Beta 3 Feature List and Community BIG THANK YOU!

[bmadcode]

Really want to thank everyone that has commented, contributed or raised issues - so much of this has fed into V3 Beta - mostly thanks to all of you for testing and trying to use these agents and workflows, and reporting also issues with different environments and platforms.

Some key items (and I am probably forgetting a lot) that will be in the final beta release (and are mostly in place)

Reminder first - It's really hard to beat the performance of the Web Agents in Gemini 2.5 pro - As much as I love chatGPT - these things really need the power of the massive 1M context window - and a larger 2M is coming soon. For document wrangling and the type of deep thinking and planning we are doing to produce these amazing detailed artifacts that will improve our agent performance in development - large context is king still!

• Streamlined IDE Agents all below 6k characters to work with windsurf
• Web Agents are now very interactive and consistent
• Architecture Preferences
• Dev Keeps Notes on Test Spiral Failures so it doesnt keep doing the same dumb thing again and again
• Improved Templates and Checklists
• A new RTE Agent (This comes from SaFE which is an abomination of agile in the enterprise - but I like the name RTE - release train engineer) - the RTE Agent is your new best friend when the dev mid project finds we need a massive change. The RTE will be who you call to right the course for the remaining artifacts and figure out if we pivot, correct course, reverse, add new epics and stories and fix the remaining ones, etc... This has been one of the biggest questions asked, about mid project changes, and the RTE is here to try and help!
• One off / not used very often type things converted to callable tasks that any agent can run, as to not bloat the IDE agents and allow for working within the Cursor total agent limits
• a new BMad Agent for web installation - an expert you can talk to that is all things BMad Method - it can propose based on your project complexity and how much you already know suggestions on how to best use the full or condensed workflow
• PM and Architect both support a YOLO mode
• PM now has 2 modes for how detailed it will make the epics and stories. If planning to use the full agile workflow (hand off to an architect for example) it will produce stories a bit more in the pure agile style without all the nitty gritty details! This honestly IMO is the best, it allows the architect to do more, and makes changes easier if the architect discovers something - and allows the SM later to really do its job pulling architecture doc info together into the stories - its hard to describe unless you have seen it done both ways, but IMO this is much superior - but the other way still exists - the PM will ask your preference, and you can even ask it for an example!
• the SM and Dev are much more aligned in the IDE - the Dev being trained to look up certain information on its own (coding standards, project structure) - so the SM will not repeat that info when populating the stories. The SM also has been corrected to not add extra stuff it randomly things might be a standard the dev needs and stick to files its pulling from (for things it will still add)
• The story template has been restructured and the SM trained to focus on putting a lot more details into the tasks and subtasks where appropriate and not repeating the information again throughout the story - there was a lotof redundant info in V2 going into the stories, and it got also unreliable in some cases where the SM embellished as I just mentioned, or would choose to leave things out which was just as bad.
• So many template and checklist enhancements.

What remains before removing the Beta tag?

• the web agents are a bit bloated - working on some optimizations
• chatGPT has an 8k character limit - we are far exceeding that with some agents, there is a work around which is to attached the agent file - but I have a better way planned I might get into V3!
• I want to produce at least 2 substantial passes through the whole pipeline of agents to produce sample artifacts (and fully test the system) before removing the beta tag
• a bit more to do for the IDE version of things - more tasks for example instead of making too many custom agent modes - while also providing more agents that do fit under 6k for users that want the cadre of agents experience in non cursor IDE's.

---

BTW - Ideas are already stewing for V4 :)

[Geserit2]

I have also tested a lot and ChatGPT currently does not come close to Gemini 2.5 pro, I use deep research with Gemini 2.5 a lot and try to verify every assumption through deep research and it has often turned out that the assumptions were wrong. 2 things that I personally find important.

1. personas - who are the users, what problem should the application solve for the users and how should this application be structured so that the target group has the best user experience. I can refer to this document again and again in every phase when decisions have to be made and this has paid off several times.

