How to Work with AI Coding Agents: Spec-Driven Development, Context and Loop Engineering, Workflows

0:00Hi everyone, welcome to this event. Uh

0:03today on the workshop, um I think I have

0:06a typo here. I need to correct it. So

0:08today on the workshop, we are going to

0:10talk about um um a few things. So I

0:14prepared a document. So I linked this

0:16document in the here in the description.

0:19So this is the document we are going to

0:21follow. Yes. Go to site. So uh this year

0:24I'm experimenting with uh releasing the

0:27course content as articles. So this is

0:30the article we're going to follow today.

0:32Um so please subscribe to um to my

0:35substack and I'll be posting more

0:38content uh from this course. So this is

0:41a course that is called AI dev tools

0:43zoom camp. So there's also a link there

0:45to the course itself.

0:48So in this course we are going to learn

0:51about how using AI for making developers

0:55more productive. So today we are going

0:57to record we're going to have the first

1:00workshop in the series. In this workshop

1:03it's called overview but I want to call

1:05it AI native development. So here I'm

1:08going to uh implement a tool end to end

1:11and I'm going to touch things um like um

1:15what is uh specdriven development,

1:16specificationdriven development, what is

1:18context engineering, what is loop

1:20engineering, what is graph engineering.

1:21So the last two I added because uh if

1:24you're on Twitter, you probably saw uh

1:26people talk about these things. So I

1:29want to cover these things too and

1:31explain what they actually are in this

1:32session today. Right? So that's uh

1:35roughly the plan. We will uh take some

1:38idea and we will implement this. Um but

1:41I want to know a little bit about you.

1:44So I want to know what is your

1:46experience with coding agents if you use

1:50one of the agents already. So please

1:52write it uh here in the live chat and

1:55also please write um

1:59um what we want to implement today

2:01because I need your ideas. Uh today we

2:04will pick one of these ideas and

2:06implement following this approach. Of

2:08course if you don't have some ideas I

2:10have a backup idea that I can use but it

2:12would be more interesting um if you if I

2:17use your input right. So if you suggest

2:18some ideas and then based on these ideas

2:20we actually build this together and uh

2:23please write this in the live chat and

2:26also I want to run a poll. I want to

2:29know what I should use today. So if I

2:31should use uh clo C code or if I should

2:36use codex I like both and the materials

2:39that I prepared today they will work for

2:43most coding agents but still I'm

2:45interested uh in knowing what I should

2:49use today called code or codex right so

2:53these are the two that I actively use so

2:57please let me know what you think and Um

3:01while you're doing this while you are

3:03replying uh what I want to do is to just

3:06briefly talk about the course. So I'll

3:08also you saw the link right? So this is

3:11um uh this substack article. It links to

3:15the course. So first of all uh please

3:17subscribe to this. I don't know why I

3:20didn't uh include the link. It actually

3:22should be uh

3:27here. Subscribe now. So I'll add a link

3:29right now. So you can do that. Uh but

3:32also there is uh yeah right now

3:36also there is a link to uh go back and

3:39there is a link to this AI dev tools

3:40zoom camp. So this is where the course

3:42will happen and right now we are only

3:46preparing for this course right. So we

3:48are preparing the materials we're

3:49recording. If you want to be a part of

3:51the course there is this link here we're

3:54registering. Um I will send it here too.

3:59So uh here in chat. So if you want to

4:02sign up for the course too um please do

4:05this. And this is the second time we run

4:08the course and um things move in this

4:13industry very quickly. So what was cool

4:16one year ago today is uh obsolete.

4:20And things I covered last year um not

4:23all of them. Some of them are still

4:25relevant but um I really want to update

4:27it. That's why we are doing this session

4:29today. That's why we have this stream.

4:31Um,

4:33and last year what I did in the first

4:36module uh of the course uh we called

4:40them introduction to VIP coding. I don't

4:44want to use this name anymore VIP

4:46coding. So what I want to do is I call

4:49it AI native developer workflow right?

4:52It's kind of more mouthful right? What

4:54does AI native mean? Developer workflow.

4:57But I don't want to call it VIP coding

5:00anymore. Even though the name kind of

5:02stuck. Uh but what I want to show you is

5:05the process, right? And this process is

5:08what kind of makes I don't know it's

5:10kind of buzz word, right? What makes you

5:13AI native if you use this process,

5:15right? So VIP coding is just um kind of

5:18shoot and forget, right? So you live

5:20only once kind of mentality, right?

5:22Right? So the agent you prompt an agent

5:24the agent is doing something and then

5:26it's good enough. So this is kind of by

5:28coding. So what we will have today

5:31instead of that is a process that you

5:32can follow to make sure that the code we

5:35create with a coding agent is actually

5:38solid. Right?

5:41Okay. So that's the plan and um as I

5:44said this is the article we can follow.

5:47What I also want to uh suggest is that

5:51uh even though the materials some of the

5:53materials are outdated I they are still

5:56relevant right so what we did last year

5:59in this introduction to w coding is I

6:02did um classification of tools into

6:05different categories

6:07um you can check this classification

6:09here I think in lessons tool map right I

6:13don't want to do this right now uh but I

6:16still want you to go through this

6:18document. So this is something I updated

6:20based on the uh previous year uh just to

6:24uh explain what things things are like

6:28there there are chat applications there

6:29are AI coding assistants uh there are

6:32what I call project bootstrappers I

6:34updated this classification a bit um so

6:36you can go uh through this yourself but

6:39I don't want to spend time on this so

6:41what I want to do today is um I want to

6:45understand uh so first of all let me see

6:47what you ordered. [snorts]

6:50So people want me to use cloud code.

6:52Okay. So I can use cloud code. Um but I

6:56still don't understand what you want me

6:59to implement. Right? So I don't see any

7:02ideas. So the ideas is like a simple

7:04project um that we can implement

7:06together today and I will walk you

7:09through uh we will work through this

7:11project together from simple ideas from

7:14the simple idea uh like how we turn this

7:18row idea into something concrete right

7:22and um then we follow the process that

7:25is outlined in this article right so

7:28this is what I want to do um so Um

7:32I don't see any ideas from there. So

7:33I'll give you some uh time to think what

7:35you want to implement. I also have some

7:38ideas. Um

7:41but since we're talking about uh using

7:44AI here. So what we can do? Why my

7:51am I still audible?

7:54Okay. I don't know what's happening. Ah

7:56okay it works. Um so since we're using

8:00AI here I can just ask Chad GPT what we

8:02can implement right and um I'll give you

8:06some time but but um JGPT is a tool that

8:10I use

8:12often very often [snorts] and this is

8:14always the first um

8:18the first step [snorts] in the process.

8:20So I see one suggestion three in a row

8:22game. So last year we had a game. It was

8:26a snake game. Uh so we implemented a

8:28snake game together. Three in a row

8:31could be um a simple enough thing that

8:34we um can implement. Two for weekly

8:40feedback

8:41for projects. Okay, I like this. So I

8:45don't want to implement a game because

8:46last year we did a game. As I said, it

8:48was a snake game. Um

8:51so I want to I want to make some sort of

8:54application like for example this tool

8:57or weekly feedback for projects. Uh we

9:01can see how to integrate my SQL database

9:03to Kubernetes using stateful sets. Uh

9:06mano uh so this is not what I want to

9:08cover here. I don't want to cover focus

9:11on technologies. I want to focus on

9:13ideas and then taking an idea and

9:15bringing this into like implementing

9:17this idea following the process. Right?

9:19So we are not focusing on technologies

9:21here and in fact uh of course it helps

9:23to select as we will go through the

9:26process it will help uh to use a tool

9:29and to use a technology a piece of

9:31technology that you're familiar with. Um

9:33but here we we will be product managers.

9:37We will be architects. Right? So we will

9:39not implement things ourselves. We will

9:41rely on agents for implementing this. So

9:44for us the more important thing is how

9:46the work is organized. How can we make

9:48sure that the output of the agent is

9:50correct? Right? So this is what we want

9:52to focus on. That's why technologies are

9:55secondary here. Um

9:58so let's create an app which will

10:00analyze articles from public resources

10:02and estimate them if they can trust this

10:04info or not. That's interesting but I

10:06think this can take some time. Uh like

10:09how can we estimate this? For example,

10:11meeting planner with movable box for the

10:14topic and realtime target tracker

10:16application tool for uh daily weekly

10:19feedback for personal goals like

10:21intelligent goal tracker. Okay, I like

10:24this. Build an AI company research agent

10:27input company domain use playright.

10:30Okay, Dominic, this is a nice idea but I

10:32guess uh for this session cuz if I'm

10:34going to use a playright, it's a bit

10:36more complicated. So I will uh take uh

10:39okay let me check a product that somehow

10:41fetches transactions and gives you a

10:43summary of your expenses in the last

10:46month for example. So this is a good

10:48application but since it requires

10:50integration with something else it will

10:52also take a bit more time for this

10:53session. Um all of that all these ideas

10:56are implementable right so everything

10:58that you mention could be applied in

11:00this uh into this framework. So we can

11:03use this framework to implement all all

11:05of these ideas. So what I will do now is

11:07I will select

11:10the tool for weekly feedback for

11:12projects. So this is a good um

11:17a good idea.

11:19Uh it's also good because it's vague,

11:21right? So if I just um if I just copy

11:25this and uh start the coding agent, it

11:29will produce something. But like we can

11:31just

11:33check what will actually happen. Right.

11:36Um, so let me I have multiple sessions.

11:41Let me actually

11:43create a new one. Um, so I'm I have a

11:47remote computer. So you don't have to

11:48have this set up. And I think we all

11:51agreed that we want to use cloud code,

11:52right? I don't know why so many so few

11:55people voted for Codex actually. Um I

11:58really like codex but since we are going

12:02to go with um clot code I'll use clot

12:05code. So um I have a remote environment

12:08where my agents are running. So this is

12:11the setup I have. You don't have to have

12:13the same setup like it doesn't matter

12:15how you set up your coding agent where

12:17it's running. Um doesn't really matter

12:19for today's session. Right? So for me

12:21it's just simpler to use it uh this way.

12:24So I will create a folder

12:27um how we will call it AI dev

12:33tools experiment.

12:36So then in this tool so uh in this uh in

12:39this folder what I want to do is um uh

12:45see what happens if I just onehoot it.

12:48One shoot means just give it a prompt

12:50and let it implement things. Okay. So

12:53this is the wrong one. Um, okay. So what

12:58I typically do is I have a timuk

13:00session. So this is a remote

13:01environment. So if my connection drops,

13:03I want to make sure that I run it in

13:06timuk. Um, so then I will uh create a

13:11session here, t-muk session, and I use

13:14clo. So the way I start clot is this clo

13:18permissions. So it's actually an alias

13:20uh that runs clo with the skip

13:22permission mode right. So but like you

13:25just this is the usual u

13:30this is the same as clo then

13:34roley skip permissions.

