On Day 1 of planning our Japan trip, the agent caught something I would have discovered at the immigration counter in Narita.
Japan’s new eVisa system — launched for Indian tourists in September 2025 — requires you to display the visa on your smartphone with an active internet connection when you arrive. Not a PDF. Not a screenshot. Live internet, at the counter. An Indian couple travelling to Tokyo earlier this year was reportedly held at Narita for 40 minutes because their phone had no data. The fix takes five minutes: buy a Japanese eSIM before you land. But only if you know to look for it.
A chatbot wouldn’t have flagged this. I asked Claude in the browser for a Japan itinerary a week earlier and got a perfectly decent list of temples and ramen shops. It didn’t know I was travelling on an Indian passport. It didn’t know the visa rules had changed. It just answered the question I asked.
An agent works differently. You give it the full picture — who’s travelling, what everyone cares about, what could go wrong — and it works through the problem. It found the visa catch. It flagged that the custom Onitsuka Tiger shoes I wanted take 6–7 weeks to make, not something you walk out with. It built a day-by-day itinerary that remembered my 10-year-old son’s Demon Slayer obsession on Day 6, not just Day 1.
This article is a no-code walkthrough of how to build one. No Python. No terminal. No coding background required. By the end you’ll have a reusable setup you can point at almost any complex planning problem — not just travel.
Not a member? Read it for free
A quick confession before we start
I’m a software architect. I build systems for a living. When I first heard people describing Claude Code as an “agent framework,” I was skeptical in the specific way developers are skeptical of things that sound like marketing dressed up as engineering. I build these things. I know what the seams look like.
What changed my mind wasn’t a demo. It was a genuinely embarrassing moment: I’d been talking about Japan with my family for two years. Two years. We never got past the idea stage. Every time it came up, someone would say “we should really plan that” and the conversation would dissolve into the Pune classic — a WhatsApp group that goes in circles for three weeks and produces nothing except a PDF someone forwards from a travel agent.
I’d also already started a spreadsheet that was going nowhere.
The Japan trip finally happened because I stopped trying to plan it and started treating it like a project I could delegate. That’s the mindset shift. Everything else is just setup.
What is an agent, and how is it different from just chatting with Claude?
Most people use AI the way they use a search engine: ask a question, read the answer, move on. That works for simple lookups. It breaks down the moment a task has multiple steps, conflicting requirements, or needs to actually produce something usable.
Here’s the core difference:
When you ask Claude in the browser for a Japan itinerary, it gives you its best single answer and stops. It doesn’t take actions, doesn’t save files, doesn’t follow up. One shot.
An agent runs a loop:
Observe → Think → Act → Repeat
You set a goal. The agent plans, researches, acts, and keeps going until the job is done. For trip planning that means: understand who’s travelling → research visa requirements → check what’s actually available in October → balance everyone’s different interests → draft a day-by-day plan → flag what needs advance booking → save a file you can actually use.
A chatbot does step six. An agent does all of them.
The other thing that makes agents genuinely different: they keep you in the loop as they work. The best outputs don’t come from a single perfect prompt — they come from a conversation. You’ll see exactly how that works below.
The mental model: a chatbot is a smart colleague you can ask questions. An agent is that same colleague given a project — they’re actually working on it, coming back to you when they need a decision, not just when they’ve finished talking.
As a developer, I’d put it differently: a chatbot is a stateless API call. An agent is a process with a goal. Once you see it that way, the use cases start multiplying fast.
Setting up Claude Code: 10 minutes, no code
You need two things.
Step 1: Download VS Code (free)
Go to code.visualstudio.com and download for your OS. VS Code is just a workspace — a place where Claude can read, write, and save files on your computer. You won’t be writing any code in it. If you’re a developer already using VS Code daily, this step is already done.
Step 2: Install the Claude Code extension
Open VS Code → click the puzzle-piece icon (Extensions) in the left sidebar → search “Claude Code” → install → sign in with your Anthropic account. Done.
