Big Idea

The argument in one line.

Vibe coding at production scale means matching the number of agents and the amount of planning to each task's difficulty, then running several tasks in parallel instead of babysitting one.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You already use Claude Code or Codex and want to see how someone runs several agents at once without them stepping on each other.
You're building a real product with AI agents and want a rule for when to use plan mode, sub-agents or a quick one-line prompt.
You want to see a full-stack job (Postgres, API, AWS, OAuth login) done end to end by prompting, including the production bug at the end.

SKIP IF…

You want a beginner explanation of what vibe coding is or how to install the tools.
You want a cheap setup. The presenter runs four paid AI subscriptions and uses his own product suite throughout.

TL;DR

The full version, fast.

The presenter rates every task from 1 to 10 for difficulty. Easy tasks get one agent and a spoken prompt. Harder ones get more agents, a plan first, and a stronger model. He builds a NestJS and Drizzle API with GPT 5.5, a sign-in page with Claude Opus 4.8, a Jest test suite, and an AWS setup through the CLI, all running at the same time in separate workspaces. Handy habits include @-referencing project folders, annotated screenshots, voice dictation, bypass-permissions mode, and handing browser chores to a browser AI. Production broke once, and pasting the console errors plus a screenshot back to the agent fixed it.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 02:42

01 · Subscriptions and tools

The four AI subscriptions, the agent workspace app, and the difficulty-to-agents rule from 1 to 10.

02:42 – 04:04

02 · Setting up the workspace

Opening multiple workspaces, launching four Claude and four Codex agents, and why a voice-to-text tool matters.

04:04 – 06:10

03 · Voice prompting and context

Dictating a landing page task, adding the project directory with @, and organizing sibling projects in one parent folder.

06:10 – 09:02

04 · Building the API and database

GPT 5.5 plans a NestJS, Drizzle and Cognito backend while Claude builds the sign-in UI and a test framework in parallel.

09:02 – 12:30

05 · AWS infrastructure and unattended agents

An 11-phase AWS build through the CLI in bypass-permissions mode, plus a Jest test suite, with git worktrees discussed.

12:30 – 14:00

06 · Screenshots as context

Annotating a screenshot to ask for a search bar and seeing it land correctly.

14:00 – 17:30

07 · Discord auth via a browser agent

An agent writes a handoff prompt, a browser AI creates the Discord application, and the frontend gets wired to the API.

17:30 – 22:44

08 · Production bug and final result

The first deploy errors, the agent fixes it from copied console output and a screenshot, and sign-in works live.

Atomic Insights

Lines worth screenshotting.

Rate each task from 1 to 10: a level 1 task needs one agent and no plan, a level 10 task needs many sub-agents and a written plan.
If you only work on one task at a time while agents are available, you're leaving most of the leverage unused.
Referencing a sibling project's folder in the prompt lets the agent copy a pattern you already got right, like a migrations setup.
Typing @ plus a folder name makes the agent list and read that directory, so it starts with the right context.
Git worktrees prevent agents from overwriting each other, but they add a merge chore, so skip them when agents work in different parts of the code.
Running agents in bypass-permissions mode lets you walk away from a 30 to 60 minute infrastructure build.
Using the AWS CLI with a local profile avoids the extra files Terraform or CloudFormation generate that you then have to manage.
An annotated screenshot with an arrow pointing at where a change goes cuts hallucinations because the agent sees exactly which element you mean.
Ask one agent to write a handoff prompt for a browser AI, so the browser agent has full context for chores like creating an OAuth app.
Start thinking about the next task while the current one runs, especially anything you've never done before like Discord login.
The first production deploy usually breaks. Copy the console errors, add a screenshot, and give them to the agent that wrote the code.
Pick the model by job: GPT 5.5 on extra high in fast mode for backend work, Claude Opus 4.8 for UI design.

Takeaway

Scale agents and planning to the task

WHAT TO LEARN

Agent work goes faster when you rate each task's difficulty, size the setup to match, and keep several tasks running at once.

01Subscriptions and tools

Rate a task from 1 to 10. Level 1 edits need one agent and a spoken prompt, while level 10 jobs need sub-agents and a plan first.

