00:00:00
[Music]
00:00:02
If you haven't been vi coding enough,
00:00:04
you must have experienced problem where
00:00:06
you ask cursor to implement a small
00:00:08
change but just map up your whole
00:00:10
project or cursor is not aware of all
00:00:12
the dependencies in your codebase and
00:00:14
implement something that leads to loads
00:00:15
of errors. This is a very common issue
00:00:17
of AI coding agent in general no matter
00:00:20
which platform you're using. But there's
00:00:21
one technique that show promising
00:00:23
improvements that will make your cursor
00:00:24
make way less errors, which is giving
00:00:26
your AI coding agent a task management
00:00:29
system. It helps it to understand
00:00:30
overall implementation plan and also
00:00:32
control the amount of context goes into
00:00:34
each step when it is implementing
00:00:36
specific tasks. I was able to build a
00:00:38
fully functional multiplayer online
00:00:40
drawing game where both players can draw
00:00:42
image of given work and we send the
00:00:44
result to GBD4 where it look at image
00:00:46
and give evaluation and pick up the
00:00:48
winner. And this whole game is actually
00:00:50
imprinted by cursor with just one shot
00:00:52
without much arrow which is absolutely
00:00:54
insane. That's why today I want to show
00:00:55
you what my workflow is and how can you
00:00:57
adopt it for your own project. And
00:00:58
before you use any of those tools I just
00:01:00
introduced people has been hacking
00:01:02
together task management workflow to
00:01:04
improve performance for a while. At its
00:01:06
core it basically means you ask cursor
00:01:07
to break down your complex PRD into
00:01:09
small tasks and have a document where
00:01:11
cursor can have access to to track and
00:01:13
maintain what tasks are coming and what
00:01:15
tasks already been done. And this is a
00:01:17
quick example from Elle. The most basic
00:01:19
implementation is in your cursor project
00:01:21
you will create a cursor rule and it
00:01:23
looks something like this. Basically
00:01:24
rule where you tell cursor to always
00:01:26
refer to task.md to keep track about
00:01:29
what task they already been done and
00:01:30
what task haven't. And with this one we
00:01:32
can just create a task.md file and give
00:01:35
a prompt I want to build a x app help me
00:01:37
breaking down into small task of our
00:01:39
core me feature and add to task.md. So
00:01:42
cursor will create a list of task here
00:01:44
and after cursor finish every single
00:01:46
task it will just come back and mar
00:01:48
those task as completed. So it has a
00:01:50
context of overall implementation plan.
00:01:52
With this method it already help a ton
00:01:54
for executing complex task with cursor.
00:01:56
But tools like cloud taskmaster and
00:01:58
boomeran task from ruk code bring even
00:02:00
more sophisticated task management
00:02:02
behavior into the AI coding agents. For
00:02:04
example, for task master AI, it is a
00:02:06
command line package that you can run in
00:02:08
cursor windsurf data where it utilize
00:02:10
cloud 3.7 or more advanced model to look
00:02:13
at PRD you have and break that down into
00:02:15
small subtasks by running a simple
00:02:17
command line like taskmaster parse PRD.
00:02:19
And what's really amazing about it is
00:02:21
that it will breaking down tasks in
00:02:23
logical order consider all the
00:02:24
dependencies between different tasks. So
00:02:26
you won't have situation where cursor
00:02:28
implements something but require other
00:02:30
dependencies that hasn't been
00:02:31
implemented yet. It also has useful
00:02:33
command line like analyze complexities
00:02:35
to use perplexity and cloud to analyze
00:02:38
how complicated each task is and if
00:02:40
certain task complexity score is very
00:02:42
high it will allow you to expand on
00:02:44
those task further and by breaking down
00:02:46
those complex tasks into even smaller
00:02:48
bits. The success rate of it delivering
00:02:50
a functional application just increase
00:02:52
dramatically. And Roco's boomeran task
00:02:54
also something similar. It gives AI
00:02:56
agents tools like new tasks to breaking
00:02:58
down a complex project into small bits
00:03:00
and keep track about progress. Those
00:03:02
tools have completely changed my
00:03:03
workflow. So I'm going to quickly take
00:03:04
you through what's my new best practice
00:03:06
vibe coding workflow with those new
00:03:08
tools. But before we dive into that, I
00:03:09
know many of you are trying to build AI
00:03:11
agents for business. But there are many
00:03:13
pitfalls I saw people fell into at
00:03:15
delivering successful production agents.
