Forem: QuoLu

Solving Complex Logic with Claude and Research Papers

QuoLu — Tue, 26 May 2026 00:53:17 +0000

Introduction

When building apps with AI-assisted coding, you get to decide "what to build." You can set the design and the policy yourself. However, you occasionally hit a wall.

There are things I want to do, but I lack the background knowledge in logic to realize them.

I am not a programmer, nor am I an expert in signal processing. This is a story about how Claude became a reliable partner when I faced such situations.

The Problem: Speech Gender Identification Isn't Working Well

I wanted to add a feature to LiveTR (a real-time voice translation app) that switches the synthesized voice based on the speaker's gender. Male-like for male voices, and female-like for female voices.

The first method that comes to mind is identifying gender via fundamental frequency (pitch). Men are low, women are high. Simple.

I tried it. It works okay for casual conversation.

However, when I tried streaming an F1 race, the app would identify the voice as female whenever the commentator got excited. Since the pitch rises when the race gets heated, the voice is incorrectly identified as "female" even though it is a man's voice. This happened frequently.

Pitch alone is not enough. But then, what else should I look at? I didn't know.

I Asked Claude

"When it comes to speech gender identification, using only pitch leads to misidentification during moments of excitement. What are some academic methods for this?"

I had Claude investigate. By having Claude use Research mode (a feature where Claude autonomously searches the web to investigate) to look for academic papers and patents, several methods I could never have reached on my own emerged.

It turns out that by combining multiple indicators—not just pitch, but also formants (resonance frequencies of the vocal tract) and MFCCs (Mel-frequency cepstral coefficients)—one can achieve stable identification even during states of excitement.

If you asked me whether I understood the entire content of the papers, honestly, that's doubtful. However, because Claude explained, "This method works on this principle and has these characteristics," I was able to establish a direction. From there, I decided, "I'll go with this combination," and refined the structure.

Straight to Implementation

Once the policy was decided, I worked with Claude to assemble it.

When I told Claude, "Based on the method in this paper, please implement it with this structure," it wrote the code for me. I ran it, checked the results, and adjusted it if something felt off. It's the usual cycle.

When I tested it with the F1 broadcast, it started identifying the voice as male even when the commentator was excited. The stability was completely different compared to when I was using only pitch.

I Think This Is the Strength of AI Coding

Having the AI write code has become a matter of course. However, being able to pull knowledge from fields you are unfamiliar with and translate it into implementation is a different kind of value.

I don't have the capacity to search for and read signal processing papers on my own. But if I tell Claude, "I have this problem," it can investigate relevant research and turn it into working code.

Of course, I don't trust what it produces blindly. I run it, test it, and if it's off, I rethink the approach. That part hasn't changed. But it's a huge help that you don't start from zero in terms of knowledge.

One caution: even if you get good results, if you do business using logic built by referencing patents, you might infringe on patent rights. Even with paper-based logic, patents related to those methods can exist. When using it commercially, you need to check the rights. You can ask Claude, "Are there any patents related to this method?" to have it investigate. However, Claude's investigation isn't necessarily perfect, so you must make the final judgment yourself.

I'm not the person writing the code. My job is to design, decide on policies, and make judgments. Claude pulls in knowledge and turns it into working code. I think this is an example where this division of roles fit perfectly.

LiveTR: A Real-Time English-to-Japanese Audio Translation App for Videos

QuoLu — Mon, 25 May 2026 00:56:03 +0000

What is LiveTR?

It is a Windows application that recognizes English audio from videos playing on your PC in real-time and translates it into Japanese. The translation results are displayed as an overlay on the screen as subtitles, and the app can also read the translations aloud in Japanese.

Whether it is YouTube, Twitch, or a local video file, the source does not matter as long as there is English audio playing.

