Resources
Where to point a model at the literature, what the prompts here are drawn from, and a long worked example.
Publications
- Dewar JM (2026)Zenodo
Large Language Models for Science and Research: A Practical Guide
doi.org/10.5281/zenodo.19177102 - Dewar JM & Venkatesh (2026)EdArXiv
- Dewar JM (2026)Nature
- Conwell et al. (2026)Nature Structural & Molecular Biology
Connecting sources — remote MCP servers
Each of these is a hosted endpoint you add once to Claude, ChatGPT, or any MCP-capable client, after which the model queries the literature directly rather than recalling it. Endpoints checked 2026-08-25. Most need their own account; Scholar Gateway is institutional.
- Scitedocs
https://api.scite.ai/mcpSmart Citations — retrieves the sentences that cite a paper, labelled supporting, contrasting, or mentioning, plus retraction and correction notices.
- Consensusdocs
https://mcp.consensus.app/mcpQuestion-driven search across ~200M papers, returning per-claim evidence with study design and sample size attached.
- Elicitdocs
https://elicit.com/api/mcpSystematic-review workflows: screening, data extraction into columns, and evidence tables over a defined paper set.
- Scholar Gatewaydocs
https://connector.scholargateway.ai/mcpSemantic search over licensed full text (>8M Wiley articles). OAuth via institutional access — check whether your library subscribes.
A worked example
A long AI-assisted review, carrying an explicit evidence label on every one of its references.
DNA Replication Termination: From Terminal Fork Geometry to Complete and Separable DNA Products
- references
- 513references
- carry a PMID or DOI
- 100%carry a PMID or DOI
- sections
- 9sections
- words
- ~30,600words
Resolved identifiers: 504 PMID · 451 DOI · 319 PMCID
Abstract
DNA replication must end without leaving unreplicated DNA, an active replisome on the product, or physical links that prevent daughter molecules from separating. These requirements create a distinct mechanical problem at the end of replication: the remaining parental duplex shrinks while torsional stress must be redistributed, replisomes enter one another's path, the final leading- and lagging-strand junctions must be converted into continuous DNA, and fork rotation can transfer intertwining into the replicated products. The solution is not one universal reaction. Programmed fork traps position terminal encounters in some bacterial chromosomes and episomes, origin firing and fork velocity generate broad encounter zones in many eukaryotic chromosomes, and rolling-circle or end-replicating systems terminate by strand transfer or processing of a physical DNA end. Nevertheless, these architectures can be compared as transformations of a terminal DNA substrate. Forks first reach a geometry that permits or requires remodeling; residual DNA synthesis and end processing then complete the nascent strands; in characterized yeast and metazoan systems, separate structural, biochemical, and extract studies support a model in which loss of fork DNA permits CMG ubiquitylation followed by Cdc48- or p97-dependent extraction; and topoisomerases or site-specific recombination remove direct links between products. Circular-template experiments directly expose covalent and topological product states; locus-resolved chromosomal assays can reveal incomplete or delayed synthesis. Failure therefore produces different pathologies depending on the interrupted transition, including renewed replication after bacterial fork fusion, persistent terminal gaps or replisomes, catenated products, and under-replicated DNA carried into mitosis. The central unresolved questions are the native structure of an unperturbed bacterial fork fusion, how final torsional stress is partitioned between the parental interval and daughter precatenanes, the minimal DNA change that licenses CMG removal, and the order of nascent-strand closure, replisome extraction, and product unlinking after the last parental interval disappears.
Evidence label on every reference
- [F-T]0complete exact primary-source full text was inspected and used for every claim attributed to the manifestation
- [PAR]28an exact identifier-matched target-paper passage was inspected, but complete primary full text was not used
- [ABS]481claims use only the validated exact target abstract
- [SEC]4a review or other secondary source is used for framing or explicitly attributed one-hop support
Not one of the 513 references is labelled [F-T]. Every claim rests on a validated abstract or an identifier-matched passage, and the review says so line by line rather than leaving you to assume otherwise. A reader can therefore tell exactly how far to trust each citation — which is the whole point of making a model declare what it actually read.
The guide
Where the prompts in this module come from.
Large Language Models for Science and Research: A Practical Guide, CC BY-NC 4.0. The full prompt library built on it lives at prompting.jmdewar.com.