Configurei um projeto no Claude do zero para implementar uma feature completa

Author: Cortes do Rafael Quintanilha · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Configurei um projeto no Claude do zero para implementar uma feature completa · Channel: Cortes do Rafael Quintanilha · 10:47 · Watch on YouTube

Topics: AI agents, Task and workflow automation, Claude, Prompt and context engineering, Software engineering with AI

Summary

The video demonstrates setting up a Claude project to add a watched-lesson history feature to an online training program. The presenter connects the project to the program’s code repository and an Obsidian second-brain repository, then configures model and effort settings. Claude reviews recent work, recommends a scheduled lesson digest, and starts parallel threads to investigate recommendations and build viewing history. The presenter also reviews UI concepts created as an artifact and shows how to configure and run a weekly routine.

Key points

Claude projects can be created in the web or desktop app, as well as on a phone or iPad; the presenter says this workflow is not available in Claude Code.
The example project aims to add a history of watched lessons, similar to YouTube’s watch history.
The project context can include repositories, files, folders, and Google Drive; the presenter adds both the program repository and an Obsidian second-brain repository.
The presenter configures Opus 3.5 for the coordinator and thread models, setting their effort levels to high and extra high.
Claude reviews the last 14 sessions and recommends a Monday summary routine.
Claude starts parallel threads to investigate watched-lesson history and how it relates to lesson recommendations.
The presenter explains that project workflows are most useful for larger tasks that extend beyond a single session, using planning, execution, review, and acceleration.
Claude creates three UI artifact concepts: a list by day, a grid, and a compact timeline; the presenter prefers the list by day.
A weekly lesson engagement digest can be scheduled for Monday at 8:52 AM or run on demand.

Chapters

Start	Chapter	Summary
00:00	Creating the project and defining the feature	The presenter introduces the goal of adding a watched-lesson history to the training program.
01:33	Adding project context and configuring models	The presenter adds code and knowledge repositories, creates the project, and adjusts its model and effort settings.
04:17	Reviewing work and launching parallel threads	Claude analyzes recent sessions, suggests a scheduled digest, and starts threads to investigate history and recommendations.
06:27	Applying a structured workflow and reviewing UI concepts	The presenter describes when project-based agent workflows are useful and reviews three UI concepts generated as an artifact.
09:10	Setting up and running a weekly routine	The presenter discusses the project’s threads, artifacts, and routine, then demonstrates the weekly digest schedule and on-demand run option.

Quotes

I wanted to make something similar to YouTube that has your history, the videos you’ve already watched — 01:03

It read the last 14 sessions and recommended a digest to me — 04:17

Claude working on three fronts within that same goal — 05:21

this is clearly something for you to do with larger tasks — 07:01

Direction number one is the list by day — 08:37

Concepts and entities

Claude Projects, Claude Code, Anthropic, Opus 3.5, GPT-4o, Sonnet, Haiku, Obsidian, Programa AI, PERA method

Full transcript (original language)

00:00 So I want to start here by kicking off a new project with you guys. And while this project is working, I’ll show you other examples I’ve already done, alright? So okay guys, I’m here on Cloud Web, so this can be done both on the web and the desktop app, okay? Or even on your iPad or your phone, you can dispatch it from any UI, alright? It won’t work from Cloud Code, it’s not in Cloud Code at the moment. So we’re going to create a new project, okay? So you’ll give it a name here, alright? I’ll create a project for us to add a video history to the program. What is that? Let me give a little more context

00:32 here. Alright? “Programa AI” is my AI engineering training program that’s already out, it’s live, and it has the AI Engineer track, currently containing classes on prompt engineering, agentic programming, and AI engineering fundamentals, okay? So, for those who purchased the track, you can acquire it by going to the link that will be in both the description and the pinned comment. You can buy either the track or the individual courses. So, whoever bought any course within the program, you can come here and start watching the classes, alright? So, there are several classes here you can watch, and

01:03 I wanted to make something similar to YouTube that has your history, the videos you’ve already watched, for example, okay? So let’s think we want to do this, and we’ll create it within a project in Cloud Code. So I’ll state here as the goal: I want to create a history of watched lessons in my project similar to how YouTube has history of videos watched. So alright, I used the whisper flow here to edit what the goal is. Once I’ve set the

01:33 goal, I can come here, see, to the context, you see? And here in the context, I can add files, folders, or add manually. If I click on add here, you’ll see there’s an option to add a repository. So I’ll open it here, select the “Programa AI” as a repository, and I’ll also select here, a tip for you guys, my second brain repository to manage documents when I’m working with AI. in my Obsidian. I’ve actually made a video on how I use Obsidian as a second brain to work with AI, it’s one of the most watched videos

02:06 here on the channel. I’ll leave the link for you to watch later. So this tip here is cool because I have my Second Brain, and since Obsidian is basically markdown, it’s excellent for putting into a git repository. So I can put my Second Brain. Pretty cool, right? I can also, if I want to, add a local file or folder or even configure Google Drive, but I’ll leave it like this for now and click on create project, alright? So what you’ll see happen the moment I create this project is that Claude will already start

02:38 trying to set up what we’re going to build, okay? Sounds great, everyone? So , the project was created and Claude even gave me some options on what to do . But before that, I want to show you one thing, okay? If you click here, on this settings button, you’ll see all the project settings, including the icon, the objective, and there are the models and their effort levels, okay? My coordinator model is set to Opus 3.5 . Got it? That’s exactly what I want to do. However, the effort, their default recommendation is low effort.