00:03:17
That's why I want to introduce you to
00:03:18
this research HubSpot did where they
00:03:21
interviewed tons of business and
00:03:22
startups who have been launching AI
00:03:24
agents for the past 12 months to
00:03:26
understand which AI agent use case
00:03:28
actually drove huge amount of business
00:03:30
value and ROI versus ones that sounds
00:03:32
fancy but actually very difficult to
00:03:34
deliver value and which signal customers
00:03:36
are deploying huge amount of budgets to
00:03:38
buy AI agent solution. They include lots
00:03:40
of real world success stories and
00:03:42
articulate those learnings into
00:03:43
frameworks that you can use to build
00:03:45
your next agents from which use cases
00:03:47
more suitable for chatbot versus actual
00:03:49
autopilot agents. How do you determine
00:03:51
which task is best for AI agents versus
00:03:54
more traditional workflow automation as
00:03:56
well as list of common pitfalls that
00:03:57
many other people encounter including
00:03:59
myself when deploying production agents
00:04:01
and best practice of how other people
00:04:03
resolve that like what's the best
00:04:05
practice for you to build integration
00:04:06
into existing systems. This helped bring
00:04:08
a lot of clarities and many mistakes I
00:04:11
personally experienced. So if you're
00:04:12
planning to build agents, I highly
00:04:14
recommend you go and have a read and
00:04:15
this is totally free. I have put a link
00:04:17
in the description below for you to
00:04:19
download. Now let's talk about my new AI
00:04:21
coding workflow with task management
00:04:23
systems. Firstly, let's talk about ru
00:04:25
code's boomer task feature. And if you
00:04:27
don't know ru code, ro code you can
00:04:28
almost consider as a open-source cursor
00:04:30
that live inside visual studio code. It
00:04:32
is totally free to use. All you need to
00:04:34
do is just provide your own entropy key
00:04:35
and then it will just work like any
00:04:37
other AI coding agent that you've been
00:04:38
using. But what's really cool about real
00:04:40
code is that unlike cursor where you
00:04:42
only have a few predefined agent mode,
00:04:44
Ro allow you to create your own modes
00:04:45
like at default they will provide coding
00:04:47
agent architect agent that will help you
00:04:49
do the planning debug agent that help
00:04:51
you figure out where the arrow is. But
00:04:52
you can also build custom modes like
00:04:54
boomer wrench mode where it will be
00:04:55
focusing on planning and breaking down
00:04:57
the plan into smaller manageable pieces.
00:05:00
Think of it like delegating your work to
00:05:01
specialized assistant. Each subtask runs
00:05:04
in its own context. So I can choose a
00:05:06
boomer range mode that I just customized
00:05:07
and then say help me build a to-do app.
00:05:10
At top you can see that it will keep
00:05:11
track about how many token consumption
00:05:13
is as well as total amount of API cost
00:05:15
and it will firstly delegate a planning
00:05:17
task to the architect agent and this
00:05:19
architect agent will have this system
00:05:21
prompt where it will continuously
00:05:22
confirm with me about the requirements
00:05:24
and then it will start planning out the
00:05:25
project breaking down into specific
00:05:27
feature figure out things like user
00:05:28
story key feature components project
00:05:31
structure state management and many
00:05:32
more. So it has a full understanding
00:05:34
about all dependencies between different
00:05:36
functions. I can give feedback in the
00:05:38
middle and once the plan is finished and
00:05:40
breaking down to small tasks, it can
00:05:41
switch to the code mode to start
00:05:43
generating the code and the code agent
00:05:44
will start executing different tasks
00:05:46
based on the plan and then complete the
00:05:48
actual application for me. And you can
00:05:49
see the result here is very high quality
00:05:51
and it even have functionality built in
00:05:53
where the agent will be able to run the
00:05:54
application in the browser, see the
00:05:56
result to automate testing as well. And
00:05:58
with this one, the result already feels
00:06:00
better than what I got out of box from
00:06:02
cursor. But on the other hand, cloud
00:06:04
taskmaster integrate much more deeply
00:06:06
into cursor and wings surfer. First,
00:06:08
let's install the taskmaster AI. You can
00:06:10
open terminal in any folder and do npm
00:06:12
install-g
00:06:14
taskmaster-ai. And once it's finished,
00:06:16
you can run a few different commands.