2. most products need a landing page. Even if classic SEO is on the wane, it will still be around for some time and will be important. Therefore, here is an SEO checklist of what the landing page must be able to do.

SEO Checklist

1. Technical SEO:

• Core Web Vitals (LCP, FID, CLS) in the green zone (according to Google PageSpeed Insights / GSC)?
• Website loads quickly on desktop and mobile devices?
• HTTPS encryption correctly implemented and active?
• Website is mobile-friendly (passed Mobile-Friendly Test, no usability errors in GSC)?
• XML sitemap exists, is error-free and submitted to Google Search Console?
• robots.txt file correctly configured and doesn't block important resources?
• No significant crawl errors (404, server errors, etc.) reported in GSC?
• Indexing status checked in GSC (important pages indexed, no unwanted indexing)?
• Clean and logical URL structure?
• Issues with duplicate content identified and resolved (e.g., through canonical tags, 301 redirects)?
• Structured data (schema markup) technically correctly implemented and validated?
• Site architecture allows important content to be reached within a maximum of 3 clicks?

2. On-Page: Content & Keywords:

• Clear main keyword (focus keyword) defined for each important page?
• The search intent behind the main keyword is clearly and comprehensively addressed?
• Content meets the E-E-A-T criteria (Experience, Expertise, Authority, Trustworthiness)?
• Content is unique, high-quality, well-researched, and offers real added value?
• Main and relevant secondary keywords are naturally and contextually integrated in the content (without keyword stuffing)?
• Content is well structured (short paragraphs, subheadings, lists, highlights)?
• Content readability is high (clear language, understandable phrasing)?
• Clear call-to-actions (CTAs) are present where appropriate?
• Content is regularly reviewed and updated when necessary (content freshness)?

3. On-Page: Structure & Metadata:

• Each page has a unique, meaningful, and CTR-optimized title tag (including main keyword, < 60 characters)?
• Each page has a unique, compelling, and CTR-optimized meta description (including main keyword, approx. 155-160 characters)?
• Correct and logical heading hierarchy (exactly one H1 per page, followed by H2, H3, etc.)?
• H1 heading contains the main keyword and concisely summarizes the page topic?
• Images have descriptive file names and relevant, keyword-containing ALT texts?
• Images are compressed for the web and have an appropriate file size?
• Internal links are logically placed, use meaningful anchor texts, and link to relevant subpages?
• External links refer to trustworthy and topically relevant sources (if applicable)?

4. Off-Page Factors:

• The backlink profile consists of high-quality, topically relevant links from authoritative domains?
• There is a healthy diversity in the link profile (different link types, different sources)?
• Link growth appears natural?
• There are no or very few toxic or spam links in the profile (regular checking)?
• The brand is mentioned positively online (brand mentions)?
• There is a positive online reputation (reviews on relevant platforms)?

5. User Experience (UX):

• The website offers intuitive navigation and user guidance?
• The time spent on important pages is appropriately high?
• The bounce rate (especially the return-to-SERP rate) is low?
• The click-through rate (CTR) in search results for important keywords is good?
• The website design is appealing and supports content consumption?
• Interaction possibilities are accepted by users?

6. Local SEO (if relevant to the business model):

• Google Business Profile is fully optimized (NAP, opening hours, categories, services, photos, posts)?
• NAP data (name, address, phone) is consistent across all platforms?
• There are positive and current customer reviews on GBP and other relevant portals?
• Local keywords are used in content and GBP profile?
• There are entries in important local and industry-specific directories (citations)?

7. Future Orientation:

• Content is optimized for voice search (conversational queries, W-questions)?
• The website has an active and maintained presence on relevant social media channels?
• The content strategy considers topic clusters and semantic relationships?

[bmadcode]

Wow @Geserit2 I had not even thought about this - great idea, and this TBH is one of my weaker areas so this is really helpful. So you will in prep of building your app, work with the pm or design architect (v3) to detail this landing page, and then run the checklist against the plan - or are you building the page or stories for the apge, and then assessing with the checklist?