Key Features

Real-time Speech Recognition — Transcribes English audio on the fly using faster-whisper
Japanese Translation — Supports online translation services (Google Cloud, DeepL, Azure, Amazon)
Subtitle Overlay — Displays translation results in a transparent window. Position and size are adjustable, and the window allows click-through
Japanese Text-to-Speech — Reads translation results aloud using AivisSpeech. Automatically reflects the voice quality of the speaker
Automatic Ducking — Automatically lowers the video volume during playback to make the speech easier to hear
Process-based Audio Capture — Captures audio only from the specified application. Prevents loops by avoiding the re-capture of the synthesized speech

How to Use

Launch the application
Select the process you want to capture audio from
Click "Start" to begin speech recognition, translation, subtitle display, and text-to-speech

System Requirements

Windows 10 / 11 (64bit)
NVIDIA GPU (compatible with CUDA 12.x)
16GB RAM or more recommended

A GPU is required. Since the speech recognition model runs in real-time, decent hardware specifications are necessary.

Development Story

I built this entirely using Claude Code. The development period was about 4 days.

Since my previous project, OLTranslator, was an application that translates text on the screen, I started this project thinking, "Why not do the same for audio?" OLTranslator took me two weeks using Copilot, so this was significantly faster by comparison. Of course, I have become more accustomed to AI coding, but a major factor was being able to maintain project guidelines through CLAUDE.md, and the fact that the design-instruction-review cycle flows naturally with Claude Code.

Points I Focused On

I paid close attention to how to cut the sentences picked up by speech recognition before sending them for translation. Deciding where to segment a sentence and how to join parts if they are cut mid-stream directly impacts translation accuracy.

I also focused on speaker gender identification. I wanted the text-to-speech voice to match the speaker, so I built the logic with Claude Code by referring to academic papers and patents. Since AivisSpeech has multiple speaker models, it reads with a male-like voice for men and a female-like voice for women.

However, this gender identification was quite tricky. If judged solely by pitch, a man might be identified as a woman during exciting moments—like F1 broadcasting—when the pitch rises. I wrote about how I solved this issue in detail in "Logic Impossible on My Own: Building with Claude and Research Papers".

Download

LiveTR — Real-time Audio Translation App (BOOTH)

OLTranslator — A translation app for on-screen text. OLTranslator handles "text," while LiveTR handles "audio." Although they both perform translation, the approaches are completely different.
Logic Impossible on My Own: Building with Claude and Research Papers — A technical deep-dive into the gender identification logic.

OLTranslator: A real-time screen translation app for Windows

QuoLu — Sun, 24 May 2026 00:54:49 +0000

What is OLTranslator?

It is a Windows application that uses OCR to read foreign text on your screen and overlays Japanese translations in real-time.

Simply drag to select the area you want to translate, and the app automatically repeats the OCR and translation process. It works with anything that displays foreign languages on your screen, such as games, live streams, or documents.

Development took about two weeks using VS Code and GitHub Copilot. I followed a style where I determined the design and direction, and left the implementation to the AI.

Key Features

Local OCR — Recognizes text on the screen without requiring an internet connection.
Included Local Translation — Ready to translate right after installation. No API key configuration is required.
Online Translation Support — Compatible with Google Cloud Translation, DeepL, Azure Translator, and Amazon Translate.
Overlay Display — Translation results are displayed overlaid on the original text positions. Clicks pass through, so it won't interfere with your operations.
English and Korean Support — You can switch the source language.
Resident Application — Stays in the system tray and remembers your selected regions. Automatically resumes the next time you launch it.

How to Use

Launch the application.
Drag to select the area of the screen you want to translate.
OCR and translation will start automatically, and the results will be displayed as an overlay.

System Requirements

Windows 10 / 11 (64-bit)
x64 processor (AVX2 support recommended)

Challenges in Development

The most difficult part was handling the text picked up by OCR.

Characters on the screen are recognized piece by piece line by line, meaning a single sentence is often split into multiple blocks. If you don't connect these correctly, the translation becomes garbled. Conversely, if you connect them too aggressively, different sentences get mixed up.

For example, separate texts lined up side-by-side in a game's UI might be combined into a single sentence, or a single sentence might be split into three lines, translated separately, and become meaningless. Judging purely by coordinate proximity leads to incorrect merging, while making it stricter leads to fragmentation. I spent a significant amount of time adjusting these thresholds.

