Same here. I review plans w rendered diagrams. Data structures, flowcharts, uml sequences, and component architecture docs.
I have found that making the AI look the plantuml and the rendering and reviewing before presenting it to me results in much more legible designs. To me the peak of this would be a shared interactive whiteboard w the ai during design.
This is very needed in the AI coding age! I find diagramming is one of the best ways for high bandwidth alignment between my mental model and the agent’s. I have been working on this problem a bit as I see good visuals as a key bottleneck in faster development while maintaining a real architecture.
Will def be adding this to my list of diagramming tools the agent can use!
It's funny, you articulated it very similar to how I feel! Seems like we need ways to maintain mental models at a high enough level of abstraction when working with agents.
There's a skill in the repo that you can install using `npx skills` to get started with. I'm sure the integration of Reladraw with agents can be made way better, though. It's just a bit... basic/rudimentary right now!
Won’t it be a good idea to stay at the visual level?
Instead of generating visuals from text and then back to text, just communicate visually with your agents.
Something like UML provides a certain granularity to be able to code something usefully in pure visual language. Of course there are other visual programming languages that provide more depth and would probably be better suited.
Point remains that a solution based on constantly mapping better two different representations of the same thing won’t be a long term efficient solution.
For my part, I've found Literate Programming to be the best lever on this for my own work, and there has been some discussion of other folks being successful applying it to LLM work:
Maybe, but if it's GRAS and supplements/non-drug interventions, can't you do it anyway? This law is addressed at products that are regulated as drugs, biologics and medical devices
Actually I think they do. Visualizations are a form of compression and focus.
All of my ai-driven plans use a tool I built to rasterize all text based visualizations. And I make the model present the rendered version to a clean room agent for brief-back. They go in circles until the diagram is good. Then, I instruct planning tools to always look at both the text and rendered version of all diagrams. I have found it to be surprisingly powerful.
It’s gotten so good that I now have infrastructure to render all mermaid/plantuml in my project to png and have AI’s always load both text and image versions. And they are instructed to review the rendering as part of the diagramming cycle (for layout, salience, usefulness, etc). They can now produce useful diagrams that help reach shared architecture understanding.
Exploring highly interactive instrument (piano) practice to see if AI can help students practice better. Full duplex voice agent alongside your practice session. Also exploring live AI jamming partner to practice playing with others.
I have found that making the AI look the plantuml and the rendering and reviewing before presenting it to me results in much more legible designs. To me the peak of this would be a shared interactive whiteboard w the ai during design.