00:06:18
One will be taskmaster in it. This will
00:06:20
set up the project inside the folder.
00:06:22
So, you can just do taskmaster in it
00:06:24
directly. But I would suggest you set up
00:06:25
the project first. Like if you're
00:06:27
building a nextjs project with chessen,
00:06:29
you can just do this command. And once
00:06:30
it's done, we can do cursor my app.
00:06:32
Inside here, we can do taskmaster in it.
00:06:35
This will ask for the pent name and I'll
00:06:37
just call it my app. Description doesn't
00:06:39
really matter. You can skip all those
00:06:40
stuff and then just let it set things
00:06:42
up. And what will happen now is that on
00:06:44
the left side you firstly see it add a
00:06:46
few cursor rule. Some of them are
00:06:48
generic one like this cursor rule
00:06:50
basically teach cursor how can it add
00:06:52
new cursor rules. So as you go deep into
00:06:54
the implementation you can ask it to
00:06:56
reflect and creating rules about the
00:06:58
mistake it make for example and it will
00:07:00
follow this rules to create the next
00:07:01
cursor rules and this self-improve is
00:07:03
basically the same thing it kind of try
00:07:05
to get cursor to do this proactively and
00:07:07
the step workflow is where it teach
00:07:08
cursor about all the commands it will
00:07:10
need to actually check all the task in
00:07:11
the backlog and if you're using
00:07:13
windserve there will be windserve root
00:07:14
here as well inside the scripts folder
00:07:16
it provide a structure about what does a
00:07:18
prd can look like but the most important
00:07:20
one is that it will have thisv.exam
00:07:22
example file. What you need to do is
00:07:24
swap out this entropic key as well as
00:07:26
proplasticity key here. So entropy is a
00:07:28
model that will be used to break down
00:07:30
your PRD into small tasks and they also
00:07:32
use perlastity to do some research. So
00:07:34
if part of task is using a new package
00:07:36
that just released then it will actually
00:07:38
use perlasty to fetch the latest
00:07:40
developer documents and include those
00:07:41
into the task information. So I
00:07:43
recommend you add both API key here and
00:07:45
once you did that we can start creating
00:07:47
our PRD. There are many different ways
00:07:49
you can create PRDs. If you're in the AI
00:07:51
builder climate building already, you
00:07:52
will have access to tools like tanks
00:07:54
coder the dev where it will help you
00:07:56
generate PRD automatically and fill in
00:07:58
all the gaps for the features that you
00:08:00
might not think of. So if you're already
00:08:01
in the AI build club, you can use this
00:08:03
tool to get the PRD here. But if you
00:08:04
don't have access, you can also just
00:08:06
chat with cursor agent use that to help
00:08:08
you generate PRD. For example, I can
00:08:10
just say help me build an online game
00:08:12
like Scribble, but instead of a human
00:08:14
guess word, it will be large model guess
00:08:16
a word. So each round all users will be
00:08:18
given a same word and they have 60
00:08:20
seconds to draw the image. In the end
00:08:21
all images will be sent to open 4 and
00:08:24
let it choose which image is closest to
00:08:26
the word. Now play the role as the
00:08:28
engineer manager help me think through
00:08:29
what are the core features of
00:08:31
implementing such game and then it will
00:08:33
spit out the core functionalities.
00:08:34
Obviously I can chat back and forth but
00:08:36
once I finish I can just say great now
00:08:38
let's help me build the core MVP
00:08:40
features requirements into prd.txe using
00:08:43
the example uh prd.txe tx as reference
00:08:46
which is what we have uh showing here
00:08:48
and then you will see a prd has been
00:08:50
created with good amount details and I
00:08:52
will accept that. So now since we have
00:08:53
this prd generated the next step we can
00:08:56
use this command taskmaster parse drd to
00:08:58
breaking down this prd into small task
00:09:00
and this is where the power of
00:09:02
taskmaster begin. So I will do this
00:09:04
taskmaster parse prd
00:09:07
scripts/prd.txt. Okay, so I had this
00:09:10
arrow uh just need making sure you
00:09:12
remove this example and let's do it
00:09:14
again. So now it will start creating
00:09:15
task files based on this prd and you
00:09:17
will see here where we have task folder.