Another challenge was avoiding the translation of content that shouldn't be translated—lines containing only numbers, symbols, or UI labels. Sending such noise to a translation engine wastes API tokens and degrades the results. Especially with online translation, where each request incurs a cost, reducing noise directly impacts both accuracy and cost. Developing filtering rules for what to translate and what to skip was surprisingly tedious work.

Focus Points

I made it easy to try out various translation services. Local translation is included so you can use it right away, and you can switch between Google, DeepL, Azure, and Amazon online translation services simply by entering your API keys. I wanted users to be able to actually compare which service works best for them.

Download

OLTranslator — Real-time Screen Translation App (BOOTH)

Following this, I created an audio version called LiveTR. That one took four days using Claude Code. I wrote about the differences in experience due to development environments in "Copilot → Cursor → Claude Code for VSC. My Journey to the Destination".

What exactly changes with the Claude Max plan?

QuoLu — Sat, 23 May 2026 00:52:21 +0000

Introduction

Everyone knows that the MAX plan lets you "use more."

That's not what I want to know. What actually changes with MAX? When I looked into it, I found that some features often assumed to be "MAX-exclusive" were not, while other differences were not widely known.

I have organized the information while verifying with official sources.

Default model is Opus

With Pro, the default model is Sonnet. With MAX, Opus 4.6 is the default.

It might seem trivial, but it's a big deal. With Pro, there's the hassle of switching to Opus every time, and you tend to compromise by thinking, "Sonnet is probably fine..." out of concern for your usage limits. With MAX, it starts with Opus and you can keep using it that way.

Reference: What is the Max plan?

1M context window at no extra cost

Opus 4.6 supports a context window of up to 1 million tokens (1M). However, it is treated differently depending on the plan.

Plan	1M context for Opus
MAX / Team / Enterprise	Included in subscription
Pro	Requires enabling extra usage (additional charges)

With MAX, Opus is automatically upgraded to 1M context without any configuration. With Pro, to use 1M, you must enable "extra usage" and accept the additional charges.

What changes with 1 million tokens:

You can grasp dependencies between files in large codebases at once
Previous context is less likely to disappear in long sessions
Frequency of /compact is reduced

Reference: Model configuration - Claude Code Docs（Extended context）

Reference: 1M context is now generally available

Priority access to new features — This is surprisingly huge

It is officially stated: New features and models are often provided to MAX first.

Let's list which features were actually MAX-first:

Feature	Description	MAX-First
Remote Control (`/rc`)	Intensive development from smartphone using Claude Code	Feb 2026 onwards. Not yet available for Pro (as of Mar 2026)
Cowork	Delegate tasks to Claude via macOS app	Jan 2026 onwards. Pro later
Dispatch	Send tasks to Cowork from smartphone	Mar 2026 onwards. Pro a few days later
Computer Use	Claude operates the PC directly	Mar 2026 onwards. Pro 2 days later
Memory	Automatically retain conversation memory	Oct 2025 onwards. Pro a few days later

A pattern emerges. Almost all agent-related new features are MAX-first. The gap can be a few days, or in the case of /rc, more than a month and still not available for Pro.

For those who want to experience new features quickly, MAX is essentially in the position of a "paid beta tester." I use /rc quite often, so it's a relief that it's available for MAX first.

Reference: What is the Max plan?

Reference: Cowork Research Preview

Reference: Dispatch and Computer Use

Reference: Release Notes

Features often thought to be MAX-exclusive, but actually are not

These are easy to misunderstand, so let's clarify. The following features can also be used with Pro.

Effort MAX (Maximum thinking depth)

The maximum value of the Effort setting, "MAX," is tied to the model, not the plan. As long as you are in an environment where you can use Opus 4.6, you can select Effort MAX with both Pro and MAX.

However, in reality, Effort MAX consumes a lot of tokens, so you will reach the limit quickly with Pro usage. With MAX, you can run it on MAX constantly without worry. As a feature, it is available on Pro, but practically, it's tough without the usage allowance of MAX.

