typing agy) and then just explained in text what I wanted: 

If you were say using ChatGPT you could have uploaded the files, or perhaps simply provided the URL if the code was on GitHub, and so on.

3. The output was a file called ARCHITECTURE_DIAGRAMS.md 

4. I copy-pasted the diagram from the .md file into the Mermaid website.

I can use the diagram to refresh myself about how the code is architected. 

Personally I like drawing my own diagrams because I prefer reducing them for as much simplicity as I can manage, but it is still useful for a quick reference. 

I think it depends on what you are building and documenting. One does not normally implement a compiler and then "discover" what language it compiles. One invents the language first and then writes the compiler. 

System engineering starts with requirements, not implementations.

Design generally should put the human user first, and designs around the user goals and interactions. Then people can worry about how to build that.

Lower level things like API documentation should at least start with a design so multiple teams can implement concurrently. It can always be adjusted as issues arise (hey we need to include a FOO in the argument list for these functions because ...). Better editors and implementation languages allow making such changes easy (who-calls).

Thanks.  I like how the last row shows what each tool is "best for."  It's a good way to end the table.

My ire over the dialog has diminished. I'm chuckling to myself as I post this response. I did some reflection before responding. As I get older I find that seems to be more important than just replying.

I went to a very good friend of mine with a problem I was struggling with. He listened attentively asking probing questions to fully understand the issue. After a few moments he said, "I have the answer you are looking for." He then waited before saying, "I just can't tell you."

Our friendship has grown over the years. We still laugh over that conversation. It took me ten years to respond. "I would never have listened to you, if you did give me an answer." His response, "I knew that so I wasn't going to give you an answer anyway."

"I appreciate that you didn't intend it as a personal attack. My concern isn't your technical disagreement—it's the language you chose.

Calling the table 'horrible,' comparing it to astrology, and later referring to AI guides as 'completely useless trash' is dismissive rather than constructive.

The sources you later provided were far more persuasive than the original wording. That's the distinction I was trying to make."

I confess I didn't write the above prose. After reflecting on your response, I thought maybe it's just me. The reply is an AI's response to the question, "What is an object of an apology?"I also provided the dialog we've had, along with some of my thoughts. I posted the AI response because it reflected my values.

I started my technology career in 1981. Entering a site required a sign-in and sign-out. Every action taken on the site or work on the equipment was written in a site logbook. If the was an incident, the log book was the first thing ceased by authorities.

4X4 drawings show system interconnect were the norm. On training course you were given a blank schematic. By the end of the course you would have used up half your colour pencils making notation. Green for receive and red for transmit.

The lessons learned in those 18 years never left me. Every job I took after that, I carried a notebook. I typically start a project with what I call the "Napkin Drawing." The doodles sometimes on a napkin that flush out an idea. I find it really hard to discard the napkin even after the CAD's are done. Maybe I will miss something I had on the napkin. KiCAD'ing a schematic is time well invested. 

Up to this point only time has been invested. I have a scope of work, a component list and a way to estimate time. If at this point it never gets built, time can't be returned. I also discovered once mothballed it take time to recover the knowledge that got me to this point. The older I get, the harder it is to recover.

I find sharing project experiences on this site a great way to document.

Me Too Movement.

I am also a minority in 2030.  I even feel vomiting when it comes to designing hardware with AI or designing hardware software with AI.

Real world hardware design, AI, are actually different domains.

As far as I know, the most complicated robotic actions shown on video are actually fine-tuned for many months before they are shown on the TV stage.  From this viewpoint, they are just a remote controlled toys with more degrees of freedom.

I had Japanese made remote controlled F1 when I was a boy.  At that time, we called them remote controlled toys, not AI. 

I have taken course on udemy.com for LLMs and Agentic AI. I've used every Model and agentic tool available. I use them for Brainstorming Embedded projects using various languages and a wide verify of Evaluation kits. Leaning on AI for design, brainstorming, and vibe coding is a highly effective workflow for me. It is especially useful when prototyping or mapping out the logic for embedded hardware platforms —you can nail down the architecture and flow conceptually before getting bogged down in the syntax.

Currently, I use Gemini exclusively for Design, Brainstorming and vibe coding.

I have not generated documentation from AI yet . The thread here is interesting and I'm open to checking out the possibilities of using the tools mentioned in the thread.