Has anyone successfully shipped a greenfield production app (100k+ users) using llm assist?
Posted by alee463@reddit | ExperiencedDevs | View on Reddit | 11 comments
Will be starting a gig in two weeks. This will be a greenfield project, starting with another dev. I have about 10 years of full stack experience from a few large companies so I have a good idea on how the sausage is made, but this will be my first greenfield project. It’s going to be something similar to the Whot app. Has anyone had success with building something from 0 to 1 to beyond w leveraging ai tooling from the start? Any tips or gotchas?
90davros@reddit
The biggest challenge IMO is teaching agents to stop and ask for clarification if they encounter a question instead of just picking whatever they fancy. Current models tend to just drop parts of the spec without warning to finish the task rather than raise a query, since they're trained to finish their task regardless of the result. That makes human review of the resulting code mandatory for anything that needs to be reliable or scalable.
AdidasGuy2@reddit
You clearly haven't used Claude code properly
90davros@reddit
Claude is one of the better behaved models in that it'll ask upon hitting significant choices, but it still has a tendency to try to brush over implementation problems. For example it'll often fabricate defaults for missing values. It's dangerous because it won't actually mention this until you read the code.
I know "just prompt better" is a meme, but the prompt will always be more like a guideline than a hard rule. All the AI companies are currently struggling with models quietly forgetting little details in requests.
AdidasGuy2@reddit
You can write that in your md file to check for fabricated defaults used for missing values. Human review will always be needed regardless.
Only-Fisherman5788@reddit
biggest gotcha nobody warns you about on a greenfield with heavy ai assist: you move fast enough that you stop reading the output carefully, and a bug gets wired into the structural layer before anyone notices. by week 4 you're debugging a side effect two layers deep that nobody can trace back to a specific commit because the ai wrote it, the human reviewed it at 9pm, and neither remember why.
concrete practice that saved us: every ai-generated function or endpoint gets one human-written failure-case test during the same session, not later. the goal isn't coverage, it's forcing a human to state what "correct" looks like while the context is fresh. later-you will thank current-you.
xeric@reddit
Seems like it would be reasonable to me, as long as you keep the overall architecture and module boundaries in check. Make heavy use of planning mode and have a human in the loop for architecture review. Have different agents with their own specialized goals around security, scalability, performance, etc.
CandidateNo2580@reddit
Alright you sound like you have some experience with this. Solo dev at small company, just wrapped up a big project about to start another. Aiming to try to get as much AI involvement as possible (more as an exercise/experiment than anything). You have any suggested resources or reading material on getting them to plan and operate within their specific module effectively? That's roughly my plan - architecture is generally designed, plan is intentionally modular and layered.
xeric@reddit
Also highly recommend Claude Code’s visual brainstorming skills - does a good job creating Lofi mockups or architecture diagrams of the solutions it’s considering
xeric@reddit
Also make it clear from the beginning that you’re building to scale, not just making a POC. Keep it from jumping straight into the code, use spec-driven development concepts.
GOT_IT_FOR_THE_LO_LO@reddit
agree with others, you need to make the technical decisions and think of good architecture & requirements yourself.
going through phases of writing a README/AGENTS.md that outlines specifics of best practices and how the project will be structured will go along way to making sure all the code that is produced adheres to that standard.
If you just give AI tools a generic prompt about what you want to build, it’s very easy to code yourself into a corner once you are adding new features on top. If you have already defined patterns of codebase then it goes much smoother.
idontevenknowwhats@reddit
I have, thousands of users though, not 100k