Reference: Effort - Claude API Docs

Sub-agents

A feature where Claude Code launches sub-agents (clones) to conduct parallel research. This also has no plan restrictions. Anyone in an environment where Claude Code can be used can use it.

Reference: Create custom subagents - Claude Code Docs

/batch

A feature for parallel processing large-scale changes. This is also a built-in skill available to all Claude Code users.

Reference: Built-in commands - Claude Code Docs

Research (Research mode)

A feature where Claude autonomously searches the web for investigations. It is available on all paid plans: Pro, MAX, Team, and Enterprise. The Research feature itself is not enhanced just because you have MAX. However, Research consumes a lot of tokens, so it will strain your usage limit on Pro.

Reference: Using Research on Claude

So what is the real value of MAX?

Looking only at the features, there are surprisingly few MAX-exclusive ones. To summarize:

	Pro	MAX
Default model	Sonnet	Opus
Opus 1M context	Additional charge required	Included in subscription
Priority access to new features	Later	Use first
Effort MAX	Available but hits limit quickly	Run on MAX constantly
Sub-agents	Available but hits limit quickly	Use without worry
Research	Available but strains usage	Use without worry

A pattern is visible. There are many features that can be used with Pro. But with Pro's usage limit, you cannot fully utilize those features.

Effort MAX, sub-agents, and Research all consume tokens the more you use them. Pro creates a dilemma where "the feature exists, but using it hits the limit." MAX removes that limiter.

You might think this just brings us back to "you can use more." But "increasing the amount" and "fully utilizing the feature" are different. The value of MAX lies in the latter.

Summary

When I looked into it, there were fewer MAX-exclusive features than I thought. The real value lies in "using existing features without a limiter."

Opus is the default. 1M context at no extra cost. Experience new features first. And you can use Effort, sub-agents, and Research to your heart's content without worrying about usage limits.

$100 a month is not cheap. But which is more valuable: paying $20 and not being able to use the features fully, or paying $100 and using them fully? For me, it was the latter.

From Copilot to Cursor to Claude Code for VS Code: My Journey to the Optimal Setup

QuoLu — Fri, 22 May 2026 00:53:54 +0000

Introduction

There is a flood of AI coding tools. GitHub Copilot, Cursor, Claude Code, Cline, Continue... There are so many options that it is hard to know what to use.

I wandered through them quite a bit myself.

I eventually settled on VS Code + Claude Code Extension + MAX Plan, but it took a lot of trial and error and money to get here.

For those who are similarly lost, I will write down why I settled on this configuration.

Prerequisite: My Development Style

First of all, I am not the type who grinds away writing code by hand.

I develop applications, but what I do is design and policy decision-making. I leave the code implementation to the AI and review what it produces. I don't care about the language, because the AI writes it for me. It is what you might call an architect-style development style.

Another important point is that I frequently share images. I often show the AI screenshots of UIs or error screens and ask, "What is going on here?" This usage significantly influenced my final choice of tools.

Claude Pro — Actually, this is where I started

It might be surprising, but the first thing I paid for was Claude Pro.

The quality of the conversation was good from the start. It properly engages in design discussions and serves as a sounding board for policies. However, there was a fatal problem.

The limit is reached immediately, and work stops.

When you are concentrating on development, seeing "Please wait a moment" is incredibly stressful. The flow is completely interrupted. As a result, I ended up leaving Claude Pro for a while.

GitHub Copilot — An amusement park where you can use various models

Next, I went to GitHub Copilot. The integration with VS Code is natural, and I realized that I could choose Claude's models in Copilot. Huh? If that's the case, wouldn't it be better to just cancel Claude and use Copilot alone? I thought.

Moreover, Copilot has models other than Claude. I ended up trying out various things here.

Gemini 3.1 — It gets me excited, but

I was honestly excited when I used Gemini 3.1. It is fast and smart.

However, this guy had a problem. It injects code that allows it to edit files on its own outside of VS Code's functionality. Even if I tell it to stop, it does it. It's arrogant. For someone who wants to decide the design themselves, having it act on its own is the biggest stress.