Step 3: Create a project folder
Make a new folder on your desktop. Name it something relevant —
japan-trip,job-search, whatever the project is. Open it in VS Code via File → Open Folder. This is where your agent reads context, saves outputs, and builds files you can use.
Three steps. Most people are up and running in under ten minutes.
The CLAUDE.md file: the step most people skip, the one that changes everything
Before you run a single prompt, create a file called CLAUDE.md inside your project folder.
Claude reads this file every time it starts a session. It’s your agent’s standing brief — its permanent memory of who you are and what it’s working on. Without it, the agent guesses. With it, every prompt you send builds on a foundation the agent already knows.
This is the difference between a generic output and one that feels like it was built specifically for you.
For me, this was the moment it clicked technically. As someone who writes system prompts and specs for a living, CLAUDE.md is essentially a persistent system prompt scoped to a project. The difference is that here, you don’t need to know that’s what it is. You just write what matters to you, in plain English, and the agent uses it.
Here’s exactly what I put in mine for the Japan trip:
## Who we are
Family of 3: me, my wife, and our 10-year-old son.
Travelling from Pune, India. Indian passports — visa required for Japan.
Trip: 7 days, mid-October.
## What each of us cares about
Me: shopping. Specifically the Onitsuka Tiger Omotesando Nippon Made
store — the only place in the world for custom Mexico 66 Deluxe shoes.
Wife: Instagram-worthy photos. At least one iconic shot per day.
Son: anime and manga — Demon Slayer, Studio Ghibli. This trip should
feel like stepping into his world.
## Non-negotiables
- Research Japan visa requirements for Indian passport holders
(flag any 2026 changes - I've heard things have changed)
- All budget estimates in INR. Mid-range travel style.
- Research flight options from Pune (PNQ) or Mumbai (BOM)
- Factor in intercity transport within Japan (shinkansen, metro)
- Balance activity with downtime - 10-year-old, not a marathon runner
- Flag anything needing advance booking and how far ahead
## Output format
Save everything to japan-trip.md with:
- Visa checklist with deadlines
- Flight options + INR cost estimates
- Full budget breakdown (flights + hotels + food + transport + activities)
- Day-by-day itinerary (morning / afternoon / evening)
- Advance booking list
- One anime highlight per day for my son
Every detail in that file shaped every output the agent produced — without me repeating myself in any prompt. My wife’s Instagram requirement showed up on Day 6. My son’s Ghibli interest shaped a museum recommendation I hadn’t thought to ask for. The Pune-to-Mumbai leg of the flights was factored in automatically.
This is also what makes the whole setup reusable. You write the CLAUDE.md once per project. When you start a new one — a different trip, a job search, a research brief — create a new folder, write a new CLAUDE.md, and the agent immediately knows the new context. The pattern is identical. Only the brief changes.
Running the agent — and why you should tell it to stop and ask you questions
Open Claude Code in VS Code. Before typing anything: switch to plan mode. This makes Claude show you its intended steps before taking any action. You see the approach, can correct it, and only then does it start working. Always use this first — it costs 30 seconds and prevents a full restart.
Here’s my opening prompt:
We're planning a 7-day family trip to Japan in mid-October.
Three of us — me, my wife, and our 10-year-old son.
Indian passports, travelling from Pune.
Start by researching:
1. Visa requirements for Indian passport holders - flag any 2026 changes
2. Best flight routes from Pune or Mumbai to Tokyo, with INR estimates
3. What makes mid-October a good or bad time to visit Japan
Then pause and ask me questions before building the itinerary.
Save all research to japan-trip.md as you go.
Notice: pause and ask me questions. This is the most valuable instruction you can add to any planning prompt. A plan built without dialogue is a guess. One built through conversation is personalised.
I’ll be honest — as a developer, my instinct was to over-engineer the first prompt. Write a perfect brief, anticipate every variable, hand it a spec. Resist that instinct. The questions the agent asks before building are often better than the questions you’d think to ask yourself. More on that in a moment.
