I figured out a way to do this, via Alpaca.

In Alpaca you can set an LLM with a persistent prompt.

Basically, I just told the thing hey, you’re too sycophantic, often needlessly verbose, and often overly confident… can you generate a prompt for yourself to address those issues?

Roughly 30 minutes of trial and error along those lines later, now its quite matter of fact, and is at least more likely to tell me when it is aware it is making an assumption, and ask me for clarification or if i can give it more context, and it doesn’t do the kind of weird, intro and outro paragraphs where it basically just reassures you that your ideas are wonderful and you are valid and i just think the things you say are so interesting!

Then, you feed it a script, ask it do a sanity check, and it will generally go through and identify strenghts and weakness of the code, at least as it perceives such.

Beyond that, Alpaca recently introduced a … character system, that is ostensibly tailored toward making specific kinds of conversational chat bots… but it also introduced a kind of ‘dictionary’ system, where you can give it a kind of additional permanent reference knowledge, to associate with certain terms.

I have not tried this yet, but, I’d be willing to be that you could say, jam that with a bunch of examples of syntax and methods from a particular language or library… and my guess would be that you could thus tailor a ‘character’ that is more up to date or specific to some domain.

So… you could give it the main prompt of something like “You are a tsuntsun senior programmer who has nothing but contempt for any coding mistakes, and you pride yourself on coming up with entirely novel insults for each inadequacy you notice.”

… And then give it a ‘dictionary’ that pertains to syntax, methods, perhaps even broader concepts…

And that might actually produce your desired vicious asshole senior programmer persona.

Of course, this is not going to work for like, an entire massive codebase, unless you’re the one stockpiling all my the RAM.

But for smaller projects or just single scripts… it might kind of work.