GPT-5.4 — From "Wow!!!" to a realization

GPT-5.4 was also wonderful. It can do anything automatically. I honestly thought, "Wow!!!".

But as I got used to it, I realized something.

Wait, when things get complicated, isn't Claude better at solving bugs?

When the code increases, dependencies become intertwined, and the cause of errors becomes difficult to solve, it was Claude that persistently followed the context and pinpointed the cause accurately.

Well, it stops at the limit immediately, though.

Cursor — When I was paying a lot

Next, I went to Cursor. It is an editor with AI features at the forefront, and at first, I thought, "This is it."

It reads the entire codebase and makes suggestions, and the chat interaction is more natural than Copilot. You can also paste images. I used it as my main tool for a while. I paid a lot of money, too.

However, as I used it more, points of concern started to appear.

When things get a little tricky, I want to use Claude as the model. And I end up using it because I want good results. The additional charges end up being for Claude's Sonnet and Opus.

And around this time, I suddenly realized something.

Claude Code — Recovering every 5 hours, isn't this the best?

When I re-evaluated the existence of Claude Code, I thought:

The usage limit recovers every 5 hours. Doesn't this mean I can essentially use it forever?

I calculated calmly here. When I added up the monthly fees I was paying for Copilot and Cursor, it was already an amount within reach of the MAX plan.

Unifying on MAX — I tried it, and it was the best

So, wouldn't it be better to just unify on MAX?

I tried it, and it was the best.

Even if I run Opus endlessly, it rarely overflows. I am a normal person using it in a normal way, so I have more than enough leeway even with MAX 5x. The "stopping at the limit" of the Pro days seems like a lie.

I cancelled both Copilot and Cursor. Only VS Code + Claude Code extension. Since it became simple, my confusion disappeared, too.

As a result, I reached the conclusion that concentrating on one tool is more cost-effective than scattering payments across multiple tools.

Claude Code for VSC — The reason I settled here

Sharing images is easy

This was the biggest factor for my usage. I can just paste screenshots into the Claude Code panel in VS Code and ask questions. I can show error screens or the UI appearance and say, "Isn't this strange?"

I can also code with the terminal Claude Code CLI, but to share images, I need to pass the file path. With VS Code, it's just drag and drop. For someone like me who wants to show screens rather than code, this difference is significant.

Maintaining project context with CLAUDE.md

When you use Claude Code on VS Code, it automatically reads the CLAUDE.md in the project root. Development policies, code styles, things you must not do. Even if the session changes, you can start from the same context every time.

The story of creating two translation apps

I have an episode where I actually felt the difference in this environment.

The first one is OLTranslator. It is a Windows app that uses OCR to read foreign language text on the screen and overlays the Japanese translation. I developed this with VS Code + Copilot, and it took 2 weeks to complete.

The second one is LiveTR. It is an app that recognizes English audio from a video playing on the PC in real-time, displays Japanese subtitles, and even reads them aloud. I developed this one using only Claude Code, and it was completed in 4 days.

To be honest, I started AI coding in mid-February, and by the time I was working on LiveTR, I was quite used to it. So, I cannot simply say "It was fast because of Claude." However, development with Claude Code was clearly smoother. Conveying the design, leaving it to the AI, and reviewing. This cycle rotated without stress.

Conclusion: Whether it fits my usage

What is important in choosing an AI coding tool is not benchmark scores but whether it fits your development style.

In my case:

Style of designing and giving instructions → Interactive Claude Code fits.
Frequently share images → VS Code panel is convenient.
Don't want to break the existing VS Code environment → Complete with an extension instead of Cursor.
Double-billing for Copilot and Cursor → Unifying on MAX is more cost-effective.

If you are someone who works with your hands and writes hard, Copilot might be the optimal solution. Some people might find value in Cursor's unique features.

However, for people like me who "design, delegate, and review," I believe VS Code + Claude Code Extension + MAX Plan is the optimal solution for now.

