Using BMAD for a large platform project - some learnings

[Milan4747]

First I want to say a big thank to you Brian for the BMAD-METHOD. I find it absolutely fantastic.

My project is a video platform and quite big

Some learnings:

1. Gemini will lose context at some point - and the BMAD agents will jump back in steps often. To avoid that now I always copy paste the next planned step into my prompt or what they have done before. Example "Thats great - yes proceed to perform the "Sharding & Backlog Preparation" task" - like this I reduced the errors.
2. The output of some requirement documents is too long in chat for Gemini. When you notice that some chapters e.g. in the PRD are skipped or just filled in with generic text - tell the agent! Then the answer will be split in as many parts as needed.

In general I learned a lot lot lot - the checklists and inputs are great. But I am already working on this for 3 days haha

I will update as soon as the BMAD agent has created the final documents. Excited.

Also a question, I started the project with BMAD v3 - does it make sense to upgrade to v3.1 during a project?

[getpat]

If you are looking for long form context I have not found anything to match the work in Augment - it can work in a large code base and it can stay on track. BMAD is the killer app for getting Augment to stay on track.

[Milan4747]

Interesting. Can you use the Claude Max plan in Augment?

[getpat]

Augment has its own plan, it is 50 a month, claude code uses the claude max, I use both tools sort of against each other, I use Augment to engineer and make major additions, I use claude code to run tests and fix issues. The two together cost less than five days on the Claude API. Both are now using Claude 4, most of the time they are using OPUS but that is not a promise from either. I find off hours are more effective for signficant work, plan your work - work your plan. I am very impressed with BMAD and its impact on long running processes. I would always run the plan in Augment then use Claude Code to review the work and test it agentically creating a report on its findings. I liek to document code reviews and create a document on every file, telling it to review every line of code as it will only look at a few hundred lines unless you tell it every line.

[bmadcode]

I really want to try this set up soon, in a few weeks when I can give this stack a fair try / will have time. Everything I am hearing about Claud Code has me very intrigued. Also I miss using stock VSCode and the keybinding muscle memory built up over 10 years of vscode. I tried Augment early on and liked it - considering that or going back to VSCode copilot as it seems they are making some strides finally, and I gotta bet the MS is not going to let copilot remain behind for much longer.

[fabb]

I like copilot, but it cuts down a lot of the actual context that is sent to the llm to save costs for them since users only pay a monthly flatrate. Also the limited premium requests (for models other than gpt-4.1) are billed by request and not by tokens, so smaller requests make sense for them cost-wise.

cline and roo code do not have these financial incentives and send more context at once, which works a lot better with e.g. sonnet.

[getpat]

I will be checking out the code rabbit, it sounds interesting.

[Milan4747]

I have one more question, maybe someone can help.

Now I am in the last part, creating the Sharded User Story. For me this would be 46 files.

Is there a fast and easy way to create 46 Sharded User Story Files?

[bmadcode]

I recommend only sharing yhe prd architecture and design architecture if also print end. Then when in the ide after shared the snowflake just creates a single story that is next to develope in the docs/stories folder. I have much better results. But... I can add a sharder for the individual stories if people are mass producing the very detailed stories also... Will make some improvements around getting this done by the web agents also for better rendering of all stories for dev drafted for an epic

…

On Sat, May 31, 2025, 8:15 AM Milan4747 ***@***.***> wrote: I have one more question, maybe someone can help. Now I am in the last part, creating the Sharded User Story. For me this would be 46 files. Is there a fast and easy way to create 46 Sharded User Story Files? — Reply to this email directly, view it on GitHub <#134 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BOTVGQMQFPPMIP6EMHIX53T3BGTOJAVCNFSM6AAAAAB6JYVPSKVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGMZSHEYTANA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.*** com>

[getpat]

I would llike to see a code review/code improvement agent that goes back through and examines the code from multiple angles including logic, overarching project goals and of course best practices

[bmadcode]

This is great idea pat - I have been thinking about adding a refactor task and a review task, so an agent could be tied to do these. I have also started using code rabbit which has been a big help although its yet another paid subscription, so getting it right with an inbuilt agent will be ideal. Code rabbit though is wild, with the plugin, dev agent builds the code, rabbit gives feedback and suggestions, copy pasta that back to the dev agent and its usually good at making the suggested changes.

I do envision it to be very configurable to platform, preferred style etc, much of it stemming from what was planned with the architect in the operating instructions.