13:37So um if you're only getting started

13:39with um coding agents, I would not

13:43recommend to run this dangerously skip

13:45permissions like just run code in the

13:48simple way like that or codex. Um but at

13:52some point when you interact with this

13:54it keeps asking you for things like hey

13:56do you approve this do you approve that

13:58and it gets tiring right? So that's why

14:00I run this in the skip permissions mode.

14:03But for me also I run it on a remote

14:07machine that um if something happens to

14:10this machine it will not affect my

14:12computer. Right? So if you use something

14:14like code spaces or you can run the

14:16machine on EC2 or whatever you use. So

14:19then you can be

14:21safe running things there. But in

14:23general like even for local use uh

14:27dangerously skip permission is okay

14:29right um but at the beginning I

14:32recommend to run it with uh with the

14:35usual mode just to understand what it's

14:37asking uh and maybe it's okay for you

14:40okay so this is what um what I have so

14:43this is the session and I'll ask it

14:48implement tool for weekly feedback for

14:51projects right so I don't post anything

14:54here and uh oops. So actually I want to

14:59stop it. So I want to have a one shot

15:04directory. Sorry

15:07here

15:09one shot.

15:13So now I'll do this

15:16implement tool for weekly feedback for

15:18projects. Right. So I'm just curious

15:20what exactly will happen. Right. Um I

15:23see a few questions why Codex is your

15:25preference. Um cuz uh you get more uh

15:30for the same plan in Codex. So you get

15:33higher limits um usage limits. Then you

15:36also get limits limit resets quite

15:38often. Um and the models in Codex they

15:42are compatible comparable to um to cloud

15:47code. So, and there are some things that

15:49work better like one of the things we

15:51are going to talk about is loop

15:53engineering and one of the

15:56implementations of this loop engineering

15:58is the /go command. We will see it

15:59later. In my

16:02experience, it works way better in codex

16:05than in code. But for most of the cases

16:09they are compatible and also

16:13um we will do this in a tool agnostic

16:16way right so it will work what we do

16:18today will work with any coding agent so

16:20if you use codex it will also work so

16:22you can use codex you can use open code

16:24you can use whatever coding agent you

16:26prefer

16:29okay

16:34so it will implement something So you

16:36decided for uh for stack. So yeah I I'll

16:39just leave it alone. But then what I

16:42want to do in parallel is [snorts] I uh

16:44want to create specification. So here I

16:47am in this section right now

16:48specification before code. Um because um

16:51the problem with uh this one short

16:53implementations is when we give it

16:55little prompt the model uh the agent has

16:58to make a lot of assumptions right. So

17:00it does it didn't ask me anything right?

17:03So it didn't ask me what tool stack I

17:05want to use. What is this problem I want

17:07to implement? Like it didn't ask me any

17:10of this stuff. And this is a problem

17:12because it cannot read my mind, right?

17:15And perhaps I had something on my mind

17:17that

17:19um the agent will like I didn't express

17:23it properly in my prompt. So the agent

17:25will fill these gaps and the decisions

17:29it will make it will most they will most

17:31likely not be what I had in mind right

17:34so that's why we want to first build the

17:36specification so this is the first step

17:38in our process we want to make sure we

17:41really scope uh out the problem or how

17:45to say scope yeah so we define the scope

17:47we really understand what we want to

17:49build and uh there are two level of

17:52specifications um project level so what

17:55exactly this project is about and

17:57feature level so this is more like uh

17:59for each task so we are starting with

18:01the project level um specification and I

18:06really like using charg for that so you

18:07can use any uh AI assistant here and I

18:11usually use the browser you can of

18:12course uh go here and talk it's it's

18:15fine but um

18:18I prefer this way I also prefer uh doing

18:23this because usually I don't do this on

18:25my computer. I use my phone. So I just

18:28open JGBT on my phone and I start doing

18:31a brain dump. So I really like using

18:33dictation mode and this is what I will

18:35do.

18:37So I will say

18:40um I want to build this tool and I want

18:44to help me I want you to help me scope

18:48um to set the scope for this project. So

18:51I want to be very precise in what I want

18:53to build. So I want to brainstorm with

18:56you uh and understand how the uh tool

19:00should look like. So give me some

19:02options and ask me some questions.

19:06And so now what I want to do is I want

19:08to turn uh this wake prompt into

19:11something very specific. And here um I

19:15usually don't put like a lot of u the

19:19thinking mode is not like very high. So

19:21something like medium is okay. Um cuz

19:25like first of all it takes too much

19:27time. Like if I increase the thinking

19:28mode like let's say if I go to high

19:30extra high it's thinking too much. Um

19:33right? So I want to have something more

19:35interactive.

19:37Um, and then also like these uh thinking

19:41modes, they are uh very verbose. Like

19:44even this one to be honest is very

19:46verbose. So I wanted it to uh let me

19:49actually do this.

19:53Ask me one question at a time and keep

19:58your output short. Right? If I don't do

20:02this like I'll get a wall of text and

20:04then I'll have to read through this text

20:06and it's just difficult.

20:09Okay. So this is what I wanted right so

20:12I wanted to have a very simple thing. So

20:16who gives the weekly feedback product

20:18owner team members client stakeholders

20:20external testers?

20:22Um so the way I see I don't know maybe

20:25the person who decided on this idea can

20:28correct me uh but I think we best better

20:31if I just make some assumptions these

20:33assumptions might not align with what

20:35you had in mind but still we will be

20:38moving towards something that we decide

20:40not the agent decides right um

20:44uh I want this to all team members not

20:47just product own project owners uh I

20:50want all team members um to be able to

20:54contribute it to to this uh contribute

20:58the feedback and uh then perhaps we can

21:02have something like a retrospective

21:04uh where we can discuss this uh

21:06feedback.

21:10So I I just made an assumption that we

21:12as a team have these processes and then

21:14once per uh I don't know some time once

21:17per month we have this retrospective

21:19where we talk about things that go wrong

21:22things that work things that don't work

21:25things like that right so what should

21:27each team member submit every week

21:30um

21:33yeah start stop continuous looks like a

21:35nice uh framework so we want to talk

21:38about things that work, things that

21:40don't work. Um,

21:43yeah. So, I think this should be uh the

21:46first should work fine.

21:59Should feedback be anonymous?

22:01Um,

22:04yeah. Uh let's uh let each contributor

22:07decide. So they can um submit things

22:11anonymously. Um but also they can choose

22:14reveal their name by default. Let's make

22:17it um attributed but then they can tick

22:22a box uh submit anonymously.

22:26I think it makes sense, right? So if you

22:28want to give some critique maybe you

22:30don't feel comfortable

22:32talking about this um in person I I

22:36don't know like it depends but let's

22:39just go with this right so it might be

22:41awkward a bit uh and then uh during the

22:43actual retrospective uh probably

22:45everyone will understand who that person

22:47is who gave feedback um but maybe for

22:51some things um like it's okay to uh to

22:54have this possibility

22:57Um, next question. When should teammates

22:59see submitted feedback?

23:03Okay, they only see the feedback they

23:05submit uh but the project uh

23:10okay I actually need to think about this

23:12um immediately

23:15when the collection closes only [snorts]

23:17during the perspective.

23:21Yeah, let's uh go with the option number

23:23four. So when I submit I only see uh

23:26only see what I submitted. I don't see

23:28the submissions on of others.

23:38Okay. Uh Yogen uh I don't know how if I

23:42pronounce your name correctly. Um all

23:45team members right. Okay. So I hope I'm

23:47moving in the direction you uh wanted.

23:50Who can reveal and facilitate the

23:52feedback?

23:56uh that's the

23:58facilitator

24:00number two which probably is a project

24:02owner.

24:04So I I guess you can see the idea right.

24:06So we want to be very precise in what we

24:10want to build but more like from the

24:12user point of view like we are not

24:14talking about technologies here. We are

24:16not talking about implementation detail

24:18but like more what exactly the tool

24:21should do not how

24:24um how feedback should appear.

24:29So uh all feedback should appear at the

24:32same time and then during the meeting um

24:36yeah we cluster there should be some

24:37automatic clustering but also during the

24:39meeting uh as the first step we cluster

24:41them um

24:44and then um once the cluster them also

24:49each team member can vote for the things

24:51that we can talk about. So then uh at

24:54the end we have a prioritized list of

24:57things we should discuss in the

24:58retrospective.

25:15Okay. Um next question. How many votes

25:17should each person get? Um

25:21three is fine.

25:26Uh should people be able to place uh

25:29yes like if somebody really wants to

25:32talk about something uh at least in

25:34retrospectives in companies where I work

25:37usually uh you could vote for the same

25:41topic multiple times. What should happen

25:43after the discussion?

25:48Yeah, [snorts]

25:49let's just capture both actions and

25:51decisions.

25:52uh we can also do like I I don't think

25:57we can actually like what should happen

25:59after the discussion capture um

26:03ideally if we record this meeting right

26:06so then uh we can use AI to um

26:11automatically infer this from the

26:12transcription

26:14uh so let's uh capture actions and

26:16decisions and also we want to record the

26:19entire meeting so then uh at the end we

26:22can upload the can get the record

26:26uh transcribe it and then based on the

26:28transcription we can capture all the

26:30things and uh all the actions and

26:32decisions.

26:47Um

26:50so now let's go with a simple version.

26:53We just upload video or we just upload

26:56audio or transcript directly right and

26:59then it can process it and then uh as

27:02the then we can add other things on top

27:05of that when we need. So let's not let's

27:08not do um built-in recording yet. I just

27:12don't want to be very ambitious here.

27:14Right. uh at the end we can build

27:17something like that right but um we just

27:21you know go with the simplest version

27:24who confirms the extracted solution

27:26decisions

27:30um I I don't know like I'm tired of

27:33making decisions um okay I'm to be

27:35honest I'm tired of making decisions I

27:37think this should be enough for the MVP

27:40um for the rest of the important things.

27:44Uh, I want you to make some decisions

27:46and explain me why you chose this

27:49decision, why you decided to go with

27:51this option and what were the other

27:53options you considered.

27:58Like at some point like we can reply to

28:01these questions all day long, but it's

28:02been already half an hour and we are

28:05only getting started, right? Um,

28:11okay. Weekly team feedback tool. MVP

28:13scope. Um

28:16create feedback cycle. Collect feedback.

28:19Uh reveal and cluster feedback. Vote on

28:22discussion topics. Run the discussion.

28:26Discuss. Keep deferred. Process the

28:28meeting recorded. System generates a

28:31transcript. Okay.

28:34Decisions made for the MPP instructed

28:36results require facilitated approval.

