The Real AI Problem Isn't Questions—It's What They Actually Know

Maneshwar is building git-lrc, a tool that does code reviews on every commit. Nothing groundbreaking so far, right? The genius part is how it actually works. Instead of letting the AI read your code like a tourist reading a map, you feed it context first: the project structure, your conventions, key files, change history. Basically, you're furnishing the brain—you're giving the model an actual foundation to stand on.

It's like the difference between asking a friend who doesn't know your project to review it (result: generic, useless comments) and explaining the full picture first (result: feedback that actually makes sense). The AI isn't guessing, it's working with solid data.

Technically, what's happening here is way smarter prompt engineering. Instead of whispering generic instructions to the model's ear, you're handing it the instruction manual for your entire system. It's the difference between shouting coordinates and handing over a detailed map with all the road signs.

And here's the kicker: git-lrc runs on every single commit. That means every time you push, you get an automated review that knows your project—not a bot reading the code like it's stumbling into unfamiliar territory for the first time. This is especially useful for small teams or solo devs who don't have the luxury of always having a human reviewer standing by.

Maneshwar's approach proves something people still haven't figured out: AI models don't need smarter questions, they need richer context. The model isn't the problem; the silence around it is. When you give it the full picture, it magically starts saying intelligent things.

This is why generic prompts ('review this code') fail regularly, while loaded prompts (that include config files, project standards, previous code examples) often hit the mark. It's not magic, it's just doing the legwork before asking the question.