literature-review-toolkit¶
Scaffolding to drive a Claude (or other LLM) agent through a structured literature review — and keep it honest.
The agent does the judgment work: choosing what to search, what matters, how to group it, and (optionally) how to write it up. These scripts handle the API calls, the verification, and the bookkeeping — the parts an LLM is worst at and where a single fabricated author or off-by-one DOI quietly poisons a review.
The core idea: judgment vs. mechanics¶
A literature review has a few points where a human genuinely has to decide something. Everything between those points has a ground truth — a DOI either resolves to the paper you cited or it doesn't — so it should be mechanized and guarded, not left to the agent's memory.
1
Scope¶
You decide the topic and the span of the search (Phase 1) — or, in lab mode, you point it at a lab's publication corpus.
2
Families (optional)¶
You approve or edit the agent's proposed theoretical grouping before it labels every paper (Phase 6b).
3
The write-up (optional)¶
If you want a narrative review, the agent authors the prose — the one judgment step the toolkit does not mechanize (Phase 7).
In between, the mechanical steps run automatically but are guarded:
- A required antecedents pass searches the topic's methodological, empirical, and theoretical roots — the forward search is recency-biased and misses them.
- Every citation is verified against PubMed / PMC / CrossRef / arXiv. Search agents fabricate roughly 1 in 4 — wrong first authors, inverted findings, invented or mis-copied DOIs, even an entirely wrong author list for a real paper.
- Every reference is rebuilt from its verified DOI into canonical APA-7, behind a hard audit gate. No agent-typed or database-typed reference text is trusted.
- Citation counts are fetched, reconciled against a second database, and schema-checked; cross-citation mining and dedup run; the spreadsheet is rebuilt from JSON each time.
The principle
The agent supplies judgment; the scripts supply ground truth. Wherever a fact can be checked, it is checked — automatically, every time, with a gate that fails the build rather than your reader.
What you get out¶
The core deliverable is one .xlsx file — an annotated, verified, fully
formatted bibliography. Two optional deliverables sit on top of it, both rendered
from the same verified data:
:material-table: Spreadsheet¶
The always-on deliverable. One row per paper: canonical APA-7 reference, summary, tag, family, and citation counts — colour-coded by where the paper came from.
:material-chart-timeline-variant: Lineage figure (opt)¶
An interactive HTML figure (plus SVG/PNG/PDF) grouping the corpus into theoretical families on a citation-weighted timeline.
:material-file-document-edit: Review article (opt)¶
An AI-authored narrative review .docx — prose by the agent, references pulled
canonically from the verified corpus, with a mandatory priority audit.
Two ways in¶
Start from a question and search outward.
"Literature review on the anatomical connections between the visual system and the cerebellum — primate or human, any tractography method, back to the 1970s."
Start from a lab's full publication corpus, derive its research themes and how they shifted over time, then search outward to place that work in the field.
"Review the Gallant lab's human-imaging work in the context of the broader field."
Both share the same verify → canonicalize → count → families → figure → review machinery. See Topic mode vs lab mode.
Next steps¶
- Getting started — install, configure, and run your first review.
- The pipeline — the whole flow on one page.
- Phases in detail — every phase, its guardrail, and its real output shown as a figure.
- Examples — two complete worked reviews, plus a gallery of finished lineage figures.
AI disclosure
This toolkit is designed to be driven by an LLM agent, and the optional review articles it produces are AI-authored (with an explicit disclosure note in every document). The verification machinery exists precisely because LLM output cannot be trusted on matters of fact without checking.