28:39Okay. So then somebody goes through

28:41these things after this

28:44feedback is submitted as a separate card

28:46team can create

28:50okay anonymous feedback stays anonymous

28:53yeah voting is visible after voting

28:55closes yes I think it's good that people

28:58don't see you where other people are

29:00voting in reality when you have a

29:02pipboard you cannot not see right but

29:05here maybe it's a good decision

29:08automatic clustering is always editable.

29:11Okay, good. Action items have a simple

29:13structure.

29:15Uh description, owner, optional due

29:17date, action items. Yeah, as a result of

29:20the meeting, we want to have action

29:22items. That's cool. One retrospective

29:24belongs to one project. Okay. Main

29:27screens, project page, current feedback

29:29cycle, perspective,

29:32feedback form, retrospective board,

29:34media, upload page, retrospective

29:36summary, roles, team members,

29:39facilitator.

29:42Okay. Then things that we explicitly

29:43exclude from MVP. I think this is good.

29:46Like we decide what is in the scope and

29:48what is out of the scope. Like with MVP,

29:50we want to have a very focused thing

29:53such as success metrics, MVP definition.

29:56Okay. So now

29:59you can probably interact more with this

30:01but I think this is enough for to get

30:03started. Now I write save everything to

30:06a markdown file that I can

30:11download. Right. So this is how I always

30:15uh finish these brainstorming sessions.

30:18So what is doing is writing some Python

30:19code or creating producing this file

30:22doesn't really matter. What matters at

30:24the end is I will have a markdown

30:26document that I can download. Um, in the

30:29meantime, let me see what this thing

30:32did,

30:34right? Uh, is there a dependency Python

30:37CLI for tracking weekly project health?

30:40Uh, okay. So, it did something

30:43completely different, right? So, it did

30:45a CLI for tracking weekly project

30:48health. Okay.

30:51Um so project at API name building API

30:54onexi.

30:57Okay I small command line register the

31:00project you care about local short entry

31:02per project per week and get a digest

31:05you can paste into weekly update.

31:08I mean it could be useful right um but

31:11this is not what we wanted at least uh

31:14you can see my point right. So this is a

31:16very different thing at the end very

31:20very different right so you just made a

31:22lot of assumptions and it just went with

31:23these assumptions I didn't stop it I

31:26didn't ask it to uh I didn't correct it

31:28so then the result is a working app but

31:31this app

31:33is absolutely not what we need right so

31:36I'm going to stop this um and I'll

31:40create another directory which I'll call

31:45um

31:46project heatback. So this is will be

31:49actually our directory that we are going

31:52to use and I'm going to open it in

31:55visual studio code. You can use whatever

31:58like if you use cursor you can open it

32:00in cursor. If you use oops if you use um

32:05yeah whatever you use you can use it. So

32:08now I will open it

32:12as a project and right now there's

32:14nothing right. So I want the first thing

32:16I want to create will be an empty folder

32:19called docs

32:22and what I will put in this document is

32:25uh in this folder is uh this thing.

32:30Okay. Can I download it?

32:34I can also read it. uh but we already

32:37read it but I'll I'll read it not here.

32:40I'll read it in uh

32:44here.

32:46So um

32:49I should be able to just drop here and

32:52I'll call it what?

32:57Yes, I'll call it plan.

33:02Okay. So let me commit it.

33:06I think I'll just use uh here the

33:09terminal get status. Uh so this is not a

33:14repository yet. So I do get in it and I

33:17do get at g commit u

33:22md right. So this is our document. So

33:25you you should commit regularly.

33:32So the reason I add underscore here is

33:34because we can have some other do some

33:36other folders in our uh in our project

33:41right and then I want the important kind

33:43of uh uh important folders to be first

33:46in this list so then I can see them

33:48immediately.

33:50Can you drop down here the MD file? I

33:54think I can. So I what I will do is um

33:59I'll ask clot

34:02um

34:04upload the plan MD file to G.

34:10So, and I'll actually I wanted to edit

34:13this a bit, but maybe we edit this and I

34:16also uploaded.

34:18I could have also actually committed

34:20this to GitHub already and share the

34:22link with you. Um maybe it would have

34:25been better, but it already created the

34:28G. So I will just give it to you.

34:34But if you're watching this in the

34:35recording, there probably be there

34:38probably will be some folder that I will

34:41put inside our AI dev tools zoom camp.

34:46So here in overview I will probably

34:48rename this. It will be not overview but

34:51something else. But inside we will have

34:53code and then uh I will put the code

34:56that cloud code created when we didn't

34:58have any specification just for you to

35:01check it and I will also put all the

35:03other artifacts we produced

35:06okay but this is this is it right so and

35:10this is very similar to what we saw

35:12right

35:14this is actually exactly the same

35:16document that we already reviewed so

35:18here there's nothing to review uh

35:20already Yeah, because we already

35:22reviewed so there is nothing left to

35:24review.

35:25Okay, so um this is what we did starting

35:28a chat chat assistant. Um

35:31now we need to initialize the project.

35:33So I already we already bootstrapped it.

35:36Um

35:37and then I want to

35:42do this. So we didn't agree. So I'll

35:44start a new session. So we did not agree

35:46on the technology, right? Uh so we did

35:49not discuss the tech stack. This is

35:52something we could have done um we could

35:55have done um during the brainstorming

35:59session with the chat application. Um

36:02but usually I already so I use these

36:05chat applications for conceptual

36:06thinking like what exactly I want right

36:09and then um you don't have to use you

36:12can use any you can also use a coding

36:14assistant for that but for me it's kind

36:16of helps with separation of concerns

36:18kind of because coding agent cannot I'm

36:20sorry AI charge cannot really touch any

36:24files right so it kind of forces it to

36:28stay on the conceptual level rather than

36:30create code so it will not create code

36:32unless you ask

36:35okay now I'm copying this read plan

36:37propose multiple options for the text t

36:38and explain each option don't write the

36:40code yet so we want to just understand

36:43what we will implement

37:00So now it's looking for um this file it

37:04found to uh plan.

37:08So now it will read it and it will

37:12actually say okay now I understand what

37:14you want. So uh before the options

37:16here's what actually drives the choice

37:18or hard parts

37:21retrospective word

37:23media pipeline uh audio video uploads

37:27transcription and the

37:30so for video for media pipeline I think

37:33what we can do is so for example I have

37:37u an Android phone on this Android phone

37:39I have a recorder so this is just usual

37:41Google recorder

37:44And what it can do is I can just use my

37:47phone to record the conversation. So

37:50let's say we are um talking we're

37:54sitting in the room. So we have a TV

37:56screen uh and then everyone is talking

37:58about these things during the

37:59retrospective. Let's think let's say

38:01this is not a remote one. So then at the

38:04end my Google um will create a

38:07transcript my Google recorder right? So

38:09then I don't need to do this media

38:11pipeline. Maybe it will just help uh

38:13make things easier. Um although if I

38:16think about this like just sending this

38:18thing to um

38:20uh to whisper shouldn't be too

38:22difficult, right? So maybe I'll just

38:24leave it like that. Uh transcription

38:26that takes minutes needs direct to

38:28object storage upload plus.

38:32Okay, whatever. I guess it works. Uh

38:35lamb steps clustering cards extraction

38:37decision a sync provider agnostic small

38:40amount of code anonymity. Okay. Option

38:43nextj

38:45full stack plus postgress.

38:50Okay. Um

38:53junga

38:54I like junga to be honest like um I

38:57think this would be a better option if

38:59we wanted to use like something like uh

39:01bersel for hosting. It would be really

39:04good option because in versal you get

39:06out of the box you get many things out

39:08of the box. H yeah it even says versal

39:10serverless model can't ah so yeah it

39:14even suggests that deploys this is good

39:17but uh for me I can read JavaScript I

39:20can read Typescript but I'm not uh I'm

39:23still like I really have to focus for me

39:26when I open the open Python code for me

39:29it's way easier to to read the code.

39:31We're not going to read any code today.

39:33That's why like you can just choose

39:35whatever is more comfortable for you. Um

39:40then I like like jungle more than uh

39:42fast API.

39:45Um because I know it better, right? I

39:48know it better than uh fast API. So I

39:50would just go with the technology that I

39:54know. So I don't even live life elixir.

39:58Okay, that's uh that's very interesting.

40:03Um

40:07kind of exotic, right? Why did it uh

40:10suggest this thing? I don't know.

40:14And it says time to MVP with jungo is

40:16fastest. Okay, like I was kind of

40:19leaning to this, but now uh since we

40:22want to move fastest, we don't have a

40:24lot of time on this section. Um, and it

40:26says recommendation option B. Um, yeah,

40:30but I would still make my own suggest uh

40:33option, right? So, if you're like, okay,

40:35I don't really care. I ask clo or I ask

40:38Codex based on what you think what is

40:41the best suggestion or you can outline

40:43some um some ideas like where you want

40:47to host it and things like that, right?

40:48And then it will help you to to decide.

40:50For me, uh let's go with Jungo. I like

40:54Jungo. I have many websites uh that are

40:57implemented in Django. So for example uh

40:59for our courses we use this platform. So

41:02this is Jungo. Uh then I also run uh

41:05this community AI shipping labs. It's

41:07also Jungo. So I have experience with

41:10Jungo. So for me this is [snorts] a

41:12natural choice because I already uh is

41:15familiar with this. So I see a comment

41:18fast API plus React. So if Alexi uh

41:21wants to use that um use that right. So

41:24here we're not really um it's not really

41:28about technologies, right? So any of

41:29these options will work. I wouldn't go

41:32with this Phoenix though.

41:35Um

41:37because it's very exotic.

41:41I don't know why why Alex here.

41:44Like I would go with Rust or go if I

41:48really wanted to do something like

41:50unusual.

41:53Okay, let me lay lay out the concrete

41:56architecture. I will write it to doc

41:57architecture even though I didn't tell

41:59it uh to write it into file. It made the

42:02right call. Uh I would eventually ask it

42:05like after we discuss all the things I

42:07would say at the end of the session, hey

42:10like let's now document everything. I

42:12didn't need to do this. It just made

42:13this decision itself.

42:17Okay. So it's taking some time to do

42:20this. Let me see what we have. So, so,

42:23so far we don't really have anything,

42:24right? So, LA is the only um document we

42:28have. I'll also tell it to

42:32commit after we finish.

42:36Commit after you finish.

42:43So, language Python 3.12.

42:48I would go with Python 3.13

42:51but I don't think it matters posgress I

42:55think right now it's version 17

43:01let's go with Python um with last Python

43:07and posgress

43:09because I think last one is this 14 that

43:14junk templates alpine

43:17js I have no idea what is But

43:20I will like when it comes to um

43:25front end technologies I'll just trust

43:26agent to pick up whatever want

43:30uh authentication email uh plus invite

43:32links. So I want to a bit the scope it

43:35so I don't want to um go like full um

43:41like I don't think we need these things

43:43right now.

