literature-review

Literature Review

Write a comprehensive, structured literature review on a user-specified topic. This skill produces a standalone review of what the field knows, where it disagrees, and what remains open. It is not for positioning a specific paper's contribution (use literature-positioning-map for that) or for writing a related-literature section within a manuscript (use research-paper-writer for that).

When to use

• The user wants to survey a topic or research area
• The user wants a literature review for a dissertation chapter, qualifying exam, survey paper, or personal reference
• The user wants to understand the state of knowledge on a question before starting a project

Inputs to collect

Before starting, confirm these with the user:

Input	Required?	Default
Topic or research question	Yes	—
Output format: markdown / latex-section / latex-standalone	Yes	markdown
Scope: quick (~15 papers, field orientation) / focused (~25 papers) / comprehensive (~50 papers)	No	comprehensive
Target .tex file (if latex-section)	If applicable	—
Existing .bib file path	No	Auto-detect or create new
Known key papers to include	No	—
Time period filter	No	All years
Specific journals to emphasize	No	—

If the user provides a topic and format in their initial message, proceed without asking. Fill defaults for anything not specified.

Workflow

Phase 0 (quick scope only): Field Orientation

If scope is quick, skip the full review workflow. Instead:

1. Run the architecture search (sortBy: "citedByCount", matchCount: 15) and frontier search (minYear: 2020, matchCount: 15).
2. Use get_paper_details_batch on the top 10 results from the architecture search.
3. Produce a field orientation document at output/field_orientation.md:

# Field Orientation: [Topic]

## 10 Must-Read Papers
[Ranked by citation count. For each: author (year), title, journal, 1-sentence contribution.]

## 3 Main Debates
[What the field disagrees about, with papers on each side.]

## 3 Dominant Methods
[How this field typically does empirical work.]

## 3 Common Datasets
[What data most papers use, via search_datasets.]

## 5 Frontier Questions
[What the recent papers (2020+) are working on that remains unresolved.]

4. Save the paper set to output/paper_set.json and log searches to output/search_log.md.
5. Log to lab notebook and suggest next steps: /lit-review [topic] with focused or comprehensive scope, /brainstorm [topic], or paper-reader on the top 3 papers.
6. Stop. Do not proceed to the full review phases.

Phase 1: Search and collect

Target ~50 unique papers for comprehensive scope, ~25 for focused scope. Execute searches in this order:

Step 1 — Architecture search (always first):

• search_papers (query: the core topic, sortBy: "citedByCount", matchCount: 15) to immediately see the field's citation hierarchy: which papers define it, which are the most influential.
• This search reveals the foundational papers. Every subsequent judgment (what to discuss at length, what to group parenthetically, what to cut) is informed by this hierarchy.

Step 2 — Frontier search (always second):

• search_papers (query: core topic, minYear: 2020, matchCount: 15) to find the latest published and working papers that represent the current edge of the field.

Step 3 — Thematic angles:

• search_papers (query: first sub-question or thematic angle, matchCount: 15)
• search_papers (query: second sub-question or thematic angle, matchCount: 15)
• If the topic has a third distinct angle, add a third search.

Step 4 — Topic-filtered journal search:

• top_cited_articles (journalNames: relevant top journals, query: the specific topic, compact: false) to find seminal papers on the topic within those journals that may not have appeared in the keyword searches.

Step 5 — Adjacent or methodological:

• search_papers (query: methodological approach or adjacent field angle, matchCount: 10) for breadth.

Step 6 — Verify and enrich:

• get_paper_details_batch (up to 25 paper IDs per call) on the top 30-40 unique papers to read abstracts, confirm relevance, and extract key findings. Use 2 batch calls rather than 30+ individual calls.
• Deduplicate across all searches.
• If the user provided known key papers, verify they appear. If not found via search, include them manually and use get_paper_details to confirm details.

Save to shared data files:

• Write all collected papers (with id, title, authors, year, journal, citedByCount, abstract, fullText when available, doi, source_queries) to output/paper_set.json. If the file exists, merge and deduplicate by id.
• Append all search queries with parameters and result counts to output/search_log.md.

