What is this?
imozuru is a Claude Code skill that digs into an unfamiliar topic by tracing citations, links, and authors back to original sources, instead of just summarizing top search results.
imozuru is a Claude Code skill that digs into an unfamiliar topic by tracing citations, links, and authors back to original sources, instead of just summarizing top search results.Is this a standalone app or something that plugs into Claude Code?It is a Claude Code skill: a folder of scripts and instructions you drop into Claude Code's skills directory, not a separate application you launch on its own.
It is a Claude Code skill: a folder of scripts and instructions you drop into Claude Code's skills directory, not a separate application you launch on its own.What tools does it need to work fully?The core scripts only need Python 3.10+ and its standard library, but it also uses pdftotext/pdftoppm with tesseract to read and OCR PDFs, yt-dlp for YouTube subtitles, and optionally ax for cleaner HTML-to-markdown conversion; the citation-chasing script chain.py calls the OpenAlex API, which needs no key.
The core scripts only need Python 3.10+ and its standard library, but it also uses pdftotext/pdftoppm with tesseract to read and OCR PDFs, yt-dlp for YouTube subtitles, and optionally ax for cleaner HTML-to-markdown conversion; the citation-chasing script chain.py calls the OpenAlex API, which needs no key.What happens if a source is paywalled or a field has no papers to trace?It never bypasses paywalls; anything it can't legitimately fetch is reported as unreachable rather than obtained another way, and in fields without papers it instead treats blog posts and videos as the original sources to trace.
It never bypasses paywalls; anything it can't legitimately fetch is reported as unreachable rather than obtained another way, and in fields without papers it instead treats blog posts and videos as the original sources to trace.Does it keep my research questions private?No — its search sub-agents send the content of your question out to the web as search terms, so the README warns against using it for confidential or personal research.
No — its search sub-agents send the content of your question out to the web as search terms, so the README warns against using it for confidential or personal research.Why doesn't it just summarize the sources it finds?It follows Kuhlthau's (1991) finding that summarizing too early causes premature closure and cuts exploration short, so imozuru only points to where in the original text a claim comes from instead of condensing it.
It follows Kuhlthau's (1991) finding that summarizing too early causes premature closure and cuts exploration short, so imozuru only points to where in the original text a claim comes from instead of condensing it.How does it decide when to stop digging for sources?It stops once every claim has a verified citation to an original source, or once 70%+ of the sources its two search agents find are already on its map — a saturation estimate borrowed from overlapping independent searches (Kastner 2007).
It stops once every claim has a verified citation to an original source, or once 70%+ of the sources its two search agents find are already on its map — a saturation estimate borrowed from overlapping independent searches (Kastner 2007).How does it avoid chasing one wrong search direction?It runs two search sub-agents with different hypotheses in parallel and compares their results before continuing, based on research (HypoSearch, Zhou et al. 2026) showing single-direction AI search tends to lock onto an early guess.
It runs two search sub-agents with different hypotheses in parallel and compares their results before continuing, based on research (HypoSearch, Zhou et al. 2026) showing single-direction AI search tends to lock onto an early guess.Where do the results end up?Output is saved under ~/research/, including a map.md of terms/people/venues, a sources/ folder of original texts, and extraction pages that pair each claim with a verbatim, line-checked quote; a sample run is in examples/demo/.
Output is saved under ~/research/, including a map.md of terms/people/venues, a sources/ folder of original texts, and extraction pages that pair each claim with a verbatim, line-checked quote; a sample run is in examples/demo/.What are similar projects?feynman, LLMSearchRecommender, deep-research-skill, cy-carousel-skill, explainer, Kami, qiaomu-anything-to-notebooklm, OmniScientist
feynman, LLMSearchRecommender, deep-research-skill, cy-carousel-skill, explainer, Kami, qiaomu-anything-to-notebooklm, OmniScientistWhich topics is it in?information-retrieval, deep-research, japanese, rss, arxiv, claude-code-skill, ocr, agent-skill, web-scraping, academic-research, skill, pdf
information-retrieval, deep-research, japanese, rss, arxiv, claude-code-skill, ocr, agent-skill, web-scraping, academic-research, skill, pdf