lit-review-initiate

name: lit-review-initiate description: Initiates a literature review — generates search queries from a spec, retrieves papers, filters and ranks results, and presents them for human review. Use when starting a literature review with no existing paper collection.

Literature Review: Initiate

Overview

This skill runs the first pass of a literature review when no papers exist yet. It takes a literature review spec as input, generates search queries, retrieves candidate papers, filters and ranks them against the spec criteria, and presents results in three tiers for human selection.

Expected Input

• Required: A literature review specification with the following fields:
  • Research question: The guiding question the literature should answer
  • Topics: Include / Prioritize / Exclude by topic (match by substantive focus, not mere mention)
  • Sources: Include / Prioritize / Exclude by source type
  • Dates: Include / Prioritize / Exclude by publication year
  • Language: Include / Exclude by language
• Format: The spec can be provided inline or as a reference to a file

Expected Output

• Deliverable: A ranked list of candidate papers for human review
• Format: Three-tier list (see Output Format below)

Step 1: Query Generation

Read the spec carefully. Generate a diverse set of search queries that together cover the breadth of the included topics and the research question.

Query design principles:

1. Decompose first: Break the research question into 2–4 concept blocks (e.g., actor, mechanism, domain). Generate synonym-rich term sets per block — including abbreviations, spelling variants, and related terms.
2. Adapt to the platform: Semantic Scholar uses semantic ranking, not Boolean matching — use simple, concept-focused queries (2–5 key terms) rather than exhaustive Boolean strings. Save broader synonym exploration for WebSearch queries.
3. Vary systematically: Generate queries that cross concept blocks in different combinations. Include queries at different specificity levels (broad concept vs. narrow sub-topic).
4. Cover the breadth: Span different framings and synonyms of each included topic. Include sub-topics implied by the spec even if not explicitly named.
5. Avoid excluded topics: Do not generate queries that would primarily surface excluded topics.
6. Err toward sensitivity: Retrieving too many candidates is preferable to missing relevant papers. Precision rates of 1–5% are normal in systematic reviews.
7. Target 10–20 queries for a typical spec.

Step 2: Retrieval

Execute each query using the WebSearch tool and the Semantic Scholar API. For each result, collect: title, authors, year, venue, abstract, URL/DOI, and citation count.

Web search — use the WebSearch tool directly (not via Bash/curl). Use it to find government reports, think tank publications, and other non-academic sources. Use WebFetch to retrieve content from specific URLs when needed.

Semantic Scholar API — use the Bash tool with curl. If available, set the API key in .env as SEMANTIC_SCHOLAR_API_KEY for higher rate limits; the API also works without a key at lower rate limits (100 requests/5 minutes).

• Search: GET https://api.semanticscholar.org/graph/v1/paper/search?query=<query>&fields=title,authors,year,venue,abstract,citationCount,externalIds,openAccessPdf&limit=50
  • Header (if key available): x-api-key: $SEMANTIC_SCHOLAR_API_KEY
• Paper details: GET https://api.semanticscholar.org/graph/v1/paper/<paperId>?fields=title,authors,year,abstract,citationCount,references,citations
  • Header (if key available): x-api-key: $SEMANTIC_SCHOLAR_API_KEY
• Optional filters (use only when appropriate, not by default): year=2015- (range), minCitationCount=10, fieldsOfStudy=Political Science, publicationTypes=JournalArticle,Review

Deduplicate across queries. Aim to collect at least 5–10x more candidates than the final list will contain.

Step 3: Filter and Rank

Filter out results that:

• Are in excluded source types or languages
• Fall outside included date ranges (unless seminal)
• Focus primarily on excluded topics (even if included topics appear)
• Are clearly not substantively relevant to the research question

Rank remaining results into three tiers:

• Highly Recommended: Strong match on relevance to the research question, hits prioritization criteria (source type, recency, topic focus), and appears high quality
• Recommended: Solid relevance and quality, meets include criteria without hitting priority criteria
• Optional: Plausible relevance but lower confidence — tangential focus, older, lower venue quality, or unclear fit

For each result, write a brief rough note explaining what led to it (e.g., which query surfaced it, which included topic it addresses).

Step 4: Human Review Presentation

Present results in three sections:

Section 1 — Highly Recommended: Strong fit, read these first. Section 2 — Recommended: Solid fit, worth reviewing. Section 3 — Optional: Lower confidence, human can skim or skip.

Each entry: title, authors, year, venue, link, and rough note. Human selects papers to download and add to the collection.

Output Format

## Highly Recommended

1. **Title** — Author(s) (Year). *Venue*.
   Link: [URL or DOI link]
   Note: [what led to this result, e.g., "query: climate adaptation governance; directly addresses institutional design for resilience"]

2. ...

## Recommended

1. **Title** — Author(s) (Year). *Venue*.
   Link: [URL or DOI link]
   Note: [rough note]

2. ...

## Optional

1. **Title** — Author(s) (Year). *Venue*.
   Link: [URL or DOI link]
   Note: [rough note]

2. ...