I care about creating stuff. How it gets from the idea in my brain to running on the computer, is immaterial to me.
I really like that I go from idea to reality in half the time.
THAT part doesn't mesh too well with AI, since it's still really bad at autonomous wholistic level planning. I'm still learning how to prompt in a way that results in a structure that is close to what I want/reasonable. I suspect going a more visual block diagram route, to generate some intermediate .md or whatever, might have promise, especially for defining clear bounds/separation of concerns.
Related, AI seems to be the wrong tool for refactoring code (I recently spent $50 trying to move four files). So, if whatever structure isn't reasonable, I'm left with manually moving things around, which is definitely un-fun.
> …I suspect going a more visual block
> diagram route, to generate some
> intermediate .md or whatever, might have
> promise, especially for defining clear
> bounds/separation of concerns…
So can my automaton bud [2]…
_____
MODEL
…
The Verdict: If you provide a clear instruction like "Before you touch the code, read architecture.puml and ensure your changes do not violate the defined inheritance/dependency structure," the agent will be very effective at following it.
If you just "hope" it bears it in mind, it probably won't.
_The agent is a tool, not a mind-reader; it will take the shortest path to a passing test unless you wall that path off with your architectural models_.
…
To make it actually work, you need to turn the UML from a "suggestion" into a "blocker." You should add a section to your AGENTS.md (or CLAUDE.md ) that looks like this:
1. Tool Trigger: By using words like "…"
Why this works:
…
_____
I've been exploring some computer vision recognition stuff. Being able to reason through my ideas with an LLM, and make visualizations like t-SNE to show how far apart a coke can and a bag of cheetos are in feature-space has been mind blowing. ("How much of a difference does tint make for recognition? Implement a slider that can show that can regenerate the 512-D features array and replot the chart")
It's helping me get an intuitive understanding 10x faster than I could reading a textbook.
Everyone is hoping to be part of those 25%.