It's a bit ironic that I'm building a
- nondeterministic shell/deterministic core app
with it's exact opposite,
- deterministic shell (harness)/nondeterministic core (model).
Hmm. Seems like we're going backwards.
I think that "coding" as it's morphing into will become an unsustainable long-term career choice if only because of the context switching. Mental-physical toll that it takes for a developer to maintain sustained attention in the midst of rapid context switching just plain sucks. I get headaches now!
Sitting and reading a spec before approving? Not bad. Maybe 2 or 3 in a day so you have time to verify the models work before you send it off to your perfectly-tuned, highly context-scoped, former-{... meta, FB, Netflix} Principal of Principal Engineers Mega-Agent (I've seen .mds \w 1000+ lines). Right?
How are you all handling this? I imagine this would be on someone's radar to solve. Hell, I'm trying to figure anything out to stop the pain! Lol
It's kind of a cheesy example but I think it drives home the point. I had a few .md files that I wanted Pi to format uniformly. I gave specific instructions for one of the lines of the file: "Insert a space between the subheading (###) and the following line that terminates with a `:`... do that for all documents."
And it followed the instructions to the tee. What I didn't realize is that one subheading (in the middle of the document) was NOT proceeded by a line that terminated with a `:`. So those did not change. But I wanted them changed.
I asked it to revert the changes. It did. All messy again.
This time I started a new session. Same prompt but this time I left out the specifics and just said to make them look uniform with a few general comments about spacing. Result: they look great, readable. I made sure to tell the agent to make it easy for humans to read.
All I keep reading is "lightning speed" this and "swarm of whatever-theyre-naming-them-now" will fix all your problems.
Is anyone being intentional anymore? Is taking one's time (with the help of models) to build something well designed suddenly a liability in the workforce?
That would be insane.
I've been looking at different SDD frameworks for weeks. Everything from the hobbyist repo to Github's Speckit. It seems like every SDD implementation fundamentally does the same thing.
Idea => Plan => Implement => review => GitHub workflow
I know that this is simplistic, but it really is all there is to it.
So, I came up with another implementation. But this one has a narrower scope: the .Net space. And I think that's where we are headed. Narrowly-scoped agents and skills leading to workflows made of chained units of work/skills. Smaller models that specialize. I would rather have a small, specialist model that does one or two task well and fast than using behemoths like Claude. It's overkill.
Anyways, just ranting. I'd appreciate your thoughts. And feel free to checkout the repo. I know. Another SDD "framework". But it has worked extremely well for me and I thought I would share it.