03Voice prompting and context

Point the agent at a sibling project that already solves the problem, like migrations or auth, so it copies a proven pattern.
Give the agent exact folders with @ references so it loads the right files instead of guessing the project layout.

04Building the API and database

Choose models by job: a strong coding model for backend, a strong design model for UI, and plan mode for anything touching several systems.
Run separate agents on separate parts of the codebase, and reach for git worktrees only when two agents could really overwrite each other.

05AWS infrastructure and unattended agents

Use plan mode and a stronger model for infrastructure, and let the agent run in bypass-permissions mode so you can leave it for an hour.
The AWS CLI with a local profile builds infrastructure directly, without the extra Terraform or CloudFormation files to maintain.

06Screenshots as context

Annotated screenshots show the agent which element you mean, which cuts wrong guesses on UI changes.

07Discord auth via a browser agent

Have one agent write a handoff prompt for a browser AI, so one-off chores like creating an OAuth app run without you clicking through.

08Production bug and final result

Expect the first production deploy to fail. Copy the console errors, add a screenshot, and hand both to the agent that wrote the code.

Glossary

Terms worth knowing.

Vibe coding: Building software by describing what you want to AI coding agents and steering the result, instead of writing the code by hand.
Plan mode: An agent mode where the AI writes a structured plan for you to approve before it changes any files.
Git worktree: A separate checkout of a repository so an agent can work on its own copy without touching other agents' files. You merge it back later.
Handoff prompt: A prompt written by one AI agent that gives another AI, such as a browser agent, all the context it needs for a task.
Bypass permissions mode: A setting where the agent runs commands without asking for approval each time, so it can work unattended.
Drizzle: A TypeScript ORM used to define database tables and run migrations from code.
Cognito: Amazon's managed sign-up and sign-in service, used here to handle Discord authentication.

Resources

Things they pointed at.

01:00toolClaude Code

01:00toolCodex (GPT 5.5)

01:00toolCursor

01:00toolPerplexity

01:40productBridgeSpace ↗

03:45productBridgeVoice

14:10productBridgeShot

15:30toolComet browser

06:40toolNestJS

06:40toolDrizzle ORM

10:20productAWS (ECR, ECS, RDS, VPC)

06:40linkBridgeBench ↗

Quotables

Lines you could clip.

06:00

“If you only are working on a singular task, you're not a really good vibe coder. You need to be working on multiple things at the same time.”

blunt claim that stands alone→ TikTok hook↗ Tweet quote

12:40

“Use git work trees if you have to, but don't complicate things with git work trees if you don't have to.”

tight, practical rule→ newsletter pull-quote↗ Tweet quote

20:10

“When you are pushing stuff to production, it won't necessarily work the first time.”

honest and relatable→ IG reel cold open↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

This video is going to be a complete step -by -step guide to my vibe coding workflow. I'm going to show you what tools I'm using, what subscriptions I have, and most importantly, the actual process that I go through when building a production software. For those of you that don't know, BridgeMind is currently at $18 ,000 in MRR and gross volume over the past three months is at $74 ,000.

And this is all from vibe coding. If you want to learn how to do this for yourself, just watch until the end of the video and make sure that you drop a like and subscribe to BridgeMind to learn even more in the future. The first thing that I want to talk about is the subscriptions and the tools that I'm using to actually power my workflow.

And one thing I want to note is that you guys do not need to be using the same subscriptions that I'm using. Everybody has their own budget. I am a power user and I'm using these tools like 12 to 16 hours a day.

So for me, I do not want to hit any limits. So I have multiple subscriptions from multiple providers and here is what they are. So I have a $200 a month cloud subscription.

OpenAI subscription, Cursor subscription, and Perplexity subscription. Now, I label these as AI subscriptions because I use Perplexity for research. I use Cursor for some different tools that they have, and I'll get into that in the video.

I use OpenAI to access GPT 5 .5 through Codex, and then I use Cloud to be able to access their frontier model, Cloud Opus 4 .8 via Cloud Code. Now, those are the AI subscriptions that I use to actually get the access that I need to the models, but there are tools that you should be... using if you are a vibe coder the first tool that you need to be using in your vibe coding workflow is an agentic development environment and I'm going to show you the agentic development environment that I use which is bridge space and it looks like this so as you guys are familiar with quad code and codex one thing that you need to know about these tools is that you are able to work on multiple tasks at the same time what has become incredibly important for workflows is that you get really good at passing off tasks to multiple agents and then managing those agents until the completion of those tasks.