43:46Let's make out very simple.

43:50No emails here.

43:54And then background job plus radius. Um

43:59this is too heavy.

44:02Uh file storage.

44:06File storage.

44:09We don't need file storage.

44:13Process recording. and throw

44:17it away.

44:20Uh, deep gram API. So, I have no idea

44:23what's that. Let's use whisper from open

44:28AI

44:30llm of course because I use clo cloud

44:33code it suggested uh this you use

44:37open ai too

44:42deploy um

44:46let's remove that part and only include

44:51docker compose right so I want to disco

44:54cop it a little bit so it's more

44:55implementable.

44:57Um

45:02in the text stack we talk about

45:04trade-offs too. Yes. Um and this is what

45:07I'm you you still need to like I I think

45:10of myself and uh this is what I want you

45:13to also do of I think of myself as a PM

45:17a product manager plus architect. Right?

45:20So product manager is thinking about the

45:22think from the user point of view and

45:24architect is thinking about technologies

45:27from high level right so you don't

45:29really go and implement all the single

45:31all these things by hand but it helps to

45:35um

45:38to be aware of the technologies and

45:40trade-offs you're making right and let's

45:42say you don't have this experience right

45:44uh that I do so I already for me this is

45:47not the first time I do this kind of

45:50things, right? So for me, I already know

45:52what I want to do like what kind of

45:53technologies I want to use. You might

45:54not have this background. You might not

45:56have this knowledge. So then what I

45:58would suggest is to challenge every

46:02single line of this decision file here

46:05and ask, hey, do you think this is good?

46:07Are there

46:09um newer versions? Because we know that

46:11LLMs have this knowledge cutff. So they

46:14suggest so for example it suggest to use

46:16GPT40

46:17while we have P4 P46 right right now. So

46:22um I would actually challenge every line

46:26here and ask hey do you think this is a

46:28good idea like

46:30is it complicated or not? Can there be

46:32simpler versions? Right? So I would

46:34really ask I would really challenge here

46:37every single line.

46:41Um okay so what I also want to do is um

46:46I like using dictation mode here too. So

46:49this is what I do. So I'm on Windows

46:51that's why if I press window key plus H

46:55I have this thing right this this is a

46:58transcription. So um I want you for

47:01every single line of your technology

47:03choices I want you to see if there are

47:06newer versions available for that

47:14file transcription is unreoverable the

47:16media is already gone or whatever like

47:18we don't worry about this things um we

47:20will deal with them later as we um

47:27as we make it let's say more mature for

47:29MVP key. We don't need most of these

47:30things.

47:34January 2026. That's interesting because

47:37for O was released like 2 years ago. So

47:40I think it's kind of lying.

47:42Or maybe it just didn't want me to use

47:44it like a proper OpenAI model because

47:46it's an anthropic uh model.

47:50Okay, I'll let it do it, but then I want

47:52to continue.

47:55Uh so what do we have next? So next now

47:58we have the text tag and now with this

48:03now with this text stack and uh

48:07with the plan I want to create a backlog

48:10of tasks

48:12right so what I want to have is I want

48:14to take like everything that we did here

48:17plan architecture all these things and I

48:20want to take them and put them into uh a

48:24list of tasks such that each task is um

48:30what do I say here? Each task is

48:32independent, small enough to finish in

48:34one session. So it shouldn't be too big,

48:36shouldn't be too small. Um so I want to

48:38have like a very clear decomposition

48:43and I think after this session I'll need

48:44to update this article a little bit,

48:46right? Because like we have uh selecting

48:50technologies like I spent a lot more

48:52time on this than just this prompt.

48:55Okay. So, uh now I will probably do this

49:00continue doing this in the same um

49:04in the same session. I think it's busy

49:07with checking the technologies. I think

49:09it's okay.

49:13Um so I don't want it to like go crazy

49:16with like all these choices.

49:19So uh now uh I continue doing it in the

49:22same chat because it already has some

49:24context about um what we want to do. Um

49:28so I'll just continue. Uh alternatively

49:30you can start a new chat. It's also not

49:31a problem.

49:34Um but first please commit what you

49:39have.

49:43Uh it's is it time to update the

49:46application spec? Uh uh Alexa asks I

49:50don't think it is right. So the from the

49:52user point of view um

49:56it didn't change much right. Um but

49:59maybe yes some things like some things

50:02we did uh here we decided here they do

50:05influence the decisions we made.

50:10And I see that it didn't update. Uh I

50:13also see that you didn't update

50:17our architecture MD. Please do this

50:21before you commit.

50:24So it could be like uh when architecture

50:27things that we decide in our

50:29architecture they can influence our

50:32original plan our original spec right so

50:35then of course we need to update it. Um

50:37I don't know if um anything we did

50:41actually uh

50:44all the decisions we made here

50:45influenced this. Um so for now I'll not

50:49uh

50:51we can actually combine these files too

50:53right we can take the architecture plan

50:56put them together

51:01um okay it's

51:03producing a backlog of tasks okay it's

51:07fine

51:12so uh let's see tasks

51:19Project skeleton with a passing test.

51:22Um, okay. Docker compos environment.

51:26Okay. Base layout and front end assets.

51:29Okay. Authentication without email. Uh,

51:32project membership join. Ah, I think

51:34Alexi what you meant is like this

51:36decision that we made about

51:37authentication without email, it does

51:39influence uh the original spec. That is

51:43correct. Right. Um so it does help to um

51:48keep this in sync but for me this plan

51:50and task and architecture. Um so they at

51:54some point they will um kind of say

51:58disynchronize with the actual content.

52:00So I mostly use it for seeding the tasks

52:03these tasks

52:05uh and uh to be honest I don't really go

52:09back to this file often. So uh in the in

52:13the article I included

52:16did I include it? Yeah, this um

52:21SQLite search um like the the way I

52:24approached it

52:26and um

52:29there I followed the same process right

52:31so I created uh I talked to JPT then I

52:34defined what I want to build and then uh

52:38I think yeah I have plan here in the

52:40root and actually this thing kind of

52:45disynchronized

52:47from the codebase So I just used it as a

52:51starting point and then like I can just

52:54actually delete it right so it's not

52:56really needed anymore because many many

52:58things happened since this file was

53:00created so like what we can do is we can

53:02treat this files right now as just

53:04seeding

53:05uh point kind of like we will create

53:09tasks from this and then um

53:14yeah we will not need to actually

53:18update this file all the time, right?

53:20So, at least I don't do this. Sometimes

53:22it can help uh but for me, we have this

53:25uh these tasks and this is what matters

53:27right now because we can take these

53:29tasks and actually uh

53:34and actually start um yeah working.

53:39So, we have the tasks. This looks

53:42reasonable. Authentication without

53:44email. Project membership and join

53:45leaks.

53:47Uh user can create a project and invite

53:49others. Permission.

53:52Uh feedback cycle,

53:56feedback cards.

53:58I think it looks okay. Um like each of

54:00them is a concrete feature with concrete

54:04goal and concrete description. So it

54:06doesn't seem too small. I'm not sure

54:08about this. um some things like

54:11permission predicates

54:13uh but I I will just go with this right

54:16so to me like here you need to wear your

54:19kind of product management manager hat

54:21and understand uh

54:24like is this task good enough like is it

54:27not too small not too big and you think

54:29more about like from this from the

54:31product perspective not from technical

54:33perspective

54:35um okay so we have 28

54:41uh things here.

54:45Okay. So now what we want to do is we

54:47want to put these things into GitHub,

54:51right? So they are here but this is not

54:53really useful. So the the reason I asked

54:56it to put this into the um to the

55:00markdown document is because I wanted to

55:02review it, right? But once these things

55:03are done, now I want to put this to

55:05GitHub.

55:07Um, so I'll just do clear. I'll start a

55:10new session. And in this session, I want

55:12to so I'll just dictate. So I want to

55:15publish this project on GitHub. Uh, but

55:18I don't like the name. Can you please

55:19help me select the best name for our

55:23tool that we're building? You can check

55:25plan.md

55:26uh to see what exactly we're building.

55:28So um yeah, I want to have a nice name

55:32that reflects u the purpose of the

55:35project.

55:42Okay, let's see what I suggest and then

55:44we will uh create um

55:48I also want to check if

55:52yeah get lo. So we

55:57we already have uh some things here.

56:04Okay.

56:09So, what is it doing exactly?

56:12Weekly loop. So, why why it was doing

56:14this

56:19retrol loop retro cycle?

56:24I don't really need an organization.

56:27I don't need an organization. And I just

56:29want to create it in my personal um

56:33space, right? So just um suggest a name.

56:37Uh retrol loop is okay. Uh so let's do

56:41this.

56:45I have a skill for creating GitHub repo.

56:48Like if you don't have this skill, like

56:50you can just ignore that it exists. Um

56:54but yeah it your agent will do the same

56:57thing without the skill. So for me it's

56:59just I have a flow for creating um

57:02projects but yeah just pretend sorry

57:05pretend you did not see the skill.

57:13So create

57:16make it public.

57:29Okay. So now we have this repo. I'll

57:32share it with you.

57:37So for now we don't really have much

57:39here.

57:42noted it mean no license whatever no I

57:44don't want to do these things uh

57:47actually it can do this by itself even

57:50if you don't ask it so at least here it

57:53it asks

57:55okay so what do we do next we create a

57:57GitHub issue for each task

58:03tasks

58:05MD

58:08so I quite often um start a new session

58:13because um so first of all you're kind

58:17of you're spending tokens right so every

58:19time you continue a session um you have

58:23a new task but you continue a session

58:25from the previous task um not only like

58:28there's already some context so it can

58:31um make your model

58:36um can confuse your model um but also

58:39you're spending tokens right So like for

58:42each new task I try to start a session a

58:45new session.

58:50So now it's going to create issues. So

58:53it's creating a script. Um and it's

58:57running the script. So I think if I go

59:00now to

59:03this retrol loop

59:06I see this issues here.

59:12I don't know why I decided to add the t

59:14the the number here. Um whatever. Like I

59:19would like on the real project I would

59:21ask it to uh remove the the numbers

59:24because makes no sense. We already have

59:26a number here.

59:29Okay.

59:31Um

59:36clear

59:40so uh implement task number one. So now

59:44we want to bootstrap the project. So we

59:47want to actually um create uh what where

59:51is it? Let me close all these things.

59:58So we want to implement um the things

1:00:06okay.

1:00:08So

1:00:10this is what I do. So we create the

1:00:12first task. So this uh bootstrapping

1:00:14actually took more time than I expected

1:00:18but this is time um so this is time well

1:00:22spent. Uh so I think I will take more

1:00:25time than I initially planned. So I

1:00:26planned initially to do it for 90

1:00:28minutes. I think it will be more like

1:00:30towards two hours.

