# Capstone: Plan a Trip to Japan

Hi everyone. Welcome to module four for the course. I made a change in my approach from what I explained during the intro video. I decided to hold off on the advanced section for now and bring some topics from it that are more relevant into this section and demonstrate them during the Capstone project. The main reason was that I wanted to get this course to market as soon as possible. The advanced use cases are a bit niche and may only be relevant to a small subset of people — things like collaborating around projects with colleagues using Claude Code, using Claude Code on mobile, or using advanced terminal commands.

I'm also not going to cover image generation in this module. I'll keep that in a bonus or advanced module and release it later.

You might say: for the Capstone topic of planning a trip to Japan, I could just use ChatGPT. I'm actually going to demonstrate that first to show you that yes, it's one approach — ask a question and get an answer in one shot. But I want to show you how to build a system around a simplified topic. The reason I chose a simple topic is that I wanted you to focus on the core mechanics and system-building approach rather than the topic itself. Planning a trip is something everyone can relate to. I'm going to make it complex and build a system around it, so you can see how to think about building systems with LLMs around an easy-to-understand use case — and then apply that way of thinking to your own work.

In this module I'm also going to show you a different tool for running Claude Code. We're going to use the Claude desktop app. So we'll use Claude Code inside the Claude desktop app, and for the user interface we'll use Obsidian, which you'll see here. I'll show you how to install it, how to set it up, and how to build your project there. You can build your whole knowledge base there. It's a very simple interface.

Many of you might prefer this over Cursor. The limitation is that it can only show markdown files, not other file types. So if you have CSVs, generated code, or a JSON file like you saw earlier, those won't be visible here. But you can see PDFs, images, and markdown. I personally use this interface for my personal use cases and my knowledge base. I can open up my knowledge base and show you how I'm using it toward the end of the project.

So having said that, you're going to learn a new way of using Claude Code, see how to set up Obsidian, and then I'm going to walk you through building a system around planning your trip to Japan.

First, I created a brief and I want to compare the approach most people take. They go into ChatGPT, say "help me plan a trip to Japan and Korea," and paste in some criteria — four to five weeks, two days of remote work per week, budget, interests, and a request for four alternative routes (two Japan-only, two mixing Japan with Korea). And that's basically it. When I did my own Japan travel planning, I recorded myself talking through my preferences and pasted that voice directly into the prompt.

So we send that to ChatGPT to get the plan.

Look how quickly it generates results. You have all this information now — sources it pulled, maybe 15 to 20 of them. And from here you're stuck. You'll probably go back and forth, give some comments, and it generates more text. You lose the thread. It becomes very difficult to take action or know what the next step is. Many people get stuck in this back-and-forth in chat with ChatGPT.

So now I'm going to take a different approach. Instead of a one-shot answer, I'm going to use both Claude and Gemini in Deep Research mode. For Claude this is called Research mode, and for Gemini it's called Deep Research in the tools. The difference from a single prompt is that these modes go through a lot of sources — I've seen runs that went through 400 websites. Much more comprehensive and broad.

Before invoking Deep Research, I'm going to transform the original prompt into a research prompt. I changed the objective: I'm asking it to use the brief to come up with a comprehensive research prompt. The goal of this research is to feed agents with comprehensive, credible, and recent data to guide me in trip planning. The itinerary is not the goal here — the objective is figuring out routes and logistics with destinations. It should look at credible travel blogs, social media, Reddit, official tourism guides, and other relevant sources.

I feed it the same material and ask it to produce a prompt I can put into Deep Research mode.

It went ahead and created the prompt. It actually created an interactive artifact so I can see different sections of the prompt and click through to the full text. I'm going to enter this into Deep Research mode with web search on for Claude, and do the same in Gemini with the full prompt. This can take up to 45 minutes to run.

Before getting the results, Claude Deep Research asks a few clarifying questions to make sure it has full context. For example, it asked about travel dates — I hadn't specified I'm going in October 2026. I also answered whether including Korea is optional (genuinely optional — I want to be helped to decide) and what matters most when routing conflicts arise. There are usually about three questions based on your message, and then it goes off and does the research.

