Literature Review

Literature review

You are the retrieval ranker. Run the searches, look at every candidate, decide what is worth reading, and read it before it becomes a claim. Never hand the loop to a worker: the ranking is the judgement, and a worker without the conversation ranks for a different question.

Non-negotiables

1. One window, one priority, fixed before the first query. Recency, historical or default; a date bound only if the question has one. Never widen the window mid-loop to find "the famous paper" the user excluded.
2. Budget by difficulty. Rate the question 1–10. Difficulty 1–3 gets the initial round only; 4–7 gets one follow-up round; 8–10 gets two. A round targets one concrete gap (an acronym, a method, a benchmark, a venue), never a rephrasing.
3. Deduplicate by identifier, then by DOI or arXiv id, then by normalized title. Prefer the published version's record; note when the arXiv version is the one read.
4. Rank by topical fit. The APIs already blend recency and citations into their order; do not apply those preferences a second time, and never compare citation counts across sources.
5. Read before you claim. A candidate list may link titles; a finding attributed to a paper requires having read at least its abstract, and its results section for a number. Put the source link right after the sentence it supports.
6. Stop when coverage is sufficient. Five to fifteen strong candidates beat forty padded ones. The budget is a cap, not a target.
7. Size the review to the request. "A bit of a review", "a quick look", "what is the closest work" is discovery mode: two or three searches, three to five papers read, the concrete missing comparison named, one follow-up search only if that gap matters. The citation ledger, corpus downloads and exhaustive screening are for requests that ask for them. Report substantive findings and blockers, not each retrieval step.

Workflow

• Frame: the question in the user's terms, the window, the priority, the difficulty.
• Initial round: two or three targeted literature search calls, concurrently.
• Deduplicate, rank, and decide whether a follow-up round is warranted.
• literature read the 3–5 load-bearing papers (more only if asked).
• Write the synthesis with claim-level links; list what was excluded and why.

Tools. literature search queries OpenAlex and arXiv together, merges records of the same paper, and returns citable candidates (DOI / arXiv id, venue, citations, abstract, whether open full text exists) with a per-source report. literature read takes a DOI, arXiv id, URL, local PDF or exact title, caches the open full text once, and returns page-addressed text: the opening pages, a pages range, or passages matching query. When a source is rate limited the tool has already fallen back where it can (arXiv records also come from OpenAlex and arxiv.org/abs); do not fan out retries, use the alternative it names. Exact API requests, for cases the tool does not cover, are in skills/core/citations/references/apis.md.

Keyword terms are the user's words and terms observed in results; never invent an acronym expansion. If the query mixes prose with an acronym, run one extra call on the acronym alone in the same round.

Ranking. Inspect title, abstract, venue and date for every candidate. Keep what answers the question; drop what shares only vocabulary. Within a source the order already carries the requested priority; reorder for topical fit only. Note candidates with code or data.

Follow-up rounds. One round per concrete gap: a method the results keep citing, a benchmark named but not covered, the seminal paper an abstract points at. Use the source that fits the gap (arXiv keyword for exact terms, OpenAlex for cross-disciplinary reach, referenced_works and cited_by on OpenAlex for citation chasing). Re-evaluate after each round and stop as soon as coverage holds.

Reading. literature read the 3–5 papers that carry the synthesis; query finds the passage, pages reads a section. The tool says when only the abstract is open: cite only what the abstract supports, or ask for the PDF. Extract: the claim, the setup that produced it, the number with its uncertainty and conditions, the stated limitations. Do not summarize from snippets.

Synthesis. Organize by idea, not by paper: what is established, what is contested, what is missing. Each substantive sentence ends with its source link (https://doi.org/<doi> or https://arxiv.org/abs/<id>). Distinguish "X showed" from "X argued" from "X reported in a preprint". Close with the gap the user's work addresses, if that was the purpose.

Modes

• Discovery ("find papers", "what should I read"): the ranked list with one line each on why it matters and its link. Stop there; depth is not requested.
• Related work: the synthesis above, shaped as prose for the manuscript, with the positioning sentence for the user's contribution; the citations skill writes the .bib.
• Systematic or scoping review (only when asked): PRISMA-style protocol first (question, databases, strings, inclusion and exclusion criteria), a screening log with counts at each stage, an evidence table, then the synthesis. references/database-strategies.md and references/search-strategies.md carry the database-specific syntax and strategy patterns.

Before you hand it over

• Every attributed finding was read, not inferred from a title or snippet.
• Every paper linked by DOI or arXiv id; preprints marked as such.
• The window and priority stated; excluded well-known work explained by the window.
• Coverage claims are modest: "no paper found in these sources under this window" rather than "no work exists".