1:00:35Okay. So it's working on this thing

1:00:38right now and while it's doing this um

1:00:45so I want to move to context

1:00:46engineering. So context engineering is

1:00:50um

1:00:52so first I want to start with prompt

1:00:53engineering. So prompt engineering is

1:00:56the prompts you write here implement

1:00:58task number one right and then uh when

1:01:00you do this when you start a new session

1:01:02the agent needs to every single time the

1:01:05agent needs to figure out what exactly I

1:01:07want from it like what is the task

1:01:08number one where it is uh what do we use

1:01:12uh do we use GitHub for this so it needs

1:01:14to um understand what exactly is

1:01:17happening right so in this case um well

1:01:21I don't know exactly where it took Task

1:01:24number one. So I think it found it.

1:01:30Yeah, I think it found it this uh in

1:01:33task. So it didn't even go to um

1:01:38it did not even go to um our uh tracker,

1:01:42right? Um so agent is making assumptions

1:01:46here. um what I wanted it to do to

1:01:48actually go and um

1:01:51take this issue from um GitHub. Right?

1:01:55Then I see uh some other things.

1:01:59So um

1:02:03where let me check what it's doing.

1:02:08So it was checking versions.

1:02:10Uh then it used Python 3.14

1:02:16uh EVP index. Okay.

1:02:26Yeah. Here everything is fine. But um

1:02:28usually so when we start a new session

1:02:31we want the agent to understand what is

1:02:34happening and what it needs to do

1:02:36because if we just give this simple

1:02:37prompt this is not enough. Right? So

1:02:40then our prompt should be more explicit.

1:02:44I should have said implement task number

1:02:46one which is a GitHub issue. Um and then

1:02:50maybe add more things. So for the agent

1:02:52not to think about these things all the

1:02:54time, we need to do what we call context

1:02:58engineering, right? So we need to give

1:02:59the agent the right context. And uh most

1:03:02of the time context engineering in case

1:03:04of coding agents is about creating this

1:03:07file called uh agents.mmd. So I'm going

1:03:10to create it here

1:03:14and write in things that are important

1:03:16for the agent um in this file. So this

1:03:19is just a simple markdown file where we

1:03:21describe all the things that are

1:03:23important for the agent like what are

1:03:25the um

1:03:28what are the tools that we use. I think

1:03:29I have an example here.

1:03:32All right. So

1:03:34comments uh tooling rules uh constraints

1:03:38pointers to documents and things like

1:03:40that. So here I have an example. You see

1:03:43in this example I don't use any uh

1:03:45markup. So I don't use uh any u like

1:03:49headers. I don't use bold formatting cuz

1:03:51this is not really needed. So um what I

1:03:54will do now is I'll take this example

1:03:59and I will ask uh is it done? No I think

1:04:03it's still doing. So after it's uh it

1:04:06has finished so I will ask it to create

1:04:09agents uh MD file with the this example

1:04:16from this example

1:04:18um but um since here I use clot I don't

1:04:22use codex so codex would go and read

1:04:24agents but clot does not clot needs a

1:04:27file called clot md and I use both I use

1:04:31both codex I use clot code I use other

1:04:33agents

1:04:34Um so only for CL we need to have like a

1:04:36separate file. I want to make it

1:04:38possible for me to use any coding

1:04:41agents. So let's say I'm using clot and

1:04:44I run out of limits. So then I can go to

1:04:46codex and continue what I'm working on.

1:04:48Right? So I don't want to really u go

1:04:50between different agents. So I want my

1:04:52setup to be uh tool agnostic. That's why

1:04:55what I do in cloud is I have this line.

1:04:58So this is the single line I have in my

1:05:01cloud code. It says, "Hey, go read

1:05:03agents.mmd." Right? So then it goes and

1:05:05reads it.

1:05:09Okay. Um

1:05:12test exists version. Okay.

1:05:16Commit and then create agent um and then

1:05:20create agents MD with uh content similar

1:05:25to this.

1:05:29And I'm doing it in this session because

1:05:31it already has some context about what

1:05:33we um want to do, what kind of

1:05:36technologies we use, how to run some

1:05:37things, right? So the next agent does

1:05:39not need to rediscover these things. It

1:05:41will not need to rediscover. Okay, this

1:05:44is jungle project. Okay, we use UV for

1:05:46testing or for for dependency

1:05:48management. Uh so it will not need to

1:05:50rediscover this. um it will just get it

1:05:54from agents.m MD.

1:05:59So it flaged some things but I kind of

1:06:01need to move faster that's why I am

1:06:04ignoring this. You shouldn't you should

1:06:06actually read and see what it wants and

1:06:09then like if it's not clear you just ask

1:06:12it hey what do you mean here like what

1:06:13do you want for me like what kind of

1:06:15decision you want and sometimes uh in

1:06:18many cases uh you don't really need to

1:06:21decide you can ask hey what are the

1:06:23possible options

1:06:25and then you ask it what is the best one

1:06:27and then you just say okay let's go with

1:06:29this

1:06:33and Um from what I see it added so this

1:06:38one is not really needed

1:06:41right uh jungo app for weekly stop start

1:06:44continue cycles and resp perspective the

1:06:46whole

1:06:49so this one we don't need

1:06:53then I don't like uh this extra

1:06:57um extra markup because it will cost us

1:07:00tokens

1:07:03UV sync [snorts] uh run server migrate

1:07:06pi test rough check okay

1:07:12um I would actually also create a make

1:07:13file but this would be separate thing

1:07:16all uh authorization leaves here uh

1:07:24okay it kind of becomes big I don't know

1:07:27like do we really need uh

1:07:30do we really I will just dictate

1:07:35uh do we really need all these rules? So

1:07:36the idea behind this file is that every

1:07:39agent session that starts in this uh

1:07:42repository needs this information. Do

1:07:45you really think that all the agent

1:07:47sessions will need that? Keep only the

1:07:48most important ones

1:07:53and then also like I'm not sure how

1:07:56important is this thing, right? uh if

1:07:58agent needs to know what is uh what

1:08:00we're talking about it can just go here

1:08:02and I think it's just duplicated this

1:08:06and I'll remove this from here too I

1:08:10think uh our readme for readmi I have a

1:08:13several separate process how what should

1:08:16go in readmi and it should be different

1:08:18from agents and actually uh since we are

1:08:21already on my substack I have an article

1:08:24about that

1:08:26Um

1:08:30let me go to archive. How to write a

1:08:33good readme. Right. So this is a

1:08:35separate thing. Um I would like read me

1:08:39and agents MD. They are very different

1:08:41files.

1:08:44Okay. So now it uh

1:08:50so then agents sometimes they write

1:08:53things that do not exist. So they say

1:08:56okay we only have posgress we don't have

1:08:58this but what is the point of writing

1:09:00what we don't have right so then I

1:09:02usually trim this

1:09:09okay so this is our uh agents MD so next

1:09:12time when we create a session uh it will

1:09:15know um

1:09:17it will get this information uh but the

1:09:20important part is we also need to

1:09:24um to say that things are in uh in in

1:09:29these issues, right? By the way, I think

1:09:30we should close this one.

1:09:33So, I'll just close it.

1:09:35Um

1:09:37so, um we need to say this is the

1:09:40process we follow. So, for this I

1:09:42usually create a document called

1:09:44process.md.

1:09:46So, I will put it here

1:09:49process.md where I describe the process.

1:09:51So, right now it's very simple. So I

1:09:53just say uh tasks are in GitHub uh read

1:09:58the acceptance criteria. This is

1:09:59something that we will add later. Um,

1:10:02and commit regularly, right? So maybe I

1:10:04don't even need this right now. Um, so

1:10:08this file will grow but I want to see

1:10:11the file because this is the process.

1:10:13This is how we work. So this is

1:10:14something that agents if we want to

1:10:17implement something or do something the

1:10:19agents will read this file and I refer I

1:10:22don't we can put them here in agents.mmd

1:10:25but because I know that uh this file

1:10:27grows

1:10:29uh and then for these files as I said uh

1:10:32they are more like um things that I use

1:10:35for seeding the tasks right so for me

1:10:38they are more like one of things that I

1:10:41that will became stale that I will want

1:10:43to eventually remove. So there is no

1:10:45point for me to mention these things. So

1:10:48then what I will do is I will uh add a

1:10:50section called files

1:10:53on documents.

1:10:56Um so right now um we don't have this.

1:11:00So I think this is just example. So

1:11:02typically I have a lot of different

1:11:04documents here in the docs that describe

1:11:07okay what is the process how to write

1:11:09tests how to uh work with API how to

1:11:15design UI and so on right so each of

1:11:17these things each of these as aspects is

1:11:19a separate documentation file and then I

1:11:22link these documents in agents.mmd so

1:11:25then if a task is about

1:11:28process is about implementing something

1:11:30then it goes and reads the process. If

1:11:32the task is about fixing tests, it goes

1:11:34and uh here and reads the testing

1:11:38guidelines. If the task is about UI, it

1:11:40goes and reads about our design system.

1:11:42Right? So it helps our um so we don't

1:11:47put everything in agents MD. So we um

1:11:50here we manage our context in a way that

1:11:52the agent knows that these files exist

1:11:55and if it needs for this specific task

1:11:57it knows where to find for the

1:11:59documentation for implementing this

1:12:01particular task. Right? So um right now

1:12:04we keep things simple. So we only have

1:12:06process and the process is this right

1:12:09tasks are in GitHub

1:12:12uh one at a time. I don't think it

1:12:14actually

1:12:16matters here. Okay. So let's commit

1:12:22commit and push.