Relative citation tiering: After deduplication, sort all collected papers by citedByCount and assign influence tiers using relative ranking within the collected set:

Tier	Label	Rule	Treatment in the review
1	Foundational	Top 10% by citation count within the collected set	3-5 sentences each. Describe what they found, how, and why it mattered. These anchor the review.
2	Established	Next 30% by citation count	1-2 sentences each, or grouped into synthesized claims with 2-3 papers per sentence.
3	Emerging	Bottom 60%, especially papers published in the last 5 years	Grouped into frontier paragraphs. Cited parenthetically to support collective findings.

A paper that appears across 3+ separate search queries is likely a network hub. Promote it one tier (e.g., Established to Foundational) regardless of citation rank.

When a paper has fullText available in the paper set, use it (not just the abstract) to make more informed judgments about mechanism, method, and contribution.

Ranking criteria (for deciding which papers to keep when cutting to target count):

1. Direct relevance to the topic
2. Influence tier (Foundational papers are never cut)
3. Recency (recent work that shifts the field gets priority over older low-citation work)
4. Methodological contribution (papers that changed how the field studies the topic)

Phase 2: Propose strand structure

After collecting papers, propose 4-6 thematic strands. Present to the user for approval before writing.

Deliver this structure:

# Proposed Literature Review Structure

## Topic: [user's topic]
## Scope: [focused/comprehensive] — [N] unique papers collected
## Influence tiers: [X] Foundational / [Y] Established / [Z] Emerging

### Proposed strands:

1. **[Strand name]** — [1-sentence description]
   Key papers: [3-5 author-year citations]
   Narrative arc: [what story this strand tells, from early work to current state]

2. **[Strand name]** — [1-sentence description]
   Key papers: [3-5 author-year citations]
   Narrative arc: [story]

3. ...

### Cross-cutting themes:
- [e.g., methodological evolution from X to Y]
- [e.g., a key empirical debate between findings A and B]

### Identified gaps:
- [Gap 1: what the literature has not addressed]
- [Gap 2: where findings conflict without resolution]

### Papers that don't fit neatly:
- [Paper] — could go in strand X or Y; recommend [placement]

Strand organization principles (in order of preference):

1. By debate or tension — group papers by which side of a disagreement they support
2. By mechanism or channel — group papers by the economic force they emphasize
3. By methodology — group papers by empirical approach when methodology drives different answers
4. Chronological within strands — within each strand, order foundational work first, then development, then frontier

Never organize by topic label alone ("this literature relates to X, Y, and Z" with no internal structure).

Checkpoint: Wait for user approval or modifications. Do not proceed to writing until the user confirms the strand structure.

Phase 3: Write the review

Per-strand writing protocol

For each strand, write in this order:

1. Opening frame (1-2 sentences): State the strand's central question or contribution to understanding the topic. Why does this line of work exist?

2. Foundational work (Tier 1 papers, 3-5 sentences each): Describe the key papers that anchor this strand. These get the most individual attention. State what they found, how they found it, and why it mattered. Do not mention citation counts in the review prose; let the depth of treatment signal the paper's importance. Citation counts belong in the reading list, not in the narrative.

3. Established evidence (Tier 2 papers, synthesized, not enumerated): Group the body of work by finding, not by author. Write about what the literature collectively shows, with citations supporting claims. Example:

   • GOOD: "Subsequent work established that credit constraints amplify housing cycles, with effects concentrated among low-income borrowers (Author 2015; Author 2017) and in regions with inelastic housing supply (Author 2016; Author 2019)."
   • BAD: "Author (2015) studies credit constraints and housing. Author (2016) studies housing supply elasticity. Author (2017) also studies credit constraints."

4. Recent frontier (Tier 3 papers, 2-3 sentences): What has the last 2-3 years added? New data, new methods, new findings that shift understanding? These papers may have low citation counts simply because they are new.

5. Gaps, tensions, or open questions (1-2 sentences): What does this strand leave unresolved? Where do findings conflict? What has not been studied?

