用 Claude Opus 5.5 在 2-3 轮提示词内做出 Apple Silicon 本地漫画翻译 App

I have been an avid Manga Reader for the last 6-7 years. I have had some mangas that no one was translating in English or didn't have an official release in English but there were some existing translations in Spanish, Portugese, Korean and Mandarin. Learning Japanese or any of the other languages from scratch to read manga was just out of my bandwidth.

So, I decided to spin up a Claude Code Session to build me a Native App for Apple Silicon with Some Hard limits on memory usage as well as some detailed requirements as follows:

Prompt-1:
Create a plan to create a small lightweight app to screen manga images and OCR the text in any language (Japanese, Chinese/Mandarin, Korean, Spanish, French or Even Portugese) and translate them to English in a certain given font intelligently, everything needs to happen on local device, the app needs to be limited to 8GB RAM (RAM usage needs to be under 200-250Mb Peak at any given time) but 2-5 GB of system storage. The app needs to be very performant to allow for scrolling through 2000+ Images at a time without any lag. I don't expect the translation and OCR pipeline to be super fast, but it should be able to translate any page and should allow for the text to be edited and formatted and also allow erasure/retouching tool. All Translation OCR needs to happen on device within the given memory bandwidth.

Some Refine Prompts that I provided to Fix some of the issues:

1. The Tooltip needs to be for individual Buttons, the Side Panel has 2 buttons for toggle it seems as well as the left panel seems to be floating out of the window (looks separate from the current main window) 2. There is no Lasso Tool to Create selection for translation (in case some text is unidentified by the app) 3. Also there is an inconsistency that pages that get translated are not marked so, only after some toggling around they get updated. (As Visible in the selected 4th Picture) 4. Additionaly newly added pictures to the directory to the folder don't reflect in the project unless and until the project is reloaded. 5. There needs to be some Error Handling system to display that translation has failed in some slide or something didn't happen as expected. (Having an Error Panel Like VS Code at the bottom that can be collapsed would be a great addition, it would let users retry any of the operations) 6. The Typeset Rules aren't useful can you remove that ? 7. The Erasure/Retouch Tool's UI is a bit off on the left side padding 9. The Pattern Erasure/Retouching lacks a lot with Duotone patterns causing weird artifacts 

And Voila! Claude fetched a list of Repository on github for Manga Translations written in Rust (Special Mention to Koharu App) and Hugging Face Models and Apple CoreML Libraries to create the first Outlined Plan for the first Iteration of the App.

After a couple more refining sessions with Claude I got something seriously usable App for my personal use and it works perfectly fine for my use case. Now I don't have to rely on external translations or wait for them to be released. The App runs a pipeline to cleanup, refine, translate and properly typeset the manga.

I had tried attempting the same task with Claude opus 4.6 & 4.8 Earlier but It just used to eat up my daily session limits and it was just not sustainable for me to leave everything upto claude to build something that I couldn't test. with the release of Claude opus 5.5 I have noticed that the session limit were improved and the quality of app that Claude could generate within a single session has been phenomenal. It is just impressive and mind boggling for me to comprehend that an entire app can be made with such ease if the use case well documented despite having limited documentations.

PS: This App will only work on MacOS >26 & Apple Silicon Devices and not for Multi-platform use.
Link to the Github Repo:
https://github.com/subh0x/MangaTL

Feel free to Fork and use it.

TLDR: Claude Opus 5.5 is phenomenal, I got a pretty polished app that can translate Mangas on my device locally without the need for any API Calls in nearly 2-3 planned session prompts.