1:12:29So this is our context engineering. And

1:12:32the next thing we need to do is we need

1:12:35to do something with these tasks. So

1:12:37okay this one is done but task number

1:12:39two

1:12:41is docker compose. Okay, for docker

1:12:44compost this is a technical task but um

1:12:47here let's say uh I want to take this

1:12:50one

1:12:53uh this authentication. So here I want

1:12:55to have a tasks that is very clear. So

1:12:58there is still some room for ambiguity

1:13:00here. I want to remove all possible room

1:13:03of

1:13:05all possible ambiguity for the tasks

1:13:08that we have. Right? Right? So I want to

1:13:10have the tasks to be very precise and I

1:13:14want to read the tasks to understand

1:13:17that if this task really aligns with

1:13:19what I want because if it doesn't the

1:13:22agent will make assumptions and these

1:13:23assumptions will not necessarily uh

1:13:27match your expectations right and then

1:13:29it will implement something that you

1:13:30don't need right so that's why again

1:13:33speaking about specifications uh we were

1:13:36talking about specification on the

1:13:37project level and we did all this uh

1:13:40talking to chat GPT uh thing right in

1:13:43chat assistant but there are also

1:13:44feature level specifications so this is

1:13:46exactly for tasks so this task I want

1:13:48now to take this task and make it very

1:13:51crisp I want to make it very focused I

1:13:53want to make it as unambiguous as

1:13:55possible right so then when an agent is

1:13:58taking these tasks it doesn't need to

1:14:00make any decisions it doesn't need to uh

1:14:03it just can take it and implement it

1:14:05right so then of course technology

1:14:07choice is not something we have to

1:14:09specify here. We can but it doesn't have

1:14:12to be here. Um so the agent may still

1:14:15need to make some decisions but these

1:14:17decisions should be decisions about

1:14:20userfacing features right and uh in real

1:14:23teams we typically have product managers

1:14:25who are responsible for the function

1:14:27functionality like from the user point

1:14:29of view like what happens if you click

1:14:31this button right uh or what is the um

1:14:36the the task the user is trying to uh to

1:14:39do to accomplish with um our

1:14:42application. Right. So that's why um so

1:14:46what typically product managers do is

1:14:48there is a process called grooming. So

1:14:49they take an issue that looks like that

1:14:52and they groom. They make it more

1:14:53concrete. They make it more they make it

1:14:56less ambiguous. And um what I want to do

1:14:59is I want to turn this into uh something

1:15:03that looks like

1:15:06this. Right? So there is a goal, there

1:15:08is acceptance criteria uh and things

1:15:11like that. Right? So I think I should

1:15:12also we should also keep um description

1:15:16here. It helps but we want to add other

1:15:19things right what is uh what are the

1:15:21acceptance criteria and what things that

1:15:23we don't want to implement right so we

1:15:25want to also be specific about things we

1:15:26want to implement but also about things

1:15:28we don't because otherwise the agent

1:15:30will think okay it's a good idea let us

1:15:32do this right and then it will come up

1:15:34with something that we don't need at

1:15:36least for our MVP and then we have more

1:15:40code to maintain

1:15:42okay so we need a product manager for

1:15:45that right so typically this This is the

1:15:47what uh um in teams the setup uh we have

1:15:52is uh product managers take issues like

1:15:55that and turn them into something that

1:15:58engineers can just take and implement

1:16:00without bugging the product manager all

1:16:02the time.

1:16:03Okay. So now I want to create a folder

1:16:06called team and in this folder I want to

1:16:09create a PM.

1:16:12PM will be our product manager that will

1:16:15take a task and it will turn in it into

1:16:18something that um agents can implement.

1:16:21Right? So then this is the description

1:16:24for our product manager.

1:16:27So you're a product manager. You groom a

1:16:29task before anyone implements it. Read

1:16:30the issue. Uh rewrite it using the

1:16:33template. I will now uh create the

1:16:35template too.

1:16:39So this is our task template. I think

1:16:41I'll um also create old dated documents.

1:16:47So we don't need this architecture

1:16:49anymore. We don't need uh plan anymore.

1:16:54We don't need tasks anymore. So for now

1:16:56I'll keep them in the project but

1:16:58eventually we can just remove them uh

1:17:01because we already have issues. Um so

1:17:04the this is enough for us to to

1:17:07continue.

1:17:09Okay. Um although at the beginning we

1:17:12may still need plan for the agent for

1:17:14the RPM to groom right so to actually

1:17:17not make assumptions about something we

1:17:19already talked about. Um but again so

1:17:23read the issue uh rewrite it using the

1:17:25template.

1:17:33Okay.

1:17:35Make themselves criteria checkable.

1:17:37Someone should be able to point at the

1:17:39screen and say yes or no.

1:17:42Think about the age cases. Uh the person

1:17:44who filed uh it did not. Okay.

1:17:50Um but the important thing here is this

1:17:53acceptance criteria, right? So how do we

1:17:55know that the task is done? Right? So

1:17:57this is what we want to be very explicit

1:17:59and this is very helpful for the

1:18:03engineers. Right? So this is yeah kind

1:18:05of similar to functional nonfunctional

1:18:07requirements. There are different um

1:18:09frameworks for this like this is just

1:18:11the one I use. I also often use user

1:18:13stories like when uh then uh kind of

1:18:18framework like uh as a user uh when I

1:18:21want to do this uh I do that right so

1:18:24these kind of user stories

1:18:26or sometimes you can also like when

1:18:28you're thinking about specifications you

1:18:30can also think of jobs to be done

1:18:32framework there are many frameworks

1:18:34right so but um agents know all these

1:18:37frameworks right and if you have some

1:18:40product management experience or I don't

1:18:42know so some UX experience whatever you

1:18:44can use that like you can just tell the

1:18:47agent what you want to do and then it

1:18:48will do this uh I keep things simple

1:18:51simple so this is the process

1:18:55and um yeah so now what I want to do

1:18:59I'll clear this

1:19:02and um I want you to uh groom all the

1:19:07issues we have in our GitHub start with

1:19:10issue number for and uh if some things

1:19:13are not clear, please use our plan.md

1:19:15document. It's located in the outdated

1:19:18folder. But for now uh this is the what

1:19:21we used to uh seed our issues. So for

1:19:25now for some of the assumptions you can

1:19:28check this file if you need but uh I

1:19:30think the issues should be

1:19:31self-sufficient

1:19:35and um uh please do one issue at a time

1:19:38for now.

1:19:41So I want to start with issue number

1:19:42four, right? So it will now uh

1:19:48so now it should actually

1:19:52Yeah, you see it's it's reading the

1:19:54instructions. It's reading this uh PM.

1:19:56So it knows what to do

1:19:59because we described it. We described it

1:20:01in um our agents. We described that u

1:20:05our work is organized. So it read this

1:20:08and it process. Okay, I did not describe

1:20:10it. I should have actually like it's

1:20:12it's good that uh we did this because um

1:20:16I need to also change the process that I

1:20:19didn't do roles.

1:20:24Yeah.

1:20:28Um

1:20:30cool. This is a part I forgot but good

1:20:33that the agent actually read it.

1:20:52Uh quick question about Corsor. I we

1:20:55going to use Corsor. Uh Alio, you can

1:20:57use whatever you want. You can use

1:20:59cursor, you can use codex, you can use

1:21:01clot code, you can use client, you can

1:21:03use like whatever you want, right? So if

1:21:06you like courser, you can use cursor. I

1:21:09don't use corser. I already have two

1:21:11subscriptions. I don't want to add

1:21:12another one on top of that. Um, so I

1:21:14have codex and I have cloud code and for

1:21:16me this is enough. But cursor is good. I

1:21:18don't think it actually matters what you

1:21:20use because they are more or less on the

1:21:24same level.

1:21:28So um

1:21:30why is it creating issues? Ah okay. So a

1:21:34follow-up issue that grooving number

1:21:36four required. Okay let's see

1:21:45so it descoped some things. Um so I'll

1:21:49tell it

1:21:50uh please first update the issue and

1:21:54then you create follow-up issues. Um so

1:21:56first I want to see the issue that is

1:21:59clearly um that clearly follows um what

1:22:02we want right and then if something

1:22:05according to the PM is out of the scope

1:22:08then we do this afterwards.

1:22:24And then usually when I have to correct

1:22:26the agent when it's doing something uh

1:22:28what um what I also do is uh at the end

1:22:32of the session this is what we will do

1:22:34right now together is I ask hey like

1:22:36based on um the corrections I made what

1:22:38documents we need to update. So then the

1:22:40next time it doesn't uh it knows the uh

1:22:44correct steps it knows the algorithm.

1:22:48Okay. So now it says it's uh groomed.

1:22:51[snorts] So let's see. So the goal a

1:22:53visitor can create an account with a

1:22:55username, display name and password. Log

1:22:56in, log out. No part of the flow touches

1:23:00email. There is no mail back end, no

1:23:01verification, no selfs serve password to

1:23:05that. Okay. So then we have some

1:23:06acceptance criteria.

1:23:08Um so this acceptance criteria I see a

1:23:11bit technical but um means that there is

1:23:13a page. So it renders a form uh username

1:23:18uh already taken renders the form with a

1:23:20visible error. So this is very specific

1:23:23right? So it um we may agree with some

1:23:26things, we may ask it to do some things

1:23:29but this is what we want to have right.

1:23:30So we want to have a very clear set of

1:23:33acceptance criteria that we can uh the

1:23:35agent needs the agent knows what exactly

1:23:38to implement. Uh and then we also have a

1:23:42way to test our application right

1:23:43because this acceptance criteria

1:23:44criteria is what we are going to use

1:23:46later after this agent says it's done.

1:23:49Um we can actually test this uh test it

1:23:53using the same criteria

1:23:56out of scope. So you see that it uh

1:23:59figured out some things that are not in

1:24:02scope

1:24:03and things that are in scope and out of

1:24:06scope. Um so like okay we don't need u

1:24:10brute force defenses.

1:24:13Okay like I I wouldn't included this in

1:24:16the feature at all. It's kind of

1:24:18annoying that it always includes these

1:24:19numbers, but okay.

1:24:27Account settings changes play name.

1:24:33Um

1:24:37for this uh let's uh make it uh priority

1:24:44uh I don't know post MVP

1:24:48priority

1:24:50at attack for that.

1:24:56Um because I I think this could be

1:24:58useful but like I'm not sure how

1:25:02useful it is for actually u maybe we

1:25:05don't even need this like it will create

1:25:08some things that are out of scope but

1:25:10then we will also need to review them

1:25:12and say okay like this not really what

1:25:14we needed.

1:25:20So what I want you to do is just to have

1:25:22uh MVP and post MVP labels for the

1:25:25issues we created from uh 1 to 26 I

1:25:29think mark them as MVP the rest should

1:25:31be post MVP. So whatever uh out of scope

1:25:35issues PM uh I created right now they

1:25:38should be post MVP

1:25:43and also based on uh my corrections

1:25:46based on what we uh did please find the

1:25:49relevant documentation we have um and

1:25:51add some things to this documentation

1:25:54like um yeah like what you think uh

1:25:57should be included there. first make a

1:25:59commit and then make changes.

1:26:03So the reason I want to it to make a

1:26:05commit and then make changes is because

1:26:07I want to uh sometimes look at g and see

1:26:09what exactly changed for code I don't um

1:26:13necessarily want to always do this uh

1:26:16but for changes in documentation

1:26:18especially for changes in how um our

1:26:22team works. So this is something I do um

1:26:26I want to see um because it influences

1:26:29it affects the process right that's why

1:26:31I want to make sure that um we here

1:26:35um we document it and I know what's

1:26:38happening here

1:26:41okay

1:26:45so now we have this MVP label we have a

1:26:47post MVP label

1:26:57So demo data command. I think this is

1:27:00actually uh useful. So let's

1:27:05MVP.

1:27:15Okay. And now when these things are um

1:27:19saved I want to talk about loop

1:27:21engineering. Even though loop

1:27:22engineering comes later here I think now

1:27:24this is the right time to introduce it.

