Claude skills for automated literature surveys: verified corpora, honest metrics, minimal figures, and a short PDF report that tells you who did it best and where the field is heading.
Ptolemy's Almagest, transmitted through Arabic as al majisti, "the greatest", was the definitive synthesis of the astronomical state of the art for over a millennium. That is the bar a survey should aim for: not a pile of citations, but the compendium a field measures itself against.
How a survey runs
flowchart LR
A[Scope] --> B[Recon]
B --> C{Coordinator}
C --> D[Survey agents]
D --> C
C --> E[Overseer audit]
E --> C
C --> F[Verified corpus]
F --> G[Figures]
F --> H[Hypotheses]
G --> I[5 page PDF]
H --> I
Install
In Claude Code:
/plugin marketplace add DanieleMosh/almagest-survey
/plugin install almagest-survey@almagest-survey
Or clone and symlink the skill directly:
git clone https://github.com/DanieleMosh/almagest-survey.git
ln -s "$(pwd)/almagest-survey/skills/literature-survey" ~/.claude/skills/literature-survey
Then start a survey with /literature-survey <your topic>, or just ask Claude to survey the state of
the art on something.
Skills
| Skill | What it does |
|---|---|
literature-survey |
Runs a state of the art survey on any STEM topic: parallel research agents with an overseer, a DOI verified corpus, open access PDF acquisition, Tufte style seaborn figures, extracted hypotheses, and a compiled LaTeX report of about 5 pages. |
What a run leaves behind
<topic>-survey/
report.pdf the deliverable, about 5 pages
data/scope.md the agreed scope, settled with you before any harvesting
data/papers.csv the corpus, single source of truth for every count in the report
data/refs_verified.csv Crossref and arXiv resolution result for every reference
papers/ retrieved open access PDFs, plus MANIFEST.md for the rest
figures/ regenerated from papers.csv, so prose and plots cannot drift
Every DOI is resolved against Crossref before it reaches the report, unverifiable fields stay blank rather than being guessed, and papers behind a paywall are listed with links instead of being scraped.
License
MIT
No comments yet
Be the first to share your take.