Tools for planetary science and language learning, built with Claude Code.
Founder: Shinichi Matsushita
Kasanegi — 日本の独立プロダクトスタジオ。
惑星科学と語学学習の道具を、Claude Code と作る。
How we build 作り方
Every product here is built the same way: Claude Code designs, writes the core, reviews the work against primary sources, and keeps the records. The founder decides; the models do the engineering. Other models take mechanical tasks.
設計・核心の実装・一次資料に当てた検収・記録を Claude Code が担い、創業者が判断します。機械的な作業は他のモデルも使います。
Products 製品
01
Impact-Structure Screening — research tooling, in development (no candidate output yet)
Why it matters
Japan has no confirmed impact structure yet — only one probable crater, Oikeyama in Nagano (Sakamoto et al., 2010). This tooling is built to find the first confirmed ones from public surveys alone, with one processing applied to the whole country.
Research tooling for finding undiscovered meteorite impact structures in Japan. It applies the same terrain, gravity and magnetic processing that can be run across all of Japan to known impact structures worldwide, to test whether known structures are picked up without per-site tuning, before the processing is run over Japan's public observations.
Done so far
Same red relief image map processing applied to public 1 m-class terrain models of 48 known structures (US 28, Canada 13, Estonia 5, Germany 2), out of a 354-row inventory.
At Oikeyama (Nagano, Japan), the in-house processing matched the nationwide Q-chizu rendering within 1 level per RGB channel on the central tile.
Quality of Japan-wide public observations pinned down from primary sources, e.g. national aeromagnetic main line spacing 3–4 km; 7 public gravity stations within 10 km of Oikeyama. 262 commits since 2026-09-20.
Built with Claude
Claude Code acts as the supervisor: writes the task briefs for ChatGPT Pro and Codex, checks every returned result against primary sources (papers, survey reports, distributor pages), and records the verified premises in git.
Claude Code designed the checks that decide what data exist, e.g. confirming that none of 36 Australian structures has a 1 m product in the public ELVIS portal, against a control site where 1 m products do appear.
Source: public 1 m-class DTMs (US, Canada, Estonia, Germany); in-house red relief image map (Chiba et al. 2007 definition, same constants as Q-chizu).Source: GSI Basic Map Information DEM1A elevation tiles (z17); same in-house processing; about 1.8 km square around Oikeyama.
日本には確認済みの衝突痕がまだ無い——長野の御池山が probable とされているだけ(Sakamoto et al., 2010)。この道具は、公開観測だけから、一つの処理を国土全体に当てて、最初の確認済み衝突痕を見つけるために作っている。
Alignment runs entirely in your browser — nothing is uploaded.
A mobile reader for language learners. Open the two language editions of the same manga you already read; the app aligns the pages on your device and lets you look at the other language through a movable lens over the page. Everything stays on the device — no images are stored on our servers. Android first, then iPhone and PC.
Done so far
Page-alignment core (Python; feature matching with MAGSAC and geometric verification) written by Claude: on a real volume, 33 page pairs matched with zero mismatches (Oct 5, 2026).
Android app shell on a Mihon fork: 134 Kotlin files, 22,000+ lines, with page matching and lens display running on an emulator.
PC prototype (Electron) running; HQ at 270 commits since Aug 29, 2026.
Built with Claude
Claude Code designed and wrote the alignment core and the page-matching logic that does not assume page order.
Claude Code supervises the whole build: specifications, reviews of code from other models, bookkeeping and git.
An experiment in AI-run production — two sports video channels (live; one monetized)
We run two YouTube channels as a live experiment: can a production line built and operated by Claude Code reach real audiences and revenue with a single human reviewer? It did. The line detects a game's end, assembles the facts and fan commentary, writes and narrates, and uploads as private; a person reviews and publishes. The experiment funds the studio, and what we learn about running AI operations feeds the other products.
Two channels: 158 public videos, 1.9M+ lifetime views, about 2,500 subscribers (Oct 2026).
Built with Claude
Claude Code designed the soccer line end to end and wrote the baseball line's rulebook and machine checks; it supervises daily operations.
Claude Code がサッカーのラインを端から端まで設計し、野球のラインの規則と機械検品を書いた。日々の運用を監督している。
Built with Claude Code Claude Code が担う所
Across all products, Claude Code (Fable and Opus models) is the supervisor and core engineer: specifications, core implementation, review against primary sources, and the git ledger.
Impact screening: task briefs for research models and verification of every returned result against primary sources.
Manga reader: the page-alignment core and order-free page matching.
Video lines: the soccer line end to end; the baseball line's rulebook, machine checks and operations.
Other models (Codex, ChatGPT Pro) take mechanical and long-running research tasks under Claude's supervision.
全製品で、Claude Code(Fable と Opus のモデル)が監督と中核のエンジニアを務める。仕様・核心の実装・一次資料に当てた検収・git の記帳。