1:27:27So loop engineering is a way to

1:27:33work through a pile of things you have

1:27:37uh or work on a specific thing. Right?

1:27:40Right. So loop engineering um in

1:27:42principle is just this command you have

1:27:44in both codex and um clot. So they are

1:27:49okay I have some here um some things. So

1:27:52prompt engineering this is what we say

1:27:54to our coding agents. Context

1:27:56engineering is all the files that help

1:27:58our agent work like this is agent MD all

1:28:01the processes all the things is context

1:28:02engineering and loop engineering is um

1:28:06it right now we are driving the agents

1:28:09we are typing the prompts but prompt

1:28:12engineering is going kind of one level

1:28:14more so it's more meta so we are we

1:28:18engineering a prompt in such we engineer

1:28:20a loop in such a way that the loop is

1:28:22prompting our application our agent not

1:28:25pass. So we say okay there are these

1:28:27issues. Now what I want let me take a

1:28:30step back. What I want to do now is I

1:28:32want to uh for all these issues I want

1:28:35to create uh I want to process them with

1:28:38a PM. Right? So then um for me what I

1:28:41can do is I can just um set a goal. I

1:28:44can say go through all these things and

1:28:48u create acceptance criteria for them

1:28:51and all the stuff we did right and if I

1:28:54don't use the goal here um the agent can

1:28:57just stop after one or two issues right

1:29:00so what I will do now is I'll say goal

1:29:04groom all the MVP issues

1:29:10I think it asks some lens Um

1:29:15I can do four loose ends. Please make

1:29:20clearly documented

1:29:23decisions. Right? Ideally you are more

1:29:25involved in the process but also you

1:29:27want to review these files afterwards.

1:29:30Right? So now I set a goal and what will

1:29:32it will do? it will uh go through all

1:29:35the MVP issues one by one and if at some

1:29:38point

1:29:40um

1:29:42it stops the loop will prompt the agent

1:29:44to continue right so it's not I will not

1:29:46need to babysit this agent and see okay

1:29:49did it stop did it finish the task did

1:29:52it uh groom all the tasks because I have

1:29:55this uh goal the goal will keep on uh

1:30:00bugging the agent to keep on prompting

1:30:03in the agent to continue. Right? So this

1:30:05is the idea behind loop engineering.

1:30:07There are two types of um loops in cloud

1:30:09code. One is goal.

1:30:13This is what I use. Another one is loop.

1:30:15So loop is uh a scheduled prompt like

1:30:18you can send a prompt like every 30

1:30:20minutes saying hey how's how are you

1:30:22doing? Like are you done yet? Something

1:30:24like this, right? And then cloud code

1:30:25can also stop the the loop. You can say

1:30:27you can instruct it. Um so there is a I

1:30:30sent set a loop once the job is done

1:30:33please stop the loop right um

1:30:37in case of codex you don't have loop you

1:30:39only have goals but this is something

1:30:41that you can actually implement yourself

1:30:43if your agent doesn't support it uh you

1:30:45can use stop hooks and if you run it

1:30:48your agent in a tumix session uh you can

1:30:51also send by through chrome you can send

1:30:54some messages to this t-ox session uh

1:30:57regularly but like I use flat code I use

1:30:59codex both of them support that I'm not

1:31:02sure about the others like corsor

1:31:04um I see a comment from Krishna loop

1:31:07engineering is a pretty much hype in

1:31:10frontier labs is it used uh for

1:31:12day-to-day activities I use goal all the

1:31:15time I use it very often I don't use

1:31:17loop often but goal is um this is

1:31:20something I use very regularly okay so I

1:31:23changed a bit the order so I think I

1:31:25will update this article and I put this

1:31:28um here after um

1:31:32uh after grooming.

1:31:34But now let's see what is actually

1:31:36happening. So it's working and what I

1:31:38can do in parallel is I can start

1:31:40another agent.

1:31:43So uh I need to go to TMP

1:31:47and this is um how do we call it?

1:31:52Oops. What's happening?

1:31:56It's AI dev tools experiments uh project

1:31:59feedback. Then I start another team

1:32:02session and it will be what skill

1:32:04permissions. Okay. So I'm starting a new

1:32:06session. So now I have this session.

1:32:08It's grooming this sessions.

1:32:12So there are acceptance criteria and

1:32:13stuff. Um now what I want to do is for

1:32:16the issues that we groomed

1:32:19like for example this one number four

1:32:23right? So I want to implement it. So for

1:32:25that uh I want to create a software

1:32:28engineer. Right? So we have a product

1:32:30manager. Uh now I want to create a

1:32:33software engineer. Software engineer

1:32:35will actually take this thing and it

1:32:37will implement them. So um let me see

1:32:43here. So, I'm going to create

1:32:47a software engineer

1:32:50and I'm going to put this thing here,

1:32:58right?

1:33:01Okay. Um, so now, um, I think I will

1:33:06need to maybe restart the session. uh

1:33:08because I also want to no I will not

1:33:11need to restart the session but I will

1:33:13need to update our process to also

1:33:16include

1:33:22here. So this is something that agent uh

1:33:24added. I'll keep it makes sense. Um

1:33:29so I need to add the engineer here.

1:33:32Right. And now uh in a fresh session

1:33:36I'll say implement issue number four.

1:33:40Okay. So what I expected to do is to

1:33:41discover that uh um there is this um

1:33:47role software engineer.

1:33:53So you see process. Yeah. So it found u

1:33:57the software engineer

1:33:59role.

1:34:01Um yeah, I think I am a bit early for

1:34:04that. Um maybe we will also need to

1:34:08implement

1:34:10two and three

1:34:12or four.

1:34:16Yeah, right. Cuz like uh none of the

1:34:19things that we need uh exist. Well, some

1:34:23of them exist, but uh actually I don't

1:34:26think we can even run that.

1:34:34Okay,

1:34:38I need to speed it up. So now it's

1:34:41implementing uh 2, three and four. Uh

1:34:44then I will also need to have a QA

1:34:47engineer a tester because we have this

1:34:49acceptance criteria, right? Um so we see

1:34:53this acceptance criteria but the thing

1:34:55is engineers um usually what happens in

1:34:58teams we have a special role for testing

1:35:02and often times if I wrote my code the

1:35:05code myself um

1:35:09then um I am less critical of this code

1:35:12right so usually you need a second pair

1:35:14of eyes to look at your code to review

1:35:16the code and u in companies where you

1:35:19don't have you usually have uh some sort

1:35:22of like uh PR review uh code review like

1:35:26these kind of things. Um so what we want

1:35:29to have is this sort of review and this

1:35:31sort of testing right. We want something

1:35:34else not the engineer to review the work

1:35:37of an engineer and we want this

1:35:39something QA engineer to say if actually

1:35:42all acceptance criteria pass and then

1:35:45the verdict will be pass right or some

1:35:47of them fail and then we will need to

1:35:50ask the certain engineer to implement

1:35:52the things right so we need a third role

1:35:55the third role will be the QA engineer

1:36:00okay so we have team

1:36:03Q engineer. So this is the role the the

1:36:07description

1:36:09um so here in description I say um how

1:36:13exactly the QA engineer should behave

1:36:15and the important thing is it's always

1:36:18either pass or fail so it's always

1:36:19binary true or false right um

1:36:24so yeah the the goal for the Q engineer

1:36:27is to read the acceptance criteria and

1:36:29check each single one uh against the

1:36:35uh test. Um I think we will need to

1:36:37update it. I'll ask maybe the next in

1:36:41the different session

1:36:44to update it.

1:36:58update for our project

1:37:01because we have npm here, right? So, npm

1:37:04is u not really uh needed.

1:37:10Okay. And this should be the quote.

1:37:15Okay. Um

1:37:18so we have the Q engineer and we need to

1:37:21also add the Q engineer in the process

1:37:27right so now we have these three roles

1:37:29we have PM we have engineer and we have

1:37:30QA right so PM grooms the task makes it

1:37:34very concrete so we use this kind of

1:37:36specifications uh for making the task

1:37:40complete concrete so the engineer

1:37:41doesn't need to um make a lot of

1:37:44decisions Then engineers implement the

1:37:46groom tasks and QA checks the work of

1:37:49the engineer. And what it can do is it

1:37:52can decide to whether the work is good

1:37:56and it passes the criteria or the work

1:37:59is bad and the criteria are bad. Right?

1:38:02And what we happen when we put all these

1:38:05three things together, all the three

1:38:07agents together is um what currently

1:38:10people call graph engineering. If you

1:38:12open Twitter, you can see. So we have

1:38:14this graph. So this sequence of steps

1:38:16that we have right so first we take a

1:38:19thing from the pool right our pool is um

1:38:25uh this set of issues so we take one

1:38:27thing from this pool this is an issue

1:38:29and if this issue is not groomed yet we

1:38:31groom it. So this is the first step in

1:38:33our uh pipeline right? So once it's

1:38:35groomed then an implement the engineer

1:38:38takes this right. So the implementer

1:38:41takes this and uh works through this

1:38:44then the next step is test. So we have

1:38:46this quality assurance and there are two

1:38:49uh two possible outcomes. One outcome is

1:38:51pass then the task is done right.

1:38:54Another outcome uh is the task is not

1:38:58done. It fails uh the verdict is fail.

1:39:01So then we um loop it back to the

1:39:03implementer. The implementer needs to

1:39:05fix the task right and then um so this

1:39:09is a graph right. So we have different

1:39:11responsibilities and the task goes

1:39:14through uh these responsibilities. So um

1:39:17this is not new. the term appeared only

1:39:19like I don't know yesterday or when was

1:39:22it uh a couple of like I don't know

1:39:26everyone on Twitter is talking about

1:39:27this but this is a pretty simple concept

1:39:29and um I have actually been using this

1:39:33kind of thing for quite some time and uh

1:39:37there is uh this article

1:39:41um I built an AI agent team for software

1:39:44development um you can check how I do do

1:39:48this. So in in addition to PM software

1:39:50engineer and tester I also have a Q

1:39:52engineer you can check how I organize

1:39:54the process. So here I just show you

1:39:56like a simplified version of that. Um

1:39:59but um in this document there is a more

1:40:01um a version that I actually use right

1:40:04so that the process that I have is a bit

1:40:06more uh complicated but you can check it

1:40:09but right now I want to actually

1:40:12implement

1:40:15this graph. So I want to make sure that

1:40:18um we can follow the steps and for that

1:40:21we need to have an orchestrator. So the

1:40:23orchestrator

1:40:24is uh the main agent uh the main session

1:40:28of our agent that can start the PM that

1:40:32can start the implement that can start

1:40:33the tester. So we don't have to do this

1:40:36ourselves right. So here I um have to go

1:40:40here and type right here I also have to

1:40:43go here and type. So in a way for me

1:40:46this orchestrator is both loop engineer

1:40:49and graph engineer right? So it knows

1:40:51how to prompt these agents. So I don't

1:40:53need to pro to to prompt them myself and

1:40:56it follows this process. It follows this

1:40:58graph right. So we need to um

1:41:02to take this the rules

1:41:06uh process.