This is the way that I approach it. As a task gets more difficult, that will make it so that you need more agents and also need a plan. So for example, hey, if a task has a difficulty level of one, you just need one agent and you don't even need a plan.

For example, like a task with a difficulty of one is like, oh, I need to go change. the color of a button right or oh i need to change like this copy or i need to change this text or i need to adjust this here you know this is not something that's going to need more agents and it's not something that is going to need a plan you can just speak your intent with your voice to text tool and it's going to be able to get it done now as that difficulty goes up like let's say to a level five you're gonna need more agents so that they can kind of like gain in a better understanding of the code base and then work together to be able to to get that task done.

A plan might be needed, but if you go to a level 10, hey, you're gonna need a lot of sub -agents, and you're definitely gonna wanna abstract away from the initial prompt. You're gonna wanna create a plan. So I wanna show you guys how I do this from scratch.

So I just opened up BridgeSpace, and what I'm gonna do is I'm gonna select this right here, BridgeSpace. I'm gonna click on this, and I'm just gonna open up a preset, which is already set to a particular directory on my computer, BridgeMine. And I'm gonna open up another workspace because I'll work with multiple workspaces at the same exact time So now what I want to do is now I want to CD into let's see into vibe Academy Okay, so now I'm in vibe Academy and let's say I want to launch eight agents and let's say I want to do I want to do half Claude and then I want to do half codex Okay, so I'm gonna launch for codex for Claude's and we're gonna open up these in an additional workspace I'm about to start prompting these agents So I just want to highlight another tool that you really do need to have in your vibe coding workflow, which is a voice to text tool.

I use bridge voice. So I'm going to go back over to bridge space here and I'm going to begin prompting my quad code agent. I need you to launch three deep dive agents to do a thorough review of the website's landing page and note that we need to add in sections for demo videos for bridge agent, bridge voice and bridge space.

Note that currently we do not have sections on the homepage for demo videos. So I need you to use the front end design plugin to create these new sections, which should have the icon of the app and the name of the app. And then the iframe below for now, it can just be a placeholder.

Okay. So you guys just heard my prompt. And then what I'm going to do is I'm going to add the website directory that I'm actually working in.

Okay. And you can even go as far to do an ultra think as well. And I'm going to submit.

this prompt but i want to show you guys that hey this launched three deep dive agents and this will work for about the next five to ten minutes as it reviews the website comes up with a good way to design it and there's a couple things i want to draw your attention to that i did when i prompted that okay one important thing is how i'm organizing my workspace so if you look over here you can see that in this bridge mind directory there are a ton of different projects there's bridge mind api there's the bridge mind app there's the bridge mind mc There's the bridge mind UI and I do this for a purpose I do this because it makes it very easy for me to then go to that particular directory on my machine and then I'll just do an app bridge mind UI and average mine API and average mine MCP and then I would be able to just prompt and Claude would know that okay.

These are the directories that are being referenced So watch this when I enter this you're gonna see did you see how it says listed directory listed directory listed directory? So if you do an at and you add a particular folder or directory or file.

It makes it so that the instance, your coding agent, has the context that it needs to be able to properly pull all the files that it needs. Now that we have some of the basics out of the way, like setting up your workspace, speaking your intent, and adding context, I now want to work on a fairly challenging sprint that I've been planning on doing for a while now.

So we're going to be working on BridgeBench v3. This is the world's number one vibe. Coding Benchmark.

It's been reposted by Elon Musk four different times. And building out BridgeBench v3 is actually a goal on my August 31st goals. We're going to build out a database, we're going to build out an API, and then we're also going to be building out AWS infrastructure all in this video.