[Geserit2]

I'm still experimenting with it, but I gave the architect this list and told it to include these points in the planning when creating the high lvl epics and stories. On the one hand, some things have to be taken into account when choosing the technology, e.g. if the landing page is built in React, this has SEO disadvantages that can be solved by prerendering. Other points are, for example, something for the acceptance criteria or others when creating the right texts. However, I derive these texts from my personas and these texts are not only tailored to the target group, but also optimized for SEO.

[jasonvanwyk]

Great stuff Geserit2! thank you for this!

[wfjt]

Foremost, thank you so much! I'm still on v2 but wanted the highlight story ordering instructions in the SM prompt and sequencing in general:

• In my case, I introduced v2 to an existing service-layer oriented Python codebase. I started in Cursor.
• When I got to the SM, it analyzed the story dependencies and concluded that 1.3 must be implemented first. This was correct.
• Then it went something like 1.3 -> 1.5 -> 2.1 -> 1.2 -> 1.1 -> 1.4 -> 2.2 -> ... all based on correct analysis of dependencies. So in v3 SM template, the instructions to assume numbering implies ordering would NOT work in my project. I did do a full PO validation, and it did not flag any sequencing issues.
• I'm sure you know, in agile it's common in more complex projects to run into blocking stories in later epics that must be pulled from the backlog before the earlier epic can proceed. I think handling this should be in v3 or later version. Emphasizing, having dependencies clearly mapped in epics/stories is essential in my opinion. Dependencies should define ordering, not file naming conventions. Things fluctuate too much in any non-trivial project for ordered numbering to work.
• What I also found helpful was to have the SM generate a Mermaid diagram of story dependencies. I can link it in Cursor on every new SM chat when the next story is to be picked, and it's of course, updated if, e.g., arch artifacts have been updated due to changing requirements or discoveries.
• This is not an expectation for v3 or BMAD method in general, and I know you mentioned using MCP as an interface before as a planned future refactoring. However, once I got the time, I'm going to see how this works with Linear or Jira MCP. And GH issues as well.
• I also noticed the SM in v2 assumed the n.m numbering implied ordering in some cases, and I needed to update the prompt. I'm also using 001 instead of 1, just a personal preference, not a proposed change.
• I'm also using a short desc in names to make it easier to manage for me, and AI seems to have no issues with this. I just replaced, e.g., epicN with epicN[-desc].md and worked fine in Cursor. It could be this would be an issue in Gemini web, haven't tested it yet.
• Another thing I do is export Cursor chats to docs/chats mostly for humans, and AI can analyze them to generate a development narrative summary, which I found useful, at least as a human. I sometimes pass it to agents when starting a new chat so they understand the context fully. Of course, I also use Cursor rules generated from relevant chats as applicable. Maybe could be added as an optional tip to instructions or README or whatever? I tend to forget why some particular change was made two weeks ago when now it seems wrong. Being able to read the relevant chat outside Cursor is helpful, also for sharing context in a human team and for transparency in general. Especially in an open-source project where using Cursor team functionality is not used, and in internal orgs where IDE choice is up to the dev.
• These chat logs also enable analyzing failure modes/loops better as often the context spans multiple chats to really understand and generate better rules, and update docs as well.

I also had to update Cursor agent instructions to always use date before updating changelogs or anything else with dates. Without this, the arch, for example, would take the current date I gave it and override all previous changelog dates with it. I had to have it analyze git logs to restore the mess (I didn't notice this the first few times it did this and committed the wrong dates).

What are your thoughts on sequencing? Should this be sorted out before epics ever reach the SM, or should BMAD method follow Scrum/agile more closely (as I had to do in my case)? I haven't tested v3, so maybe this is not even an issue anymore.

