I have worked over and over practicing with Cursor in Composer Agent mode to work out a refined workflow on a somewhat interesting project that is a bit more fun than your standard TODO app (as summarization of hacker news daily sends email summary of top 10 articles and comments, hosted in AWS) - and wrote up a lot of details of how to work with the AI Agent in the workflow I can AiADD with me as the PM/Technical Approver, and the Agent as the developer. I have refined this and it works very well through increasingly complex projects. I also follow a similar pattern in very large repositories for work and have good luck with this workflow.
I keep YOLO on with this workflow.
The High Level is:
- With this method your role is Project Manager / Architectural Review Board / Approver (and with the right personality for the AI, commander of a Enterprise class Starship
) - .cursorrules at the base of the project drive the AI to work off of a PRD file 1 item at a time.
- a .ai folder is the project mgmt history and memory for the project or task at hand.
- with each item, it creates a story file with breakdown in great detail of how it will meet the objective, and then waits for my review of it.
- once I approve it will execute and generally also use TDD
- sometimes it thinks it is done, so i will tell it to run the tests (if it did not) and it will iterate on testing until it is all working
- Once satisfied, I will tell agent to push all changes, it will update the story with progress, write a good commit message, and push all to the remote
- Claude 3.5 still is the only model this all works well with, even in writing the MD files for the stories as the other models screw up and fail to generate the files.
- Between stories, I start a new chat - the nice thing is I can just greet the AI or ask it whats up, and it will know and suggest to generate a new story file or continue on the current one.
- There was almost no code I wrote manually in the sample repo, and maybe 3-4 times I used cmd-k.
Cost Savings:
- Using Agent mode in cursor of course it not the most cost effective, but it will give you a glimpse of where I think the future is of work with AI Agents in a collaborative workflow, and it works very well!
- To save cost, don’t burn tool usage on simple things that you can be scripted into simple commands - and then just user the terminal to run them yourself (its not as fun, but it does save on credit burn)
- Spend the MOST time on your plan / what you want to have built in the end, a good PRD with tech stack, constraints, preferences, clear goal, and details Story breakdown.
- More Granular and Detail Stories in the PRD will produce better story results
- Once Other cheaper / free models bugs are worked out, they will be better to use for generating stories and updating the PRD in an iterative fashion with the AI in cursor - in the meantime, feel the credit burn, or use external AI tool (especially for the PRD)
More details and best practices in the repo - you can get an idea after a few days of running it through this process - it implemented the 9 stories in the .ai folder. Not saying at all its the best structured project or way of doing this specific application (also details on better version and ideas in the readme) but its just generally a great exercise to work our your own flow of what works for you.
Your project idea sounds great, have fun with it and learn what works and what doesnt.