And I'm just going to do the best that I can to walk you guys through this step by step using these principles. The first thing that I'm going to do is I'm going to be building out the API and database. use gpt 5 .5 extra high fast extra high is by far the best model for back end and when you put it in fast mode it kind of gets around the slowness of gpt models but let's give it the prompt so i'm going to add that context i'm going to do bridge bench ui and i'm going to add bridge bench as well i need you to do a quick review of the bridge bench ui as well as bridge bench and note that we need to begin building out a back end and database for bridge bench you can call it and then i'm just going to give it the slug that i want to call it which is bridgebench api this needs to be a nest js api and i want to use drizzle for the orm and i want the orm and the migration system to be built into the api you can reference vibe academy api for how this is done so there's another project that i have that does this really well and i think that once you start to build up kind of like a portfolio or you have things working in other projects it does help when you are building something from scratch if it is similar to something else that you've built i will reference

something else that I've built that has done it correctly. So I'm going to reference Vibe Academy API because it does that really well. I need you to start out with a very simple MVP with a singular sign up and sign in endpoint that uses Cognito for the auth system.

Note that the only way that we are going to allow users to authenticate is via discord. So I need you to assist me in building this out. There should only be one table called users and let's create.

a structured plan first. So you guys can see that this is going to be a difficult task. So I'm going to put it in plan mode because, hey, if I just had it go out and build it, it probably wouldn't do that good of a job.

This is a fairly difficult task because you're thinking about this and you're saying, okay, it's doing database, it's doing API, it's looking at another project, it's going to use Cognito. So it's a fairly difficult task. So I'm not going to use any sub -agents on this one.

I'm just going to let GPT 5 .5 work and we'll see how it does, but this I would say is probably about a difficulty level of like six. So this one is now building out the API in the database.

And this is a perfect example of what I referenced about being able to work on multiple tasks at the same time. If you only are working on a singular task, you're not a really good vibe coder. You need to be working on multiple things at the same time.

So the next thing that I'm going to do is I am going to reference bridge bench UI, and I'm going to use Claude code for this because Claude Opus 4 .8 is one of the best UI. design models right now. So you're going to want to use it for UI.

I want you to build out a page for signing up and signing in to bridge bench. Don't worry about hooking this up to any backend. I just want you to focus strictly on design, make it as simple as possible.

Add in these routes to the sidebar of bridge bench and note that the only way to sign up and sign into bridge bench is via discord. Okay. So that's going to be my prompt.

And this one here, it's not really going to need a whole lot of planning. It's going to be able to do a good job. This is like a difficulty level of like two.

So I'm not too worried about it. Now, one thing to note is I do have my Claude codes in a high effort level. For the most part, you probably should be using extra high when you're using Claude code right now.

I'm not using extra high. I'm using high, but let's switch this one over to extra high. This time I want to actually start building out a testing framework.

So I'm going to, again, reference a bridge bench UI. I need you to assist me in building out a complete testing framework for bridge bench UI, launch two deep dive agents and come up with a structured plan to build out a testing framework so that we can perfect the UI.

All right. So I'm going to throw this in plan mode and then we'll submit this plan. So I now have three agents that are working for me, right?

And what's really important when you're vibe coding is that you remember exactly where each agent is working. And this can be a challenge, but one of the most common questions I get asked is, hey, how do you run all these agents without them like conflicting with one another? And you can use get work trees, okay?

If you need to use get work trees, you can. And a get work tree, basically, if you don't know what that is, it's a way for you to be able to have an agent that runs on it. its own isolated work tree that is not going to interfere with any of the other agents.

And bridge space makes this really easy. But the only thing about this is that if you do work in a get tree, you are going to have to handle actually getting that merged back into whatever branch you want that merged into, but it will help for the overwriting. But most of the time, I think that get work trees can be a little bit overkill because there's no way that these three agents are going to overwrite each other right now.

This one's working in the API. This one's working on the UI. a test suite and this one is working on like the sidebar so they're not going to conflict with each other so remember where they're working and it looks like our API is done now but remember where they're working and use git work trees if you have to but don't complicate things with git work trees if you don't have to So now the next step is actually building out our AWS infrastructure and getting this into the cloud.

But before we do that, I do want to create a robust test suite. So I'm just going to prompt this agent down here. Great work.

I now need you to build out a complete test suite that uses Jest and any other packages that you deem necessary. I want you to do a thorough review and build out a scalable test suite that has full coverage and full unit test coverage across all. all endpoints and services.