[bmadcode]

@wfjt this is so insightful - this really would be the best way - would also make changing up the sequencing easier. I will say when working with the PM on the PRD, I really try to help it get the ordering right - but there will always be some detail. Even in the v3 beta demo output - after I thought I had it sequenced all correctly - it still generated a story where it just staight up said, will call X function (TBD 2 stories from now) Part of it was because I introduced some bad suggestions that confused the llm that i then didnt notice in the push to finish the demo output quickly that it spread throughout and kinda ruined the planned updated stories for a few epics!

[bmadcode]

a lot of great ideas here though from you both - I will push a few more minor fixes just to fix the level of detail issue I am still tweaking. I think getting in this sequencing though is doable for V3, ill try.

oh BTW - @wfjt - I have this all working (with some unpublished modifications) really well with custom Trello MCP (nothing special re the specific mcp for trello - I just wanted to try making trello mcp since they have such an easy to use API)! For 1 person or very small team, the artifacts in the folder work well enough - but now I am thinking of going back to trello, as reordering things gets pretty easy when its in that system

I have used Cursor + MCP + enterprise Jira and the integrations are rea with some modifications with MCP + jira - pretty cool stuff!

[jasonvanwyk]

Brian - thank you for your Bmad Methods again and for starting this project, YouTube and repository and making it available to all of us aspiring developers! And of Cource, I just also want to thank of of you others out there contributing to this - I't's really awesome!

Things are moving fast, I just finished a website using the origional V1, my next project is goiong to use V2, but i see we are already getting onto V3 BETA! - such good stuff!

I guess the only think a can contribute right now, is my 50c - I agree, we need that 1M context window that Gemini 2.5 offers! What i have discovered is that these models all seem to be carving out niches for themselves, and I find that I'm using more and more different models for different needs, e.g. Gemini 2.5 for big context window, GPT for images etc.

Like most, ive been IDE hopping, but i see big potential again for terminal based AIDER, but it's still hard to see a clear winner here. My take, we should focus on owning what we do to avoid vendor lock in (We also need to make a living, they cant have the whole cake to themselves).

finally, I like the control these new web agents give us. I actually feel like i'm contributing as an engineer! or should i say, i'm able to x10 my junior engineering skills. I still dont trust YOLO mode - I like to do things step by step, yes, it slows me down a bit, but I get to wrestle with all the ins and outs of my project - I truly own it! and I can speak to it!

Thanks all!

[bmadcode]

Yes @jasonvanwyk exactly agree with your points - and again thank you also - I honestly did not expect much engagement on these ideas, so I again just cant say how happy it makes me to see others resonate with this. But I am also happy people are taking the message of pragmatism, there is no one true way, gotta figure out what works individually, models and context and prompting skills at the end of the day are all that really matter in this crazy new ai dev world. Unless some crazy clear value winner emerges that does it all (lets see what windsurf does with that openAI $$$ lol) or the AGI uber model comes - this is all about avoiding vendor lock in. When the agi arrives - well I guess then its time to take up wood working again LOL.

I like to use a combination of cursor, roo and copilot (mostly to see how its coming along) - each has strengths. I still think value and when I dont mind slow requests, cursor can be good when choosing its cheaper models for example and know they can get the job done.

MCP is really the next step for all of this, a way to really keep it independent and easily usable regardless of the tool - although with the v3 addition of tasks, that really sets up a potential good path forward also that doesn't require ide specifics.

But if this is going to get to using true agent orchestration with a bit less human intervention (which I am not really on board there yet, but maybe it will come) - MCP or other agent or orchestration frameworks will have to play a bigger role.

Totally agree @jasonvanwyk re yolo - i added it as requested, but I do not use it either. I do use the cursor yolo mode, but I also work in very small increments and always commit and push between working features.

And in brownfield, I do a lot more hands on coding - as a long time engineer, I still enjoy actually writing the code, planning it, designing the patterns - and using the agent more as a paired partner or tool - or even just using the inline assistance for the file I am in.