Gemini doesn't ask questions and just starts gathering sources immediately. Both are very good: Gemini is strong at getting recent sources, and Claude is thorough in going broad across many sources.

While those run, I want to show you how to get YouTube video transcripts. Actually, let me first show you how to download Obsidian and open a space where we can start collecting notes. Then from there we can gather the context we need for the trip. I want a space where all of our files live, so I need to show you the combination of Obsidian and Claude Code running inside the Claude desktop app — and how to view and work with it in Obsidian.

Okay, so Obsidian is a markdown viewer tool. Download it for your device and open it. When you first open it, it asks you to create a vault or open a folder. I'm going to create a new folder called "Demo Vault." I'll open it in Obsidian. Once you open it you get an empty page.

First, I'll create a file called "Japan Trip Resources."

Now I'll go to YouTube and search for "best places to visit in Japan." I'm going to grab a few videos — I'll sort by relevance. I'll grab this one, then "14 Days in Japan," and a couple more. You'd normally watch a bit of each video to check quality, but you don't need to take detailed notes. One of them is an hour long and 1.7 million views, so it's probably a good deep resource. Let's say we have four videos. This is another source of context. Now I'll show you how to use the transcriber we set up to transcribe all of these videos and save the transcripts in this folder.

To do this, I need to launch Claude. Let me bring it up. Open the Claude desktop app and switch to Claude Code. Select the folder — choose the vault folder we just created. Now I want to check if the YouTube transcriber skill is in the context. I'll run the context command and confirm the skill is accessible.

Claude Code is now running inside this folder and has access to all of the files.

I'll ask it: "Do you have access to the YouTube transcriber skill?" It confirms it does. Then I say: "In the Japan Trip Resources file, can you transcribe the first YouTube video and store the raw transcript in a separate markdown file?"

It found the file, found the video, and is using the skill to transcribe it. The transcript file now appears in Obsidian.

Now I say: "Run three parallel agents to transcribe the other three YouTube videos from the resources file, store their transcripts in separate markdown files, and put all transcripts into one folder called YouTube."

Claude Code can run agents in parallel — it figures out which tasks can run simultaneously. You can see two agents running at once. One handled cleanup, the others are transcribing. The long one was the hour-long video, which also used up a lot of the context window. You can check how full your context is with the `/context` command. Once the YouTube transcripts are done, I'll clear the context and move on.

All four transcripts are now in Obsidian, organized under the YouTube folder as markdown files.

Let me go back to the Deep Research results. Gemini is done. What you want to do is export the result as markdown and store it in your vault. Google doesn't have a direct markdown download, so you create a Google Doc from the result, then download it as markdown from there. I've done that and dropped it into the vault.

Claude is still doing its research — it's already gone over 600 sources and is 30 minutes in.

The next stage is getting Reddit data. We have facts from Deep Research from both Google and Claude, and rich information from YouTube bloggers. Reddit is another good source for people's opinions. I'll clear the context first since this is a new, unrelated task.

The provider I'm going to use for this is called Bright Data. Create an account there. It's similar to other tool setups we've done — you find the MCP section, copy the Claude Code installation command, and run it. I'll ask Claude Code to give me the user-scope version of the command so it's installed at the user level.

With the token added, it runs in the terminal and installs the MCP. You can verify by checking the claude.json file — the MCP entry will be there. Clear the context and run `/context`. Bright Data is now available.

Now I'll set the Japan Trip Resources file as context and ask: based on the information here, find all subreddits related to Japan and Korea, find the hottest posts, grab the top three comments for each, and create a comprehensive doc covering people's opinions on different places, hidden gems, great restaurants, and things to avoid.

I'm in plan mode so Claude will show a plan first. It identifies the target subreddits, runs eight parallel Reddit searches using the Bright Data MCP, scrapes them as markdown, and then compiles everything into one comprehensive doc. I approve the plan and it executes.