After the research, the agent came back with this before writing a single day:
Agent — questions before building:
- Tokyo only, or do you want Kyoto and/or Osaka? Each adds meaningful travel time.
- For the Onitsuka Tiger Omotesando store — the custom NIPPON MADE order takes 6–7 weeks for delivery. Do you want the shoes shipped to India, or is this mainly about the in-store experience?
- Your son and anime — Akihabara (merch, figures, manga) or teamLab Borderless (immersive digital art, very Ghibli-feeling)? Which is the priority?
- Your wife’s Instagram shots — iconic and recognisable (Shibuya Crossing, Fushimi Inari gates) or off-the-beaten-path spots that feel more discovered?
- Rough total budget in INR for the full trip, excluding flights?
Question 2 is the one that mattered most. Six to seven weeks. Without knowing that, you walk into the store expecting to leave with a box. Instead, you’d be ordering shoes for December delivery. This is exactly the kind of constraint a developer would think to check — and exactly the kind of thing you forget to check when you’re also managing a visa application, a 10-year-old’s wishlist, and a wife who has already decided which Fushimi Inari photo she wants.
The agent didn’t just tell me where the store was. It thought about my specific situation.
I answered all five questions. Only then did it build the itinerary.
What the research phase surfaced
Visa — two things I didn’t know
Already covered the eVisa internet requirement above. The second thing: walk-in VFS submissions are no longer accepted at most centres in India as of March 2026, including Mumbai. You need a pre-booked appointment, and slots fill fast during peak season. The agent flagged a 6–8 week lead time. That was the first thing I actioned after reading the output — before I’d looked at a single day of the itinerary.
One more thing the agent caught: Japan takes the itinerary submission seriously. A vague “I plan to explore Japan” is grounds for delay or rejection. The detailed day-by-day plan I was building wasn’t just for travel — it was a visa requirement. Planning and paperwork became the same document.
Flights
No direct Pune–Tokyo flights — which most people in Pune already know, but it’s worth spelling out for anyone who hasn’t made this calculation yet. Best route: drive or catch a train to Mumbai (BOM), then fly to Tokyo Narita (NRT). Air India and ANA both fly direct in roughly 8–8.5 hours. Vietnam Airlines and IndiGo via connecting hubs are cheaper but add 3–5 hours. For mid-October, round-trip per person runs ₹55,000–₹85,000 booked 8+ weeks ahead, or ₹85,000–₹1,20,000 if you wait.
October timing
Mid-October is genuinely one of the best times to visit Japan — 15–22°C, low humidity, autumn foliage beginning. Hitachi Seaside Park near Tokyo hits peak colour with its red kochia fields in mid-October, which made it into the itinerary as a day trip on a slower day rather than wasted downtime.
The full budget breakdown
The agent built this before writing a single day of the itinerary. That’s the right order — it means every activity it recommends is already calibrated to what’s realistic. And frankly, it’s the only way to have a real conversation with your wife about a trip without someone doing mental maths at the dinner table.
The budget estomation table form the report
JR Pass note: The 7-day JR Pass costs ~₹50,000 per person (₹1,50,000 for three). Tokyo–Kyoto–Tokyo individual shinkansen tickets for three people run ₹42,000–₹52,000 total. For this specific route, the Pass loses money. Always calculate your exact routes before buying — the agent flagged this unprompted.
Onitsuka Tiger note: The ₹47,000–₹52,000 figure is for the custom NIPPON MADE Mexico 66 Deluxe — handcrafted at their Tottori factory, personalised in-store at the Omotesando flagship, the only place in the world this service exists. This isn’t a sneaker purchase. It’s the point of the trip for me. Worth every rupee, budgeted accordingly.
Exchange rate used: approx ₹0.55 per ¥1 JPY.