03:11 Why? Because the coordinator doesn’t execute much; it really just coordinates, it just takes the input and dispatches tasks, right? Typically, in agent orchestration, you have a coordinator that is even more intelligent. The logic here is the opposite, okay? The logic here is a very simple, very direct coordinator, and the work really happens in the threads. Anyway, since I have credits to spare, I’ll set the coordinator to high, and my model that will execute in each thread, meaning the sub-agent, is also set to Opus 3.5. And here I’ll set it to extra high, okay? See, this

03:46 will spend more tokens, but I’ve been running good tests with Opus 3.5, which in my view is the best model for programming today. I even did a review of Opus 3.5. I’ll leave it here for you to watch later. This way works well for me, okay? So, all right. I’ve configured the models and their effort levels here. And I can now, right here, ask Claude to “setup from recent work.” Let’s see what it does, right? Configure this project based on my

04:17 recent work, okay? Look what happened. Remember? I asked it to configure the project, and what did it do? It gave me a setup recommendation. Look, it read the last 14 sessions and recommended a digest to me, right? A summary of watched lessons every Monday. I could set this up right here. I’ll actually do this: I’ll click update setup because you’ll see that it will create a routine inside my project. I have the threads, which are my sub-agents, I have the libraries here, which hold the resources, and I also have routines,

04:49 meaning tasks executed on a schedule at a specific time. And another thing it thought of was: “Well, I can start working on two things. I have two threads I suggest you build: the watched lesson history and aligning history with lesson recommendations. Got it? Let’s do this: let’s click on start two threads. And what will happen is that it will automatically initialize two tasks at the same time. So, look, we now have Claude working on three fronts within that same goal,

05:21 which is our objective of creating the lesson history, okay? So it’s very elegant, a very cool way for you to see it working, you know? And you can separate the work well, you can orchestrate it well and see exactly what each thing is doing, including the steps, see, what the steps are? The first thing it’s doing is reading how the recommendation works, right? And the other one is reading how the lessons, videos, and progress are modeled, okay? All of this using GPT-4o with reasoning extra, and I can change

05:55 it, I can switch to Sonnet, to Haiku, and change the reasoning as well. So it turned out to be a very interesting interface. And look at how that session is doing. One thread is working, look, it has already done four out of seven tasks and one is solved, see: align history with lesson recommendations. So , basically, what it did was investigate what needed to be done and how the recommendation works. And in the end, it sent the result over to the build lesson watch history thread, right? So you can see here that in the

06:27 original conversation with our orchestrator, one thread finished and the other is still running. I can click here at any time to see it. Got it? And look how cool, it started here with the back end, okay? And then, I can already see that one of the next steps here is drawing on the canvas, meaning creating an artifact within Claude for what the UI would look like, okay? So, before it actually builds the UI, it will be reviewed by me. This is great for ensuring that whatever it creates will respect my design system. Now, folks,

07:01 look, this is clearly something for you to do with larger tasks, it’s not meant for everything, right? If it’s something you can solve with just the right prompt. Actually, if you think from an agentic programming perspective , I even made a video recently where I talk about why you should learn SDD, I’ll leave the link here for you to watch later. In that video, I introduce the Pera method, right? The method I even show in the agentic programming course for the program. And what is this method? It’s a way for you to

07:33 program with AI agents. First, by planning what you want to do, then executing the tasks, reviewing what was done, and finally, when the scheme is fully functional, you go ahead and accelerate development. So, this Pera method is a method that allows you to develop tasks much faster, okay? And that is exactly when the project scope grows that this makes sense. Look, Claude Projects is an excellent tool for this, right? For you to work beyond

08:05 the scope of a single session, right? So here in our example, it worked by researching what needed to be done and started with the back end. I can see here that it is now in the stage of designing what the UI will be, right? So we’ll see very soon what it’s going to prepare for us. And the last thing I wanted to show here regarding projects is precisely the canvas part, right? How you can also have an artifact right here within the project view. And look

08:37 there, it created the artifact for us, which we can see right here in the library. It created, see, a class history artifact. And we can see exactly what concepts it created. So, it created three directions for me, see ? Direction number one is the list by day, right? Everything I watched each day, including, look, continue watching again, see? Pretty cool. Concept number two is like a grid, and concept number three is a compact timeline. Wow, all three were cool, okay? I think the list turned out better, and coincidentally,

09:10 it even matched his recommendation. All right? So I can select here, look, list by day. And then I’ll also say here, the problem is we don’t have a thumbnail. Can you try getting a frame from the video? Just so we have a way to display an image here, right? Because otherwise, it just keeps appearing, you know? A placeholder. It feels like something is missing, but look how cool, we managed to bring the artifacts directly into the project here as well. So here we have, look, PRs, artifacts, our conversation with

09:41 all the threads, and even routines that we could create. In the Haiku and Managed Agents video, I showed how you can have an agent running in the cloud, on the cloud infra, spending the credits here. Of course, it’s drinking from your subscription, right? The credits here are your subscription’s credits. So you can create this routine which, as I said, is a job, right? It’s a task that will be executed at a specific time. Here it will be every Monday, at 8:52 AM, precisely with these instructions. Use the program’s

10:13 site post, summarize the last few days of lessons watched, the 10 most-watched lessons, etc., etc. Post a summary here , okay? So, this is the routine that will be executed right now. Or I can click here, look, and run now. And if I click run, then I can anticipate this routine and execute it whenever I want. All right? We’ll go back here, look, and see that it’s running here, look, weekly lesson engagement digest. It’s also here in one more translated. So, folks, this interface turned out really good, okay? We have to take our hats

10:47 off to the Anthropic team. Indeed, when it comes to UI and UX.