BMAD + VSCode + Roo + Gemini

[ai52-coder]

Hi, I've just been getting into this over the last 48 hours - thanks very much to Brian and anyone else involved and what is cleary shaping up to be an excellent framework.

I have had some great success with the configuration I've set up but wanted to check for the general consensus of how everyone else is going as I'm not sure if I've done it correctly.
I've had some very inconsistent results happening in certain places and feel I may be missing some fundamental Instructions or setup direction.

Initially I ended up getting an LLM to browse the repo to try and distill down exactly what I should be doing with VS Code + Roo Setup and the exact method to approach it - below is pretty much what it told me. I just want to check if this is what everyone else is doing?

Firstly was setting up a Gemini Gem with the instructions from a specific file and then talking with it at length, - we spent a whole day practically - at the end it produced three documents which were
PRD.md,
architecture document.md
front end component design.md

Then I asked it when I should take these files over to my IDE (VS Code) and indicated that a good time is after these documents are completed.

I found that there was extreme difficulty trying to get these three documents out of Gemini Pro at the end because the chat context had got so large after hours of back-and-forth on the design that it was really struggling to give me these document contents in markdown format for me to create files (also was very surprised that Gemini could not give me an MD file to download - it could only try and list the .md text in our chat in a code block for pasting, but that didnt always work either!)

In VS code I copied the bmad-agent folder to the root of an empty folder, setup a docs folder at the root also and put the three documents in it.

Based on the earlier LLM advice I then set up different Roo modes, as shown in the screenshot below for each of the modes and copied the contents from the personas folder into the appropriate ones…

I then set the ide-bad-orchestrator.MD as the main system prompt for Roo.

I got one final prompt out of the Gemini LLM to get me started which was

_Hello Dev!

The project documents (PRD, Architecture Document, Frontend Component Design) are now organized in the /docs folder within the mdmagic-project root directory.

I'm ready for you to guide Roo to set up the initial project structure for xproject.

Please provide Roo with the step-by-step instructions to:

Initialize the React Frontend project. Please recommend the most suitable tool (e.g., Vite, Create React App) for a modern React setup, considering our established project structure and component organization. Provide the exact command(s) for initialization.
Initialize the FastAPI Backend project. Provide the exact command(s) for setting up a Python virtual environment and installing FastAPI.
Create the recommended top-level folder structure within both the React frontend and FastAPI backend projects, as detailed in the "Frontend Directory Structure & Naming Conventions" section of the Frontend Component Design, and the "Overall Architecture Philosophy & Patterns" of the Architecture Document (specifically, src/ for frontend and backend/ for backend, and their respective sub-folders).
Specify the default local development port numbers for both the frontend and backend.
Confirm the initial requirements.txt and package.json dependencies that Roo should install to get the basic setup running.
The goal is to have the empty project structure ready, adhering to our architectural decisions and naming conventions, so that we can then start implementing the first user story._

And off it went and overall did a really good job creating lots of stories and felt it was going great.

Although a little way in it it said it was missing key documents such as project-structure.md, operational-guidelines.md and tech-stack.md, which were never generated in my initial chat with the LLM however once in VS code it would not proceed without these documents until I went back to Gemini and asked it why it didn't have the documents and then and apologise and gave me some promptss to give back to Roo to tell it where to look in the other documents it had created out of those three.

So I'm not sure why, when there seems a very specific and seemingly excellent methodology, that it does not produce at the start what it says is mandatory to proceed with and how to direct it - or again it may just be my lack of understanding or the lack of documentation.

Also at times it will say "what do I do next?" and I kept asking it what it thinks it should do next, because it has the methodology and apparently is the expert on the development flow… It seems it doesn't want to create all the user stories first it wants to create some of the stories and then Develop some of the code and then interface and testing… But I'm just working through that now.

The strangest thing was when I ultimately had to start a new chat as although the context compression was on, it was getting unwieldy and expensive… And when I changed the chat it seemed to be completely different. It referred to all the different agents by different names and seemed to switch modes without changing the Roo mode - as the Roo mode then stayed on BMAD ID Orchestrator - but it would just change modes by invoking text… and in this scenario my developer was called James!

In the initial session I did which ran for hours it was actively changing the Roo modes as you can see at the bottom of the screen and the developer was called Dev.

So all of this is entirely confusing and I'm not sure why, although I can only guess that there are some settings/modes for the IDE usage of this and some for the web use of this, and I think this is where there seems to be a lack of clarity in the documentation, that might be obvious to those that have designed it but coming into it doesn't make sense to me.

When I went looking this morning on the repository and in the discussions, I tried to read more information I saw the document instruction.md where it talks about it utilising a build script to package various agent assets etc etc... and now I'm not sure whether I need to be running something to build it rather than just putting folders in certain locations… it seems to be able to do that ....
Again there seems so much that is excellent here but it's totally confusing and I'm unsure whether I've got it configured correctly even though it is doing some excellent work for me.