Okay, so I'm gonna give it this prompt and I'm just gonna have it build it out since we only have two routes right now, a sign in and sign up. I'm not too worried that that's not gonna be super complex. But while this works here, I'm gonna go over to this cloud code agent now and I'm gonna reference the new BridgeBench API, okay?

So you can literally see, if you look here, you can see BridgeBench API, right? So I'm gonna reference BridgeBench API and I'm now going to give it the prompt that it needs to actually build out the AWS infrastructure. I need your help in building out the AWS infrastructure for BridgeBench.

I need you to review the API and note that we want to use Docker and we want to use ECR, ECS, and we also would like to use RDS for a Postgres database. And we also will need to handle the migrations for the database. I need you to review Vibecademy API and how this is done in Vibecademy API for the migrations.

But you can... start to plan out a structured AWS infrastructure for BridgeBench. Completely review the API.

The UI is in Vercel, so don't worry about that yet. But let's build out the VPC. Let's build out the security groups.

Let's build out the ECR. Let's build out ECS. Let's build out the load balancer.

And I need you to create a structured plan first to build out a complete MVP for our AWS infrastructure using best practices. and integrating cybersecurity best practices. Okay, so we're going to do a plan first, okay?

And honestly, what I'm gonna do is actually, I'm gonna switch this up because this is a difficulty level of like 10. So what I'm gonna do is I'm actually gonna switch this prompt right here and I'm gonna paste that prompt. What's nice about Bridge Voice, it does save your prompts once you speak it.

So I can go over to Bridge Voice anytime and just copy those prompts. It's very nice. But let's go over here.

This now has that pasted prompt. I'm gonna put... Bridgebench API backend.

I will reference Bridgebench UI just so it has it, but then I'm also going to give it this. I want you to use the, and then I do have a, it's called like an AWS profile, okay? But I do have an AWS profile on this computer.

AWS profile. on this machine to actually access my AWS account and build out this infrastructure using the CLI. So I'm going to use the AWS CLI.

I think that this is the fastest way to build infrastructure. A lot of people use Terraform or CloudFormation, but sometimes this just builds a lot of extra things to manage. And once you build your AWS infrastructure, if you're using agents, and one thing I want to do actually is I do want to launch it in plan mode.

And I'm going to say, I want you to launch five agents to assist you with this task. This cloud code agent is now provisioning my AWS infrastructure, and this is an 11 phase plan. So it's going to take a while for this to build out.

But one thing that I want to highlight just for you guys to know is that I never have my cloud code agents or my codex agents in like plant, like where they're asking me things. I always have them in YOLO mode or in bypass permissions mode because it allows you to literally just be able to walk away. and go somewhere else.

And, you know, I'll come back to this, you know, in like, this will probably take about 30 minutes to an hour to actually provision all of this infrastructure correctly. So I'm going to be able to walk away now. I think this is a great example of being able to work on multiple tasks or projects at the same time, because now that this is working and working on provisioning the infrastructure, there's nothing else with BridgeBench that I necessarily want to work on right now.

So for example, this is when I would just go over to another project like vibecademy and we could start working with vibecademy while we wait on the aws infrastructure in bridgebench to finish provisioning While we wait for the infrastructure to finish up, I do want to show you guys another tool that I do think that you should be using in your vibe coding workflow, which is a screenshotting tool.

Okay. I use BridgeShot, which is a product in the BridgeMind ecosystem. The reason a screenshotting tool is so important is because it allows you to add context to what you're looking at.

So for example, let's say that I want to add a search bar on this page here. I can do a screenshot. I can go down here and I can draw an arrow.

to where I want the search bar to go, and then I can say, I want there to be a search bar here for me to be able to search models. So I'll just enter this, I'll save it, and then I'll go back over to an agent. Let's pull back BirdSpace, and I'll just drop in this screenshot.

I need you to add a search bar here, respectively. And then again, I'm going to have to add BridgeBench UI just so that it knows, okay, it's this project that he's talking about. We'll enter that.

And that, again, that's going to be a super easy prompt now that I added the context. But if you do use a screenshotting tool and you add that context, I promise you, it will cut down on hallucinations so much. Here's our search bar.

