Skeptic impressed by colleague's AI workflow

I've been in the AI-code-assistant skeptic camp for a while, but a recent conversation with a dev I know and respect may be changing my mind. He's solo-programmed a few popular online games and is working on another one. So I asked him for a specific feature example and to see his AGENTS.md file.

Here's what he had to say about his workflow that prompted our discussion:

"I’m (we’re) kicking out feature after feature that is being done right because I have an agentic AI flow that not only writes code virtually exactly the same way I would manually (after many, many prompts to update the AGENTS.md file), but I’m confident in it due to the review flow itself.

I am using OpenAI’s for their codex cli tool which is my main one, but also Claude Code almost as much - the extra veracity you get by having 2 frontier products that can review themselves is probably worth it. And I use Gemini free tier for its extra review some times. Oh and have vs code copilot for its AI autocomplete."

Here's his prompt to create a new feature (truncated for HN char limit, full prompt in comments):

  "ok now i want you to work on the offer, confirm and decline player draws events in game feature. there’s already a userSubmitsDrawOffer file on the server, rename this and fix references. call it “userSubmitsDrawRelatedEvent”. on the client...

  ...lets get it to that point and then when we’re at the point of an accepted draw we’ll do the “end game” style updates after you’re all done with this. go ahead and try to do this now plz."

And here's the AGENTS.md file (active link in comments):

  https://hastebin.com/share/vewunonulo.sql

I told him that he's probably in the top 2-5% of AI-coding experiences I've read on HN. So he asked Claude why they thought he was having a better experience.

His question: "question, it seems i get much more productivity gains using you than other people. is this related to the fact that this is a one-language monorepo or no?"

Claude:

  Good question. A single-language monorepo likely does help, but it's probably a combination of factors. 
  Why monorepo + single language helps: 
    - Consistent patterns throughout 
    - once I learn how you do things in one area, it applies everywhere 
    - No context-switching between different languages, build systems, or paradigms 
    - Shared utilities/types are discoverable and reusable 
    - Import paths and dependencies are predictable 
  Other factors that might matter more: 
    - Codebase quality 
    - if your code has clear patterns and good structure, I can learn from it and stay co nsistent 
    - Task scoping 
    - how you frame requests (clear, well-bounded tasks vs vague asks) 
    - Iterative workflow 
    - some people work with AI in tight feedback loops, others try to get everything in one shot 
    - Domain familiarity 
    - web apps with common patterns (React, Express, etc.) are well-represented in trai ning data 
  What might hurt others' productivity: 
    - Polyglot repos with lots of glue code between languages 
    - Legacy codebases with inconsistent patterns 
    - Heavy reliance on proprietary/internal frameworks I haven't seen 
    - Very specialized domains (embedded, HPC, etc.) Your setup (looks like a TypeScript/JavaScript full-stack app based on the file paths) is pretty close to an ideal case 
    - popular stack, single language, monorepo structure.

I have to say I was also pretty impressed by these answers. These are all things that also make life much easier for human devs and extend the lifespan of the app before things eventually turn into a slog. So it sounds like getting AI to work well might actually enforce tight design up front, which is an aspect I'd never considered.

So anyway, I'm just curious if others out there are having similar experiences and if they've run into any hidden downsides or upsides.