If you plan to build a relativly complex app , which approach have you found most helpful?
a) just give it the whole idea/plan and let it run with it and keep iterating
b) give the agent the whole vision/structure of the planned future state but ask it to execute the structure and just a couple of the first features
c) choose some initial steps out of your plan and only give those steps to the agent, and progressively feed more steps.
My question comes from wondering whether there is a benefit to showing the model the whole picture so it understands the whole even if building only the initial part, or whether showing the whole project confuses the model more than help.
I believe the best way to work used to be brainstorming your app with something like Claude Opus.
Then getting Opus to write a series of development stage prompts that would layout the best tech stack and architecture for what you’re trying to build.
From there you would give the Replit Agent stage 1 of the prompt.
Bug test Stage 1. Get everything working on Stage 1 and then move onto stage 2.
This is what I did to build the apps I have currently built.
I’m not sure if this is the best way to work now that the agent can run continuously for 200 minutes.
I am looking forward to getting the budget to start building another project with Replit from the ground up using Agent 3. Giving it my full list of prompts in the beginning, leaving it to run autonomously and seeing what it cooks up.
Agent 3 feels worse when using it to work on existing apps, but perhaps it works great on new apps that it builds itself from scratch?
Allsop, I follow a similar planning model as what you do, plan extensively outside of Replit, howeverI tried a new app build with agent 3 yesterday, it had all the relevant documents added, I specifies every part of the build, tech stack the lot, high quality first prompt. it took the md docs, changed them, then tried to build the app out, with instructions not to. It looped out at 48min in. Told me it’s sorry for the mistakes. So I would wait or a Replit response before you try. As for working on an existing app, id clone and test, IMO.
@alespi Progressive, step by step, turn off autonomy, 1 thing at a time, as it normally does one of 2 things correct, and if you need to roll back, you kill the good fix with the bad fix, in my experience, so I stick to 1 feature at a time, but also backed up by my outsdie replit AI chats. Dont let agent add its suggestions, stick to your pre-planned plan.
So I would wait or a Replit response before you try. As for working on an existing app, id clone and test, IMO.