MaryCache/imozuru

What it does

imozuru is a Claude‑powered “skill” that helps you conduct deep, source‑first research on an unfamiliar topic. Instead of asking an LLM to summarise a few top‑ranked web pages, it builds a research map of key terminology, influential people, and hubs of knowledge, then follows citations, links, and authors until it collects the original documents (papers, blog posts, videos, etc.). The original texts are saved unchanged; the tool only points you to the exact file‑and‑line where evidence for each claim lives, leaving the final judgement to you.

How it works

1. Prompt – You give Claude a natural‑language request, e.g. “芋づる式に調べて: Gaussian Splattingって何か”.
2. Map creation – Two parallel search agents crawl the web, gathering vocabularies, experts, and venues. Their findings are written to map.md.
3. Citation chasing – Using the OpenAlex API (no key required) the agents follow references, cited‑by links, and author pages to fetch raw sources into sources/.
4. Stopping criteria – The process stops when every desired claim is backed by a source (HALT condition) or when 70 % of newly found material overlaps with the existing map (saturation).
5. Extraction – For each claim the exact snippet from the source is extracted; report.py verifies that the snippet matches the line numbers in the saved file.
6. Output – A directory structure is created under ~/research/<topic>/ containing:
   • map.md – the research map and progress log
   • log.md – week‑by‑week notes of what was tried and why it succeeded/failed
   • sources/ – raw PDFs, HTML‑converted markdown, YouTube subtitles, etc.
   • extracts.json and report.html – the “抜き書き” (extract) page that lists each claim with its supporting excerpt.

Why it matters

• Source‑first rigor – By refusing to summarise early, it avoids “premature closure” (Kuhlthau 1991) and ensures you see the original evidence before forming conclusions.
• Hybrid scholarly‑plus‑grey‑literature search – Works equally for academic fields (arXiv, PDFs) and for practitioner‑heavy domains (Qiita, Zenn, YouTube), where the best knowledge lives in blogs and videos.
• Transparent provenance – Every claim is tied to a concrete file and line number, making the research audit‑able and reproducible.
• Modular, open‑source – Implemented in pure Python 3.10+ with only optional external tools (pdftotext, yt‑dlp, ax, etc.). It can be added to Claude via the Vercel Skills CLI.

Who should use it

• Researchers, engineers, or hobbyists who need deep, verifiable background on a topic they know little about.
• Anyone uncomfortable with LLM‑generated summaries that may omit critical nuance.
• Teams that want a reproducible “research notebook” that records exactly which primary sources support each argument.

Getting started

# Install as a Claude skill
npx skills add MaryCache/imozuru
# Or copy manually
cp -r skills/imozuru/ ~/.claude/skills/imozuru/

Make sure you have Python 3.10+, and optionally install:

• pdftotext (poppler or Xpdf) for PDF extraction
• yt-dlp for YouTube subtitles
• ax for HTML‑to‑Markdown conversion
• tesseract + pdftoppm for OCR on scanned PDFs

Then in Claude Code simply ask:

芋づる式に調べて: Gaussian Splattingってよく聞くけど全然わからない。論文とか解説の流れを掘ってほしい

The skill will create a folder like ~/research/GaussianSplatting/ with the map, sources, and an HTML report.

---

License: MIT – you can freely modify and redistribute the skill (but respect copyright of the harvested source documents).