00:09:19
It has all the tasks that created from
00:09:21
taskmaster. What we can do is we can do
00:09:23
taskmaster list and this will show you
00:09:25
the list of tasks that it has been
00:09:27
created. What's really cool about
00:09:28
taskmaster task list is on the right
00:09:30
side you can see here's a dependencies
00:09:32
column. So when breaking down tasks, it
00:09:34
will actually list out the task in a
00:09:36
logical order and making sure there are
00:09:38
clear dependencies mapped out. So when
00:09:40
it implement, it can implement in the
00:09:41
right order. Meanwhile, there are also
00:09:42
some pretty useful commands. You have
00:09:44
this command called analyze complexity.
00:09:46
What this will do is that I can do task
00:09:48
master and analyze complexity. This was
00:09:51
basically send all the task I created to
00:09:53
cloud 3.7 as well as perplexity.
00:09:55
Basically ask it to just evaluate how
00:09:57
complicated or how difficult it is to
00:09:59
implement this feature. And once it is
00:10:01
done, I can do taskmaster complexity
00:10:04
report. It will show me the evaluation
00:10:05
of each task and its complexity score.
00:10:08
But what's really useful is for those
00:10:09
complex tasks, it also give you the
00:10:11
prompt that you can copy. At the moment,
00:10:13
you can't really copy this directly. If
00:10:15
the UI just break up, but you can just
00:10:16
copy the first one and go to complexity
00:10:18
report and find the specific ID which is
00:10:21
this one and we can copy the expansion
00:10:23
prompt. So here it generate a prompt
00:10:24
detailed technical implementation of
00:10:26
HTML 5 canva component including drawing
00:10:29
tools input handling across all devices.
00:10:31
So I can just do this and now it
00:10:33
breaking down that specific task into
00:10:35
smaller ones. And as we know once a
00:10:37
complex task breaking down to smaller
00:10:39
one it is more likely to succeed without
00:10:41
any error. And you can continue doing
00:10:43
that um by adding another one for the
00:10:45
task number five. So you get the drill.
00:10:47
Um basically you can do this process
00:10:49
back and forth for a few times until
00:10:51
you're happy with the backlog here. You
00:10:53
can also do things like update as well.
00:10:55
So if later down the row you decide to
00:10:57
like change the plan, you can also do
00:10:58
taskmaster update ID equal to like four
00:11:02
and prompt equal to something like make
00:11:04
sure we use three js. And what will
00:11:06
happen is that once you give this
00:11:07
prompt, it will actually update the
00:11:09
whole plan based on this new instruction
00:11:11
which is really really helpful. But once
00:11:12
it's done, you can do taskmaster list
00:11:15
with subtasks. This will show all
00:11:17
subtask here directly. So it will make
00:11:19
it easier for you to review all the task
00:11:21
here. So that's pretty much it. Once we
00:11:23
done this, you can just go to the cursor
00:11:25
agent and say let's start implementing
00:11:27
the app based on task we created using
00:11:28
taskmaster. Let's check the next most
00:11:30
important task first. So you might have
00:11:32
this arrow here that it is reusing the
00:11:34
wrong command. Uh so I will just tell it
00:11:36
do not use script dev.js use taskmaster
00:11:39
instead. Just go list all task created
00:11:41
and follow the plan. Then it will see
00:11:43
the plan and decide to implement the
00:11:45
first one first. And as it work on the
00:11:47
task, it will also set the status of
00:11:48
this task to be in progress. And here
00:11:51
since I actually turn on the yolo mode
00:11:53
of cursor, which means it didn't ask me
00:11:55
for permission for running any command
00:11:57
lines, I can basically just go drink a
00:11:59
coffee and let it do the work. And as it
00:12:01
complete, it will mark this task as done
00:12:03
and move on to the next one by doing
00:12:05
taskmaster next. But now when you're in
00:12:08
Euro mode, uh I think cursor it will
00:12:10
limit to 25 coursees if you're using the
00:12:12
cloud 3.7, but you can actually put on
00:12:14
Gemini 2.5 Pro Max. This will allow you
00:12:17
to skip this limit and just do like 2002
00:12:19
coursees without any pause here. All
00:12:21
right. So, it's actually pretty crazy.
