| name | draft-survey |
| description | Turns a topic into a ranked reading list of the best papers plus a two-column, arXiv-ready literature-review/survey draft grounded in verified citations. Use when a researcher says "write a survey on X", "do a literature review of X", "give me a ranked reading list on X", "what should I read on X first", "draft a survey paper for arXiv", or "two-column related-work survey on X". Searches across DBLP/Crossref/Semantic Scholar/arXiv, ranks candidates by citation impact, venue strength, recency, and citation-graph centrality, clusters them into a taxonomy, verifies every reference, and writes an original-prose survey (intro, taxonomy, per-theme synthesis, open problems) where each claim cites a real, resolved paper. Never fabricates citations, never copies source text — original synthesis only; the author stays the author. Trigger words - survey, literature review, reading list, ranked papers, what to read, arxiv survey, two-column review. |
Draft Survey
Two deliverables from one topic: a ranked reading list (what to read, best first, with a one-line why) and a two-column, arXiv-ready survey draft that synthesizes the area in original prose with verified citations. Composes find-papers, verify-citations, study-exemplars, and draft-related-work.
When to use vs. literature-review
literature-review builds the related-work for your own paper — scoped to your contribution.draft-survey produces a standalone survey/review document (a reading list + a publishable 2-column draft) of a whole area. Use it to learn a field fast or to draft a survey paper for arXiv.
Inputs
- The topic (e.g. "geospatial data conflation"), optionally a sub-scope and a target length (default 5–6 two-column pages).
- Optional: a venue/format (default: generic two-column
article, which compiles on arXiv).
Process
- Gather candidates broadly. Run
find-papers across DBLP + Crossref + Semantic Scholar + arXiv for the topic and its synonyms; then run its citation-graph expansion so seminal anchors and direct lineages are not missed (a survey that omits the foundational papers is a weak survey).
- Rank them. Run
python3 scripts/rank_papers.py candidates.json — a composite of normalized citation count, venue tier, recency, and citation-graph centrality (weights documented in references/ranking-criteria.md). Output the ranked reading list: rank, title, authors, year, venue, a citation/impact signal, and a one-line why read this (seminal / survey / SOTA / dataset / contrarian). Keep seminal and recent both represented.
- Verify every entry. Route the list through
verify-citations so each has a real DOI/arXiv id; drop or flag anything unresolved. A survey with a fabricated reference is disqualifying.
- Build a taxonomy. Cluster the verified papers into 3–6 themes/sub-problems (the survey's section structure), each with its lineage (foundational → recent).
- Draft the survey, in original prose. Write a two-column
.tex: abstract, introduction (scope + why a survey now), one section per theme (synthesize and contrast methods — never copy source sentences), a cross-cutting comparison (a table helps), open problems / future directions, conclusion, and \bibliography. Every claim cites a verified paper. Target the requested length.
- Make it arXiv-ready. Ensure it compiles (
latexmk), uses a portable two-column class, and the .bib is clean. arXiv has no peer-review desk-reject, but it expects a compilable source and a real abstract; preflight-check can sanity-check length/structure.
Output
reading-list.md — the ranked, verified, annotated reading list.survey.tex + refs.bib — the two-column, arXiv-ready draft (original prose, verified citations).- Both written to
paper-workspace/research/ or a path the user names.
Guardrails
- Real citations only. Every reference is resolved via
verify-citations; never invent a paper, DOI, author, or year to fill a gap.
- Original synthesis, not copying. Summarize and contrast in your own words; do not paste sentences from abstracts or papers. No paper content is bundled (fetch on demand, process transiently).
- No completeness theater. State the search scope and that a survey is never exhaustive; flag themes where coverage is thin rather than padding.
- Copilot, not pilot: the author reviews the reading list and the draft, and is the author of any submission.
Source verification
Citation impact and venue facts come from live scholarly APIs, not memory, and each cited paper carries a resolvable identifier the user can open.
Memory
Uses .paper-memory/ as described by paper-memory-convention.md: read prior reading lists/lessons for the topic at start; append the topic, the chosen taxonomy, and any coverage gaps at end (deduped).