Even if I run Opus endlessly, it rarely overflows. I'm just an ordinary person. I think this is the most cost-effective way, isn't it?

I realized I was only using half of what Claude Code has to offer

QuoLu — Thu, 21 May 2026 11:07:43 +0000

Introduction

I have subscribed to Claude MAX and am using it for coding almost every day.

When developing applications, I handle the design myself and leave the implementation to Claude—a so-called "architect-style" usage. I don't really care about the language, because the AI writes it for me.

Then, one day, I realized:

I didn't even know half of the features of Claude Code.

I didn't even know about /init, and I only recently discovered the existence of /rc.

I am writing this article because I suspect there are quite a few people in the same boat.

`/rc` (Remote Control) — Develop thoroughly from your smartphone

This was the most shocking discovery.

While running a Claude Code session on your PC, you can operate the same session from the Claude app on your smartphone. All the code is executed on your local PC, and your smartphone just acts as a "window."

/rc

With just this command, a QR code is displayed, and once you scan it with your phone, the connection is complete.

Why it's great

It's not just for monitoring. You can issue instructions, have it write code, and perform reviews normally from your smartphone. I can do heavy development on my phone while drinking coffee on the sofa. I don't need to be glued to my PC. I can check the progress, stop it if it's heading in a weird direction, and issue the next set of instructions right away.

Points to note

The session disconnects if the network is down for about 10 minutes.
It ends when the terminal is closed.
There is a bug where the remote connection does not recover after hitting a rate limit (as of March 2026).

It is still a research preview, so it has some rough edges, but the direction is perfect.

`/init` — Crucial, but not guided

I want to shout this from the rooftops: It's important. It's incredibly important. And yet, it's not prominently guided anywhere. I was caught off guard.

This is an initialization command to help Claude Code understand the project structure when starting a new project. Whether you do this or not makes a huge difference in Claude's level of understanding.

/init

Whenever you start a new session, run /init first. I want you to make this a habit.

I used it for a while without knowing this. I was getting frustrated every time Claude made an off-target proposal, but it was just because I hadn't run /init. I wish I had done it from the start.

`CLAUDE.md` — Another existence you notice through `/init`

When you run /init, a file named CLAUDE.md is generated. I suspect many people realize its existence for the first time here, just as I did.

This is an instruction sheet that Claude Code reads every time you start a session if you place it in the project root.

In my case, I write my development policies and code style here. Things like "Apply code changes only after diff suggestion and approval" or "Write comments in Japanese."

Without this, you end up repeating the same instructions every time. I believe the /init -> CLAUDE.md flow is the very first step in using Claude Code effectively.

Plan (Plan Mode) — Don't fix code when I just asked a question

Claude is excellent. He is so excellent, in fact, that he sometimes starts fixing code when I only asked a question.

I just asked, "What's wrong with this error?", and when the response came back, the code had already been rewritten. Being perceptive and human-like isn't necessarily bad, but sometimes I just want to ask.

In such cases, it's better to switch to Plan mode. You can simply toggle the mode from the Claude Code panel in VS Code. Claude will then propose each action, and it will only execute them after you approve.

There seems to be a command (/plan) as well, but I don't use it.

Effort (Depth of thought) — Always set to maximum if you have the capacity

"Effort" is a setting that adjusts the depth of Claude's thinking. You can switch it from the Claude Code panel in VS Code.

Simply put, the higher it is, the better the answers you get. It supposedly increases usage, but if you look at your 5-hour or 1-week limit and have room to spare, I think you should just set it to maximum.

By the way, on the MAX plan, the highest setting is upgraded from "high" to "MAX" lol.

There seems to be a command (/effort) as well, but I don't use it.

`/compact` — Savior when context grows bloated

When working for a long time, the accumulated conversation causes the context to bloat. If you feel it's getting heavy:

/compact

You can also specify what to keep.

/compact Keep the error handling patterns

As a rule of thumb, run this when context usage exceeds 80%. If you are switching tasks, /clear is better.