00:12:23
It just continue executing tasks one up
00:12:26
another and generate a whole bunch of
00:12:28
files, probably a few thousand lines of
00:12:30
code. And let's just run that first. Um,
00:12:32
so I'll do mpm run def. All right. So,
00:12:35
if I try to open this, it has a lobby
00:12:37
building and authentication building.
00:12:39
So, I can give a name, choose avatar,
00:12:41
start playing, and my user has been
00:12:42
created. I can set how much time we have
00:12:44
for drawing the difficulty and whether
00:12:47
it's public game or not. So it's pretty
00:12:48
good. Uh I can create but okay looks
00:12:50
like the actual game inside hasn't been
00:12:52
done yet but it's pretty impressive
00:12:54
about how much it is able to deliver in
00:12:57
just one go. And I can come back to do
00:12:59
task master list with subtasks. So you
00:13:02
can see that it done the uh first four
00:13:04
tasks but it haven't fully finished the
00:13:06
development game run logic into my UI
00:13:08
component. Maybe that's why the actual
00:13:10
game room is not showing up. So I can
00:13:12
ask it continue to implement but you saw
00:13:14
that I had this arrow here and then I
00:13:16
can prompt it now refract the errors you
00:13:18
made and creating new cursor rules to
00:13:20
making sure you don't make those
00:13:21
mistakes again. Now you can see that it
00:13:23
actually adding new cursor rules about
00:13:25
nextjs app router. Uh though for some
00:13:27
reason it didn't actually create content
00:13:29
which is bit weird and accept this one.
00:13:31
So for now I just copy paste manually. I
00:13:33
probably to update cursor rules to
00:13:34
making sure it will be saved properly.
00:13:36
But now I'm going to continue task. Now
00:13:38
let's check what is the next task to
00:13:41
complete to a point where we can do some
00:13:45
quick testing of the drawing. Okay,
00:13:47
after another 15 minutes of it just
00:13:49
doing the task by itself, I got this
00:13:51
game that's kind of fully functional
00:13:52
where I can give it a name, choose my
00:13:54
avatar, start playing and also create a
00:13:56
room called Jason's room. I can set a
00:13:59
timer about how long people can draw as
00:14:01
well as number of runs, the difficulties
00:14:03
and if I create a room other people will
00:14:04
be able to see the room I created as
00:14:06
well. Click on join. We will see
00:14:08
multiple players in the room. And if I
00:14:10
start this, I will have this canva where
00:14:12
people can start drawing to describe
00:14:13
what this uh word is. And on the top
00:14:16
right corner, there's also a timer
00:14:17
document how much time is left. And once
00:14:19
it finish, it will send both result to
00:14:22
chat GPT and GPT will look at the image
00:14:24
and give the description and evaluation,
00:14:26
pick up the winner to get points and
00:14:28
then move on to the next one. So it's
00:14:30
pretty amazing that it did a whole
00:14:31
multiplayers games like this by itself
00:14:34
in 20 minutes. So this how much
00:14:35
performance gain you can get by equip
00:14:37
cursor with the right task management
00:14:39
systems and what's really exciting is
00:14:41
that this is just beginning. I can
00:14:42
imagine those tools and system became
00:14:44
way better in a few months time. I also
00:14:46
interviewed the creator of taskmaster
00:14:48
project where he gave more detailed
00:14:49
breakdown about the best practice
00:14:51
workflow to fully unleash the power of
00:14:53
taskmaster and exciting things that they
00:14:55
are working on right now. If you're
00:14:56
interested, you can join the AI builder
00:14:58
club where I put a full interview and
00:15:00
workflow inside the community for you to
00:15:01
check out as well as bunch other
00:15:03
learnings and tips from industry experts
00:15:05
for both vibe coding and building
00:15:07
production ready AI agents. And you will
00:15:09
also have access to tools like 10x
00:15:11
coderdev where it will help you generate
00:15:13
bulletproof cursor PRD as well as
00:15:15
next.js boplay that already have all
00:15:17
payment backend database set up so you
00:15:19
can launch your SAS in just a weekend. I
00:15:21
post a link in the description below for
00:15:23
you to join if you're interested. I hope
00:15:24
you enjoyed this video. I continue
00:15:26
sharing new tips and workflows I
00:15:28
learned.
!["Brasil Football Club" - A história do futebol brasileiro [2014]](https://ruhnrawpgxrzdrgaohnf.supabase.co/storage/v1/object/public/images/video/mLtRmNtde8Y/thumbnail.jpg)