A sample day — this is what the agent actually produced
I changed almost nothing in what follows. This is the agent’s Day 3 output, lightly formatted for readability. What struck me reading it back was that it remembered details I’d put in the CLAUDE.md two sessions earlier — the Instagram brief, the anime priority, the specific store — without me mentioning any of them in the Day 3 prompt.
That’s not magic. That’s just good system design. But it felt like magic the first time.
**Day 3 — Harajuku · Omotesando · Shibuya****Morning — Meiji Shrine & Harajuku (8am start)**Arrive at Meiji Shrine by 8am. The forested path to the shrine is one of Tokyo’s most photogenic walks — wide gravel avenue, towering cedar trees, morning light filtering through. Near-empty at this hour.Wife’s shot: wear a light-coloured outfit. The contrast against the dark cedar trees is striking and the light is perfect before 9am.Walk south into Harajuku’s Takeshita Street. Multiple Jump Shop branches carry Demon Slayer merchandise not available in India — figures, limited prints, the full range. Budget ₹2,000–₹4,000 for your son here.Afternoon — OmotesandoWalk to Onitsuka Tiger Omotesando NIPPON MADE (4–24–14 Jingumae, Shibuya-ku). This is the only store in the world offering the custom Mexico 66 Deluxe experience — you choose colours, materials, detailing. Price: ¥40,000 (~₹22,000) + tax. International shipping available; delivery to India in 6–7 weeks. Book the customisation slot in advance — the consultation takes 45–60 minutes and fills up.The Omotesando boulevard itself is a photography walk: tree-lined, boutique-fronted, café terraces at street level.Wife’s shot: the boulevard at 3–4pm catches late afternoon light through the zelkova trees. Strong natural framing.Evening — ShibuyaOne train stop or a 20-minute walk. Shibuya Sky observation deck: aerial shot of the crossing at sunset — book tickets ahead, ~₹1,800 per person. Then street level for the crossing itself. Dinner at Ichiran Ramen — the solo booth system that your son will find genuinely fascinating (it’s essentially an anime cafeteria).Son’s anime highlight: Jump Shop Harajuku carries the full Demon Slayer range, including items not stocked outside Japan. This is the stop.Day 3 budget: ₹9,000–₹14,000 per person (Onitsuka Tiger order separate)
Iterating — where agents actually earn their keep
After the first full draft, I sent three follow-up prompts over about 20 minutes:
Day 5 is too packed for a 10-year-old by mid-trip.
Move one activity to Day 6 and add a half-day at
teamLab Borderless instead.
Which day is Fushimi Inari, and what time for the
best photos without the crowds?
Add a "reuse this setup" note at the top of the file
for the next time I use this for a different trip.
Claude updated the same japan-trip.md — not a new file, the existing one, edited in place. It shifted the activity, added teamLab with booking instructions (3–4 weeks ahead, ~₹2,200 per person), answered the Fushimi Inari question (arrive by 6:30am, free entry — the half-hour before tour buses arrive is the shot), and added the reuse note at the top.
Two more rounds and the plan felt completely ours.
This is the part that surprised me most as a developer. I’m used to systems that do what you spec. This felt more like working with someone who understood what I was trying to achieve and kept that goal in mind across sessions. The difference between executing instructions and actually helping.
Five mistakes that’ll waste your time
1. Skipping CLAUDE.md
Without it, the agent has no memory of who you are or what matters. You’ll repeat context in every prompt and get generic outputs every time. Write it once. The agent uses it automatically in every session.
2. Not using plan mode
If you let the agent act immediately, you can’t correct course before the work is done wrong. Plan mode costs you 30 seconds. Skipping it costs you a restart. For developers: think of it as reviewing a PR before it merges, not after.
3. Writing a vague prompt
“Plan a Japan trip” produces a Japan trip. The agent is only as specific as your brief. Constraints aren’t limitations — they’re what make the output worth using. My son’s Demon Slayer obsession, my wife’s Instagram requirement, the specific Onitsuka Tiger store, the visa concern: every detail showed up in the output without me repeating it, because it was in the CLAUDE.md. Garbage in, generic out.
