A Claude skill that digs for primary sources instead of summarizing
imozuru borrows decades of library-science research on information seeking to make Claude Code chase citations, authors, and links until it lands on original texts — then refuses to summarize them.
What it does
imozuru is a Claude Code skill for researching unfamiliar fields. Instead of the usual “read the top search hits and summarize” behavior, it first builds a map of the field’s terminology, key people, and gathering places, then follows citations, links, and authors until it reaches primary sources. Those sources are saved verbatim under sources/, and the final report points you to the exact file and line for each claim you need — the reading and judging is left to you.
The interesting bit
Nearly every design decision is grounded in information-seeking research: it refuses to summarize because Kuhlthau (1991) warns that early summarization causes premature closure; it stops when either every claim has a source (HALT) or two independent search agents overlap by ~70% (a capture-recapture style saturation estimate); and it runs two search agents with different hypotheses to avoid the single-direction tunnel vision HypoSearch identifies. There’s even a report.py that verifies every extracted quote exists verbatim at the cited line — one word off and the report won’t build.
Key highlights
- Handles fields with no papers (e.g., vocal mixing, illustration techniques) by tracing blog posts, “I learned this from so-and-so” attributions, and video subtitles via
yt-dlp - Tags every source as 〔論文〕/〔本人〕/〔観察〕/〔まとめ〕/〔公式〕 so you judge credibility yourself
- A watch mode monitors discovered venues (Qiita/Zenn tags, note, YouTube, arXiv) via RSS
- Python 3.10+ standard library only; optional tools like
pdftotext,tesseract(with Japanese vertical-text support), andaximprove extraction when present - Citation chaining via
chain.pyuses the OpenAlex API, no key required
Caveats
- Search queries go out to the web, so it’s unsuitable for anything confidential
- It won’t bypass paywalled papers — unreachable sources are reported as unreachable, full stop
- Fetching refuses non-http(s) and private/LAN addresses (an anti-SSRF guard), which can break entirely behind some VPN/proxy setups unless you explicitly override it
Verdict Worth a look if you research unfamiliar domains with Claude Code and are tired of summaries that drop the details you actually needed to decide. If you just want a quick answer, this deliberately isn’t that tool.
Frequently asked
- What is MaryCache/imozuru?
- imozuru borrows decades of library-science research on information seeking to make Claude Code chase citations, authors, and links until it lands on original texts — then refuses to summarize them.
- Is imozuru open source?
- Yes — MaryCache/imozuru is open source, released under the MIT license.
- What language is imozuru written in?
- MaryCache/imozuru is primarily written in Python.
- How popular is imozuru?
- MaryCache/imozuru has 1k stars on GitHub.
- Where can I find imozuru?
- MaryCache/imozuru is on GitHub at https://github.com/MaryCache/imozuru.