So I just wanted to get this out there to see if anybody else has any experience or clarity around this. I did see the "a beginners guide for dummies" in the discussions here - and that helps somewhat, but doesn't address a lot of the things and I'm still not confident I've got this configured correctly.

Should it be changing the Roo modes automatically in Roo itself, or does that not matter as long as it says in the text that is changing mode and it stays on the orchestrator?

Am I missing anything else? Should I run a build of something?

Again I want to thank Brian and anyone involved in doing this. This is already generating such superior code and planning beyond what my coding and software development abilities are - and I am very grateful - so I don't want to come across as complaining about this - I would just love to know if I've implemented it correctly and fully.

Thanks for listening.

Roo Modes

Gemini

Changing Modes just by invoking text rather than Roo Modes

[fabb]

As far as i‘ve understood in BMAD v3, the web architect just creates a single architecture.md file which is then expected to be sliced by the ide architect with the slice/shard docs task. Since i started my project with BMAD v2 and only later migrated to v3, i‘m not completely sure this info is correct.

[ai52-coder]

Hi, thanks for that...in all my interaction with the system there has been no mention of slicing or sharding although I did see it mentioned on the video that Brian did when he announced the release of v3.

[bmadcode]

Hi @ai52-coder - thank you this feedback is very helpful and I do not take it at all as complaining as you took so much time to share a lot of details with is very helpful! Apologies on my part, I did skim a bit past that part accidentally in the video, I had a longer edit explaining it and it did accidentally get cut when I ran into audio issues and rerecorded and forgot that whole part :) I hope to complete a very short but concise video this weekend outlining these these and a bit more in a very clear workflow.

The idea as @fabb mentioned is correct - get the 2-3 large docs out and then in the ide use the sharding task. There is a potential bug where it is not properly sharding, that I am working on fixing this weekend, where it might fail to extract a few things when there is embeded markdown - its not always, but there is a bug in certain conditions and the fix should be in place tomorrow,

The longer the chat, the more potential there is for context overlfow, even with Gemini 1M - 2M window is coming, but even then, its always potential to run out. So I recommend - and should have clarified in the video - just because you can do it all in one chat potentially, it doesnt mean its the best idea always, especially if a lot of back and forth ideation in any phase.

But it doesnt hurt to try, just make sure for example when the initial PRD is done, before switching to the architect, get the full copy actually chatted out by the PM agent and save it either to markdown, or as a google doc. Similar with the Architecture.

You can always start a new chat, provide any documents you have so far, and tell bmad in the new chat for example - here is the PRD and Architecture, I now want to talk to the PO to run the final checklist etc...

It looks like you do have all the agents set up in Roo - and honestly once it Roo - you only really need to SM and Dev Agent potentially - the others are nice to have if you want to do some non standard 1 off things other than draft a story or dev a story.

BTW my recommendation (although not required) is use the SM to create 1 story, then the dev to develop it to done, then draft the next story etc...

before doing the SM/Dev cycle - you can try to use the SM /doc-shard command - or if the command does not work you can just execute the task doc-sharding-task.md - in cursor I can drag a task file into the chat and just say something like execute this task for the docs/PRD.md for example and it will create the granular epic files. similar for the architecture.md - I assume its very similar in Roo.

Hope this helps, and watch the repo for updates as I hope by the end of this weekend, many small annoying things that help clarify some of this, streamline a few things, and also ensure that the sharding is solid will help!

Good luck!

[fabb]

I‘m currently working in post-MVP and iteratively creating new epics and implementing them. It‘s not ideal having to go through the orchestrator every time when working like this, unnecessary overhead and extra messages for getting to the correct role. Especially in roo code it seems to get confused when i directly write that i‘d like to plan a new epic with the PM in the first message in the BMAD orchestrator mode i created. I first have to write PM, and wait until the BMAD orchestrator starts up and switches to the PM and only then i can write that i‘d like to plan the next epic. Since i‘m using the copilot licensed models via the vscode lm api in roo code, i have limited requests and not tokens when using sonnet or gemini, so starting a persona directly in a roo mode would be more cost effective. Especially since roo code supports mode switching with boomerang tasks anyways (even though i‘m not yet using this). The dedicated sm.ide.md and dev.ide.md work much better and support directly giving it instructions which story to tackle right in the first message. So i‘m thinking about creating direct lean prompts for the pm/architects/po as well.

[fabb]

Update: i just saw #133, this could help! Looking forward to trying this 😊

[bmadcode]

Yeah! I'll have it done today very close 😀

…

On Sat, May 31, 2025, 1:23 AM fabb ***@***.***> wrote: Update: i just saw #133 <#133>, this could help! Looking forward to trying this 😊 — Reply to this email directly, view it on GitHub <#128 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BOTVGQMBXKGNZHLZHD2NK6T3BFDHHAVCNFSM6AAAAAB6HDOZYGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGMZSG42TCMY> . You are receiving this because you commented.Message ID: ***@***.***>