`/fork` and `/rewind` — Experimentation and rewinding

/fork copies the current conversation into a branch. It allows you to say, "Try experimenting with this policy over here" without dirtying the main line.

/rewind rolls back the conversation along with the code changes. "That's wrong, undo it" becomes a one-command task.

/fork     # When you want to experiment
/rewind   # When it messed something up

Model (Model switching) — Opus is enough if you have MAX

In VS Code, you can switch models by clicking on the part where the model name is displayed in the Claude Code panel. You can also do it with the command /model.

I often see articles about dropping down to Sonnet or Haiku to save costs, but if you are on the MAX plan, you can run everything on Opus without any issues. In normal usage, I almost never hit the limits with Opus. Since you are paying for MAX, there is no need to intentionally lower the performance.

Features I haven't mastered yet

Skills — A mechanism to define custom skills and have them called automatically. Still experimenting.
Claude Code Channels — You can operate sessions from Telegram or Discord. I'm interested in it as an alternative to /rc.
/simplify — Apparently, it lets 3 agents review the code before a PR. This is what I will try next.
/batch — Process large-scale changes in parallel and automatically create a PR. I want to use this once the scale gets larger.

Summary

If you are just using Claude Code normally, you haven't even touched half of its features. The official documentation is in English, new features are buried in a single line of patch notes, and it's natural that you wouldn't notice them.

For those like me who "pay for MAX and use it heavily but actually don't know many of the features," I will continue to write based on my real experiences.

I have written about my experience actually building apps with Claude Code in the articles OLTranslator and LiveTR. For the story about how Claude Code provided value beyond "writing code," see "Building from Claude and research papers".

Forem: QuoLu

Solving Complex Logic with Claude and Research Papers

Introduction

The Problem: Speech Gender Identification Isn't Working Well

I Asked Claude

Straight to Implementation

I Think This Is the Strength of AI Coding

LiveTR: A Real-Time English-to-Japanese Audio Translation App for Videos

What is LiveTR?

Key Features

How to Use

System Requirements

Development Story

Points I Focused On

Download

Related Articles

OLTranslator: A real-time screen translation app for Windows

What is OLTranslator?

Key Features

How to Use

System Requirements

Challenges in Development

Focus Points

Download

Related Articles

What exactly changes with the Claude Max plan?

Introduction

Default model is Opus

1M context window at no extra cost

Priority access to new features — This is surprisingly huge

Features often thought to be MAX-exclusive, but actually are not

Effort MAX (Maximum thinking depth)

Sub-agents

/batch

Research (Research mode)

So what is the real value of MAX?

Summary

From Copilot to Cursor to Claude Code for VS Code: My Journey to the Optimal Setup

Introduction

Prerequisite: My Development Style

Claude Pro — Actually, this is where I started

GitHub Copilot — An amusement park where you can use various models

Gemini 3.1 — It gets me excited, but

GPT-5.4 — From "Wow!!!" to a realization

Cursor — When I was paying a lot

Claude Code — Recovering every 5 hours, isn't this the best?

Unifying on MAX — I tried it, and it was the best

Claude Code for VSC — The reason I settled here

Sharing images is easy

Maintaining project context with CLAUDE.md

The story of creating two translation apps

Conclusion: Whether it fits my usage

I realized I was only using half of what Claude Code has to offer

Introduction

/rc (Remote Control) — Develop thoroughly from your smartphone

Why it's great

Points to note

/init — Crucial, but not guided

CLAUDE.md — Another existence you notice through /init

Plan (Plan Mode) — Don't fix code when I just asked a question

Effort (Depth of thought) — Always set to maximum if you have the capacity

/compact — Savior when context grows bloated

/fork and /rewind — Experimentation and rewinding

Model (Model switching) — Opus is enough if you have MAX

Features I haven't mastered yet

Summary

`/rc` (Remote Control) — Develop thoroughly from your smartphone

`/init` — Crucial, but not guided

`CLAUDE.md` — Another existence you notice through `/init`

`/compact` — Savior when context grows bloated

`/fork` and `/rewind` — Experimentation and rewinding