The result is a doc with more than 3,500 words from Reddit, organized with top comments, hidden gems, tons of specific information. By now I've structured all the research context that will go into the planning.

Before starting the actual planning, I need to do a quick Obsidian setup that will keep the knowledge base organized. Go to Community Plugins, enable them, then browse for two plugins: Templater (for reusable templates) and Dataview (for creating structured views from your notes). Claude Code understands how to use both of these — you don't need to configure them manually.

Now I'll delete the Japan Trip Resources file since we've extracted everything from it. I'll record a voice note describing how I want the knowledge base organized, then explain it and review the plan together.

Here's what I tell it: I want to organize my files in this knowledge base. I'm thinking of having multiple object types — trips, places, and events. These objects would have properties relevant to their type. This is an Obsidian vault, so we can organize these object types with templates. Propose a way of keeping the vault organized, and think about it holistically since I'll probably add other types of notes and information later.

The plan it comes back with is solid. It's going to create a home folder, objects for trips, places, events, people, media, and projects, daily notes, fleeting notes, essays, sources research, YouTube transcripts, and templates for each type. It's also going to reorganize the files already in the vault and add YAML front matter metadata to each markdown file. For example, a place like Tokyo would have fields for country (Japan), region, city type, restaurant/cafe/bar classification, tags, and so on.

I approve and it executes.

As it runs, you can see the folder structure filling in — daily notes, essays, events, trips, sources. All empty folders for now. The templates are being written: one for trip planning, one for events, one for daily notes. Once notes are created, the Dataview plugin will load information from them into a dashboard. The research files are already reorganized under the sources section.

Now I'll clear and switch to Ask Mode. I ask: what are the alternative routes I could take for my Japan and Korea trip? Stay concise, I just want to know what options I have.

It tries to load all the research data but there's a lot of it. It's finding a better way to load it into context. I ask for a bit more detail on why each route is unique, since I can't choose just from city names.

Let's say I go with Route 2. Now I'll prompt it to create a high-level logistics document so I can decide how long to stay in each city. I say: create a document for me. I'm leaning toward Route 2. I want to see where I should start the trip, how long I should stay in each location, and how to route this from and back to Berlin. Just high level logistics first — I want to figure out logistics before getting into activities, because logistics dictate the activities. If you're missing any key information, ask me before creating the document.

It asks me two questions. I answer: I'm thinking of leaving the first week of October and coming back the first week of November. For remote work, I prefer to work when I'm in major cities where I have a longer stretch of time — something like two working days and three or four exploring days per city.

One thing I want to add: I'll start a new session and run `/init` to create a CLAUDE.md file in this vault. This is important because Claude Code loads the CLAUDE.md file in the project folder at the start of every session. Without it, it won't have the context it needs to navigate the vault intelligently.

It went and created the CLAUDE.md. Now I can talk to it to adjust that file as needed.

Now I want to create a transportation tracking document. I ask: I like the itinerary, but now I want to move to booking transportation. Can you create a separate document that tracks transportation? My intention is to go through each leg one by one, book tickets — flights or trains — and provide you with the information so you can update the vault.

I switch out of plan mode. The CLAUDE.md is visible in Obsidian, listing all trips, templates, and so on.

It creates a transport booking tracker. I can check off each leg as I book it. Let's try it: I search Berlin to Tokyo flights on Kayak, pick October 5th, take a screenshot, and say "I booked this flight."

It processes the screenshot, identifies the route, dates, cost, and arrival time. It notes that Tokyo arrival shifts everything downstream. It asks about the return flight and leaves the confirmation number blank since it wasn't in the screenshot. You can fill that in later.

Now I ask it to create an event from this flight — specifically: go to the events folder and create an events object from this flight with as many properties as you can pull from it.

It creates the event with a date start, date end, status (booked), type (flight), title, and a relation to the Japan trip. Now I can ask it to create a trip entity as well: create a trips object for my Japan trip based on the itinerary we agreed on.