Let's look up Claude fable five and boom, there it goes. And it works perfectly. Another thing that I'll do when I'm vibe coding is I like to think ahead about the next thing that I'm going to need to do after one of the tasks that I'm doing is finished.

So for example, with our AWS infrastructure, one of the things with bridge bench that you guys heard me ask for is I want to make it so that users are able to authenticate using discord. So I know that that's going to be a little bit of. a process in itself.

I've actually never done that before. So we're going to start working on that with this codex agent down here. I need you to assist me in thinking through how we can actually make it so that users can sign up and sign in using discord.

What is the process for this? Do I need to create a discord application? So let's first figure out how to do that because I actually don't know how to do that.

So I'm just going to ask the agent that worked on building out that functionality in the first place what we need to do okay the agent came back quick and what it says is that you do need a discord application so what i'll do now is i need you to write a handoff prompt to the ai that is going to be working in the browser to create this discord application so you guys are actually going to see a really cool workflow that i use and it's using comet so i told you guys at the start of the video that i use comet and what i do is i'll actually use codex or I'll use Claude code to write handoff prompts for when I need to use AI in the browser or do anything in the browser so I just had it write this handoff prompt I'm gonna copy all of this and then we're gonna go over to comment and actually have AI set up this discord application for us so here's comment I'm currently in the discord application portal and all you have to do is click assistant over here and then I can submit that in

higher prompt and then I can go over here and I can just say control browser. Do you guys see this button right here? Control browser.

So I'll click that and I can then submit this and you guys will see how this works. I think this is so cool. There are other ways that you can do this, but comment will basically take control of your browser and then it'll begin actually creating this discord application for us and we won't have to do anything because we already gave it a handoff prompt with all the context that needs to create the discord application the discord app has been created using comment and it's now time to actually wire all of this in and get it deployed so i'm going to prompt this agent i need you to now wire the front end bridge bench ui into the api so that we are actually able to sign up using discord respectively i'm going to reference the bridge bench ui and the bridge bench api and then i'm also going to put in ultra think on this one ultra think does make it think harder.

So we're going to submit that and we're going to wire all this in and then we're ready to test. Okay guys. So I did just push this to production and I'm getting an error in production here.

When I click sign up with discord, you can see that I'm getting some errors here. And what's cool is that all you have to do is that you have to copy those and then just drop them in to the agent and then drop in the screenshot. And I actually already did that and it found it.

It found the bug. You can, if we move up here, it says found it. This was the error.

And then it fixed it and then it just pushed it to the main branch and we'll be able to see if this is now fixed. But typically when you are pushing stuff to production, it won't necessarily work the first time.

You'll have to do a couple things to actually get it working in production. So make sure that you do use browser tools and that you're giving your agent the context that it needs to fix any issues that are happening in production. So I deployed that fix and I just attempted to sign up again and now it's actually working.

Okay, look at this guys. So I am going to see this is gonna be the first time.

Can we actually authorize? Let's see, guys. This is gonna be huge.

Let's see. Wow. Okay.

Look at that. Look at that. My account right here, bridge mind official.

That's my discord account. And guys, this works perfectly. So this is, um, this is how you build software in today's world.

You guys can see the account right here, bridge line official. So in this video, we literally built the entire stack. We built the database.

We built the API. We built the AWS infrastructure and we sorted out some issues that we had, but you can see that I now am able to actually sign up for. BridgeBench and log in and log out.

This is amazing. So we are going to be working on BridgeBench V3. Like I said, this is a big goal.

And one of the steps that we're taking towards BridgeBench V3 is actually allowing people to sign up for BridgeBench and participate in different things that we're going to create for BridgeBench. So go check out BridgeBench .ai, go sign up using Discord. It is live now.

So I'm super excited about this. That's going to wrap it up for today's video. If this was helpful for you in any way, shape or form, make sure you drop a like on the video and subscribe to BridgeMind and join the BridgeMind Discord community, which is the largest and fastest growing vibe coding community on the internet right now.

It has like over 12 ,600 builders. There's going to be a link in the description below. But with that being said, I will see you guys in the future.

The Hook

The bait, then the rug-pull.

One person, eight terminal panes and a pile of AI subscriptions. The video shows how that setup turns one afternoon into a database, an API, cloud infrastructure and a working Discord login.

Frameworks