4. Treating the first output as final
The first draft is not the deliverable. Every genuinely useful output I’ve gotten from an agent came after two or three rounds of follow-up. Don’t think of iteration as fixing errors. Think of it as the workflow itself. Developers already know this — it’s called a code review, not a deployment.
5. Not naming the output file
If you don’t specify a filename, output lands in the chat and vanishes. “Save to japan-trip.md” gives you a single document you can update, share, and — in this case — submit directly as the detailed itinerary Japan requires for the visa application. Planning and paperwork, same file.
How to reuse this for any other project
The Japan setup isn’t one-time. The pattern — project folder, CLAUDE.md, plan mode, iterate — works for any multi-part task. Swap the folder and the context file; you have a new agent for a new problem.
The visa checklist and budget table from this trip are now a template section at the top of my CLAUDE.md. Next international trip, I start there. The agent already knows to check these categories for every destination.
Here are four agents worth building next — each one with the specific thing it catches that a chatbot wouldn’t.
Job application tracker
You’re applying to ten companies. Each role needs a tailored CV, a cover letter in your voice, and follow-up tracking. The CLAUDE.md holds your experience, target roles, tone, and the companies you’ve already approached.
What it catches that a chatbot won’t: “You’ve applied to three companies that share the same parent group — they may share candidate data. Worth knowing before you reference one in the cover letter for another.”
Weekly newsletter agent
You publish a newsletter on a topic you follow closely. Every week you spend two hours pulling together what matters. The CLAUDE.md holds your audience, format, and writing style — enough that the output sounds like you. The agent researches, prioritises, drafts, and saves a file ready to edit. Your job is the 20-minute pass, not the two-hour build.
What it catches that a chatbot won’t: “Three of your last five editions covered the same sub-topic — your readers may be noticing the pattern. Here’s what you haven’t touched recently.”
Meeting prep agent
You have a call with a potential investor, a new client, or a partner you haven’t met before. The CLAUDE.md holds your company, your role, and what you’re trying to achieve in the meeting. Give the agent the other person’s name or company — it researches their background, recent news, shared context, and potential friction points, and outputs a one-pager before the call.
What it catches that a chatbot won’t: “They announced a pivot away from this market six weeks ago. You may want to address that directly rather than leading with the assumption they’re still in it.”
Content repurposing agent
You publish a long article or give a talk. Turning it into LinkedIn posts, a tweet thread, and a short summary takes an hour you don’t have. The CLAUDE.md holds your platform preferences, tone, and what you’ve already posted recently.
What it catches that a chatbot won’t: “The strongest line in the whole piece is buried in paragraph eight. This should be the hook, not the current opening.”
In each case: one folder, one CLAUDE.md, plan mode, iterate. The problem changes. The pattern doesn’t.
The bigger point
The Japan trip is happening. The visa application is in. The Onitsuka Tiger appointment is booked for Day 3. My son has already allocated his Akihabara budget to the decimal — in a separate spreadsheet, which honestly impressed me.
The agent didn’t just plan the trip faster. It found the things I would have found out at the wrong moment — standing at the immigration counter in Narita, or standing in the Onitsuka Tiger store realising the shoes I came for ship in six weeks, not today.
I build systems for a living. I know how to automate things. What I hadn’t done until this trip was treat my own planning the same way I treat a complex project at work — with a clear brief, a defined output format, and something working the problem while I handle the decisions.
That’s the real shift. Not faster planning. Not better checklists.
An agent thinks the problem through the way a good colleague would — and keeps you in the loop while it does. Once you’ve worked that way once, going back to the WhatsApp group feels genuinely absurd.
Software architect based in Pune. Writing about AI, building with Claude, and making complex systems actually useful. Connect on LinkedIn &#xNAN;· GitHub
Comments
Loading comments…