문헌 검토에 Claude를 사용하는 방법: Hugo Skill 가이드

It can write the synthesis, but not find the sources. Asked cold, any chatbot will produce fluent review prose with citations that look right and sometimes do not exist, because it is completing a pattern rather than consulting a database. Used properly, you supply the papers and Claude does the part that is actually hard: grouping findings by theme rather than by paper, mapping where studies agree and disagree, and arguing what the field has not settled. The Hugo skill enforces that order so the sources come first.

Because a language model predicts plausible text, and a citation has a very predictable shape: an author surname, a year, a journal title, a DOI. Producing something that looks like a reference is easy; knowing whether that reference exists requires looking it up, which the model is not doing. The only reliable fix is to remove the need to invent one, by working exclusively from sources you paste in and writing [CITATION NEEDED] wherever a claim has no source behind it.

No, and it says so rather than pretending. It has no live database access. What it does instead is design the search with you - Boolean strings, database-specific terms for PubMed, Cochrane, PsycINFO, Web of Science and Scopus, and written inclusion and exclusion criteria - and then work on whatever you bring back. That constraint is the anti-hallucination mechanism, not a missing feature.

$29, one time. No subscription and no institutional licence: you download the file and keep it. The standalone AI Literature Review Generator prompt, which used to be sold separately, is now included with the skill.

A synthesis matrix is a grid with your sources down the rows and your themes across the columns, each cell marked as supporting, contradicting or not addressing the theme. Building it before any prose exposes two things a stack of summaries hides: which themes rest on very thin evidence, and which studies contradict each other and therefore need explaining rather than averaging. It is also far cheaper to restructure a matrix than a written chapter.

With the scaffolding, yes. Hugo covers PRISMA flow diagram planning, PICO and SPIDER search frameworks, data extraction template design, and CASP and Cochrane quality assessment structure. What it cannot do is run the screening for you or vouch for the papers, so the protocol, the searching and the quality judgements remain yours. Treat it as the person who sets up the structure correctly, not as a second reviewer.

That depends on your institution, and the rules vary widely: some departments permit AI assistance with disclosure, some restrict it to specific stages, some prohibit it in submitted work. Check the policy before you start rather than after you have drafted a chapter. Whatever the policy, verifying every citation against the actual paper is on you.

A well-written prompt gets you the source-only rule and the matrix once. The skill keeps them applied across the whole conversation, including follow-ups, and carries what a prompt cannot: the quality-assessment frameworks, the discipline-specific review formats, and the systematic-review protocol structure. If you want to test the behaviour first, the starter prompt in the guide above will show you most of it.