It creates the trip and links the outbound flight to it. You can navigate between connected files with command-click — the trip links to the flight event, and the flight event links back to the trip. The bookings section of the trip now shows the flight.

There's also a graph view in Obsidian showing how everything is connected — trips linked to flights, hotels, events. I mostly use the tree view, but the graph is there if you want it.

Let's test one more thing. For my Tokyo destination, I want a list of things to do. For all of those things, add them to places and cross-connect the documents.

Obviously in real planning you'd do a lot more back and forth, but I want to show how easy it is to create and connect these documents. When you're actually traveling, all you need is the trip file, and all the information is right there — accommodations, activities, everything. You can drill into any item. The beauty is that anything you don't need you just delete: if I don't need to notify my bank, I remove that item. You can work on these markdown files alongside the agents.

It creates all the Tokyo locations and organizes them in the Tokyo note. Each location has details on top experiences, hidden gems, how to get there. The Tokyo note now has a table listing all the places, and each place has its own linked entity. With one prompt you have a trip that is its own connected knowledge structure — prioritized, typed (restaurant, day trip, neighborhood, nightlife), with notes.

Now I'm going to switch from this demo project and show you what I've built for my own personal knowledge base and Japan trip, so you can see the end product.

This is my personal knowledge base. Different note types, each with their own templates. I also have a CLAUDE.md here. To keep things organized, I have a skill in Claude Code that runs on a schedule every morning at 8 AM. As long as the app is open and your laptop is on, it processes anything I drop into an inbox folder. I can drop in to-dos, meeting notes, website links, event links, PDFs of tickets — and it routes and organizes each one automatically. It connects to my calendar, my to-do list app, and it uses Reducto to convert PDFs to markdown. For example, if I drop a concert ticket PDF, it reads the ticket, creates an event in my calendar, and stores the PDF in the knowledge base with a link in Obsidian. On mobile, that link opens directly to the ticket PDF. I get an email summary from the agent each morning with a log of what was processed and a running record of the last 50 items.

Let me close everything and go to my Japan trip folder. It has its own CLAUDE.md because there were many resources and I needed to guide Claude Code specifically for this folder. I have pre-departure checklists, pre-booked experiences, accommodation records — all with Google Maps links, hotel names, addresses, and confirmation numbers. If I shifted days from one city to another, Claude Code went and updated everything downstream.

There's also an activities list for each city — things to pick from, with icons for anything that needs to be pre-booked. A pretty organized map of what I want to do.

Let me show you one more thing: I can visualize the trip as an HTML file. I'll switch into plan mode in my Japan folder and say: take the information about the cities I'm staying in, the duration, dates, and accommodations, and create an HTML file with a map of Japan showing the sequence of destinations in a nice UI. When you click on each destination, show the details of where I'm staying.

Here's what it built. I open it in Chrome and it shows the start of the trip, dates for each leg, hotel details, links to Google Maps. You can click through each destination. Some hotel links don't open cleanly — you'd want to be more specific in the prompt — but it's a shareable HTML file you could send to friends to show them your trip plan instead of a Google Doc.

With that, module four is done. I'm not going to release the advanced module right now. With everything you have, you should be able to get going with Claude Code, Obsidian, and Cursor for both personal life and work.

The important part: understand the mechanics of context. Keep it clean, fresh, and lean. Build context about yourself, about your company, and about your projects. When you organize those into different levels, LLMs have access to all of them and build a brain that mirrors a fraction of what you already know. The goal is to keep increasing that proportion so the LLMs' outputs are more relevant to what you want to do.

You now also know how to connect Claude Code to the tools you're already using.

I hope this course was useful in getting you started with Claude Code and these powerful tools. I really appreciate your feedback — let me know in the group discussions if you have questions or challenges. I'll try to engage there and share answers from my experience, and you can share yours with the community too. I'll keep dropping new videos as I explore new topics, and from time to time I may hold live sessions with members to go through questions and exchange ideas. These tools are transforming how we work and how we organize our lives. I really appreciate your presence and participation in the course. Have a great day.
