Designing and Discovery in LLM Projects

Hi There! I'm Flavia from Brazil and new in this LLM World! 😀

My original skill is being a Business Analyst, but now I'm trying to move for AI. I mean, being a kind of solution designer, but with some coding also, because I love doing it! Well... These explantions just to contextualize who am I! To understand what is the point I'm trying to bring here! So, let's go to it!

As a BA/Requirements Analyst, my goal is to build, based on cliens needs, the modeling of the product/system. I use some tools for it, as you should know: BPMN, UML Diagrams and Use Case, free text Specifications, etc... I also use agile's tool such as Use Stories/Epics and Discovery dinamics, as Sprint Design. This is great for the building! But, this is to build strutuctred information systems, in major to automize corporative tasks and process. As we know until now. Until the advance of IA Agents, APIs, etc...

So, what is the point? Is that maybe, this "old" way to model an information system or a digital product may not be efficient for AI automations. In my research, using AI agents, I noticed, for example, that US are very low to say or document an AI. I noticed also, they work better with more details, (hello Prompt Engeneerring), with a logical and good structure text, free text I mean. But, I was interacting straight to an agent (I tried the most popular, ChatGPT, Claude, Gemini and Deep Seek), it's relevant to notice.

But the experiment have lighted a though: What would be a better way to model an AI project? Should we use the same tools that we use for the other SW projects? Should we create somenthing new?Or they already exist? Should we built some hybrid? What would be different?

Maybe some known tools are still applicable and good for it, as a PVB or Jobs To Be Done Canvas, used for the very first impressions. For the latter steps, I guess a User Journey Map can be shown as good. In the other hands, I don't think a User Story can be an ideal tool for the requirements, for the job organization it can be, but, the backlog can not provide the information whit the quality an AI project needs.

I guess the way that can be shown as a fit are what we call "Traditional", such as UML and written specifications. But I don't kown, I must study the case and, the main, listen to developers, listen to who consumes these artifacts! Listen to who does!

That's the question: What should be the best way to document/model the AI requirements? What do you think is useful and brings the quality we all need for the project?

And this is also an invitation: Let's think together? Let's build a study group for it?

Come together!!!! :-)