1:41:09The main session is orchestrator. It

1:41:11launches the PM the engineer and QA as

1:41:12sub aents. it does not groom implement

1:41:14or test itself. So typically previously

1:41:16when we were hey implement this task it

1:41:19would um start a it would do this within

1:41:23the session. Now we say you need to

1:41:25launch a sub agent for that. So both

1:41:27codex and cloud code uh and also open

1:41:30code and probably other uh engines can

1:41:33do that. They can launch sub aents.

1:41:36Um

1:41:38okay life cycle uh and we basically

1:41:41describe the graph right. So we describe

1:41:43the graph in simple terms in simple

1:41:45words

1:41:46uh one issues at a time. Um like I I

1:41:50wouldn't necessarily enforce it. We can

1:41:52say um you can work uh on up to five

1:41:59issues at the time.

1:42:02Yeah. Let's keep things simple. So in

1:42:04reality I actually um work on oops I

1:42:09work on many issues at the same time

1:42:11just to kept to keep things simple and

1:42:14see this um loop and this graph in

1:42:17action we will actually uh now do this.

1:42:21Okay so let me see um so it updated the

1:42:24key engineer so let clear it

1:42:28uh it is still implementing these

1:42:30things.

1:42:31Okay.

1:42:34So it's going through issues from one to

1:42:37four

1:42:39but I think so issues one to three um

1:42:42they are fairly technical. The first

1:42:45real user issue start with number four.

1:42:48So I think this for us is the

1:42:50interesting issue to actually um use

1:42:53this approach use this framework. So for

1:42:55these ones uh they are more like um yeah

1:42:59they they just need to I would not use

1:43:01any fancy process for them. I just would

1:43:04get them done

1:43:09docker compos environment.

1:43:14Okay. Okay. So it probably mentioned

1:43:16them because uh there is dependency on

1:43:18that right

1:43:25out of scope.

1:43:27Okay, whatever.

1:43:29Um so where are we?

1:43:32So I kind of wanted to show you the

1:43:34actual loop.

1:43:36So let me let's wait till it finishes.

1:43:39Um

1:43:40yeah. So graph is basically a workflow.

1:43:43Yes.

1:43:45How did you write this? Uh what do you

1:43:47mean? Um how did I write what? How did I

1:43:50write uh this um these things? So for me

1:43:54this is based on I've been using agents

1:43:58for quite some time. So this is the

1:44:00process I follow when working on my own

1:44:03uh projects. And what I show here in

1:44:06this session is more like a distillation

1:44:08of this. Right? So if you take a real

1:44:11project that I have, let's say um AI

1:44:15shipping locks, right? So I I think I

1:44:17should go to GitHub.

1:44:22So the process there is more

1:44:24complicated, but the idea there is kind

1:44:28of similar, right? So you have this

1:44:29process.md

1:44:31document that describes the process. Um

1:44:34so the graph here that we have is a bit

1:44:36uh more complicated, right? Uh but uh so

1:44:40what I did here is I took all of this

1:44:44from many different projects where I use

1:44:47uh this approach and I kind of condense

1:44:50it into something right and at the

1:44:52beginning you need to start simple right

1:44:55and then the file will grow itself

1:44:58because at the be at the end after each

1:45:00session you can say hey like what do we

1:45:02need to improve in the process or when

1:45:03you see that you need to steer the agent

1:45:07manually. So you ask it hey like can you

1:45:09please update our process so I don't

1:45:10need to do this right? So for example I

1:45:12say which kind of models it needs to use

1:45:15for the agents right and then because I

1:45:18use uh often I use clot and codex at the

1:45:21same time. So I have the the agents that

1:45:24are um defined in the clot format but

1:45:27then I also describe for codex how to

1:45:30actually um run these agents.

1:45:34So this is how these things appeared.

1:45:36Okay. I do are we down here?

1:45:40Uh

1:45:44but anyway, so uh now if I want to show

1:45:48you this uh graph engineering, this

1:45:50would be it. So I write a goal and write

1:45:53work through

1:45:57the back block, right? So now this would

1:46:02uh maybe I need to

1:46:06focus on MVP

1:46:10issues only.

1:46:13Right. Um

1:46:17okay. So I'll let it finish and then I

1:46:20run this. Um

1:46:25yeah because I want to show you how

1:46:26exactly it looks like how it starts

1:46:28ovations and stuff.

1:46:33Um,

1:46:35let's stop. Which issue are you

1:46:40working on?

1:46:47Okay. Um close uh

1:46:51the done issues

1:46:54and create a

1:46:59note to

1:47:04issue number four

1:47:07about what what you have done so far.

1:47:13Okay. Okay. So I want to just take this

1:47:15and run this as a separate session.

1:47:37Okay. Smart software engineer. The

1:47:40engineer doesn't close issues. K check

1:47:42some you're already that okay

1:47:52so I just do this then I start with or

1:47:58see comments

1:48:08okay

1:48:10so let's see the first. So this agent is

1:48:13the orchestrator, right? So we see that

1:48:15goal is active. So it will figure out

1:48:17the current state

1:48:20and uh it will see what is happening

1:48:24there. Right? The issue is already

1:48:26groomed has a be work in progress

1:48:29comments. Um

1:48:33so then it says handing number four to

1:48:35the engineer. You see now we have this

1:48:38thing here. So there is actually an

1:48:42agent a sub agent that started it. So

1:48:45because I did not follow the convention

1:48:48for defining agents that is in clot

1:48:52that's why it just calls it clo right.

1:48:54So if I want to have proper names here

1:48:56if I want to see that this is actually a

1:48:58software engineer so then I would need

1:49:00to do something like

1:49:03here go to clot create agents and then

1:49:07put them here. I want to make it um

1:49:11engine agnostic that's why I want it to

1:49:13work with any coding agent that supports

1:49:16sub agents and this approach will work

1:49:18right um so it will work in codex it

1:49:21will work in open code I don't know if

1:49:24if cursor supports sub aents if it does

1:49:26it should also work there right so and

1:49:30this is it um what I will do is it will

1:49:33probably take a few hours to actually go

1:49:35through this spec lock and implement all

1:49:37the things so what I will do is I will

1:49:40put all the code online and I will refer

1:49:43it and you can also look at the issues.

1:49:46Um I can also include a few issues from

1:49:48AI shipping labs so you see how I use it

1:49:51on real projects.

1:49:54Okay. Um do you have any questions? Um I

1:49:57know we took a bit more time. Uh yeah so

1:49:59it wasn't uh 90 minutes it was more but

1:50:02I think I covered everything. So um

1:50:05please subscribe to Substack. So, I will

1:50:08be putting more content here.

1:50:12Click on this button.

1:50:14I will putting more content here. So,

1:50:16I'll put the next three uh lessons here

1:50:20too. Um so, yeah, if you want to keep um

1:50:26um if you want to get notifications

1:50:28about this, so please subscribe

1:50:30and I will of course also based on what

1:50:33we did today, I am going to update it.

1:50:35I'm going to add some pictures. In fact,

1:50:37I published this like 2 minutes before

1:50:40we started. So, I will need to to update

1:50:42it because it was like a bit rogue kind

1:50:44of.

1:50:46Okay. Um

1:50:48well, that's it for today. Um

1:50:52yeah, I see some questions. So, per

1:50:54personal project. This is how it's done.

1:50:56Uh working with Markdown files. Yes. So,

1:50:58this is how I do it for my personal

1:51:00projects. Well, for me, um,

1:51:03I have both projects that I work on

1:51:05professionally and projects that I work

1:51:07on personally. And all of these projects

1:51:10kind of follow a very similar structure,

1:51:13right? So, I describe uh this thing that

1:51:16the link is at the end.

1:51:20Um, so this is more or less the process

1:51:22I use for all my coding projects, right?

1:51:25And uh when I start a new one, what I do

1:51:28is I say uh hey like there is a process

1:51:31that I use in this and this project but

1:51:34in this project the new project that I'm

1:51:36starting I use this and this

1:51:37technologies. So please take the process

1:51:39from there adjust it to new technologies

1:51:42and then commit. So this is going to be

1:51:44my first like commit uh in uh in a new

1:51:47project right even before I bootstrap

1:51:49the project.

1:51:52Of course like I also include plan I

1:51:54also include the architecture decision

1:51:56and then when it's there then I

1:51:58bootstrap the process and then I start

1:52:00working on the issues and then typically

1:52:02the first few issues as you see they are

1:52:05more technical so for that I don't need

1:52:07to follow the process but once we start

1:52:09talking about um userf facing things so

1:52:12then I want to follow the process I want

1:52:14to have like always have a QA always

1:52:17have a product manager not just the the

1:52:19implementer for some things I use this.

1:52:22So for example, for these courses, I

1:52:24started this project long time ago and

1:52:27um so here I don't really follow the

1:52:29process. Um

1:52:33but yeah, so there I need to be a bit

1:52:35more careful with tests and stuff,

1:52:37right? Because in practice I see that um

1:52:41following this process, especially the

1:52:43tester part really helps. The tester

1:52:45often finds some things that a software

1:52:48engineer missed. So this is quite

1:52:50important.

1:52:53Um one note though that if you follow

1:52:55the process you will consume a lot more

1:52:58tokens right because now in uh um in

1:53:02addition to just implementing this thing

1:53:04you have all the other things right and

1:53:06then uh naturally your token consumption

1:53:08goes like I don't know four five times

1:53:11more for each single issue because you

1:53:13need to groom it you need to implement

1:53:15it you need to test often times tester

1:53:17decides to uh fail to say that uh this

1:53:22is failing. So then you need to go back,

1:53:24you need to implement this. Uh but what

1:53:26you get in return is uh better quality.

1:53:30Okay, I don't see any other questions.

1:53:33Um

1:53:35so I think I should um stop here. Um so

1:53:39thanks a lot. The article will be

1:53:40updated so keep an eye on it and um yeah

1:53:45see you soon and we will have another

1:53:48session like that in a couple of weeks.

1:53:50I don't remember when exactly. Think I

1:53:52can check.

1:53:56Uh I think you can find it if you go to

1:53:58our LM Zoom camp

1:54:02and there is a list here.

1:54:06Where is it? Oh, wrong link. Sorry. AI

1:54:09dev tools

1:54:13here. Workshop number two SVP.

1:54:17And it happens in um what

1:54:23it's in August, so in a couple of weeks.

1:54:28So yeah, that's it. Um thanks a lot and

1:54:31see you around. My