6. Transition (1 sentence): Connect to the next strand.

Cross-strand synthesis section

After all strands, write a synthesis section covering:

• What the literature collectively establishes (the consensus, if any)
• Where findings conflict or remain ambiguous (unresolved debates)
• Methodological limitations shared across studies
• Open questions and promising directions for future work

Writing rules

Follow all project writing norms (references/writing-norms.md, references/banned-words.md):

• Synthesize, do not enumerate. The review should read as a narrative about a field, not as a list of papers.
• Cite to support claims, not to name-drop. Every citation should back a specific point.
• No filler intensifiers. Do not use "importantly," "crucially," "interestingly," "notably."
• No em dashes. Use commas, parentheses, colons, or separate sentences.
• No promotional language. Do not use "novel," "groundbreaking," "seminal" (even for truly seminal papers, describe what they did instead of labeling them).
• Precise language. Replace vague claims ("many papers study X") with specific ones ("a large body of work, beginning with Author (Year), examines X").
• Gap claims require evidence. Do not write "no one has studied X" or "the literature is silent on X" unless the search results support this. Prefer "to our knowledge, based on [N] papers reviewed" or simply describe what has been studied and let the gap speak for itself.
• Seminal papers get more space. 2-4 sentences for foundational work. Supporting papers get grouped into synthesized claims with parenthetical citations.
• Conflict is valuable. When papers disagree, describe both sides fairly and note what might explain the disagreement (different samples, methods, time periods, settings).

Phase 4: Citations and output

Citation handling

After writing the review:

1. Collect all paper IDs cited in the review.
2. export_citations (list of paper IDs, format: bibtex) to generate BibTeX entries.
3. Check for an existing .bib file:
   • If the user specified one, read it and append new entries (skip duplicates by checking cite keys).
   • If none specified, look for *.bib files in the project. If found, ask the user which to use.
   • If no .bib exists, create one:
     • markdown format: notes/literature_review_references.bib
     • latex-section format: same directory as the .tex file
     • latex-standalone format: paper/literature_review/references.bib
4. Write the .bib file.

For papers where export_citations does not return a result (e.g., the paper was mentioned by the user but not found in Corbis), construct a manual BibTeX entry from known information and flag it for the user to verify.

Output by format

Markdown (markdown):

• Write to notes/literature_review_[topic_slug].md
• Use standard Markdown headers for strands
• Use parenthetical author-year citations: (Author, Year)
• Include the reading list (see below) at the end
• Append the .bib file path at the bottom for reference

LaTeX section (latex-section):

• Read the target .tex file
• Insert the review at the location specified by the user (or find an appropriate \section{} marker)
• Use \citet{} for textual citations ("Author (Year) finds...") and \citep{} for parenthetical citations ("...as shown in prior work \citep{author2020}")
• Append new entries to the existing .bib file
• Use Edit tool, not Write tool, to insert into existing files

LaTeX standalone (latex-standalone):

• Copy latex_template/ to paper/literature_review/ if it does not exist
• Read the template .tex file to understand its structure
• Replace the title with "Literature Review: [Topic]"
• Replace the abstract with a 100-word summary of the review's scope and key findings
• Write the review body into the appropriate sections
• Create references.bib in the same directory
• Ensure \bibliography{references} points to the correct file

Reading list

For all output formats, also produce a reading list using assets/reading-list-template.md. This is a curated table of the 10-15 most important papers with one-line descriptions of their key contributions.

• Write to notes/reading_list_[topic_slug].md
• Organize into these sections:
  • Read first (3-5): The papers you must read before doing anything else in this field
  • Foundational (5-7): The highest-cited papers that define the field
  • Closest to [user's question] (3-5): The papers most directly relevant to the user's specific interest (if one was stated)
  • Best empirical designs (3-5): Papers with the cleanest identification strategies, worth studying for methodology
  • Recent frontier (4-5): The most important work from the last 5 years

Phase 5: Log and state

Lab notebook: Append an entry to notes/lab_notebook.md:

---

### [DATE] — Literature Review: [Topic]

**What was done**: Comprehensive literature review on [topic]. Searched [N] papers via Corbis, reviewed [M] abstracts, organized into [K] thematic strands.

**Strands covered**:
- [Strand 1]: [1-sentence summary]
- [Strand 2]: [1-sentence summary]
- ...

**Key gaps identified**:
- [Gap 1]
- [Gap 2]

**Key conflicts identified**:
- [Conflict 1]

**Output files**:
- Review: [path]
- Reading list: [path]
- Bibliography: [path]

**Next steps**: [e.g., use findings to inform research-idea-generator, or feed into literature-positioning-map for a specific paper]

Project state: If notes/project_state.md exists, update the literature positioning section with the strand names and key gaps.

Phase 6: Coverage report

After all outputs are written, present a brief report in chat:

# Literature Review — Coverage Report
## Topic: [topic]
## Scope: [focused/comprehensive]
## Papers cited: [N] (of [M] unique papers found)
## Influence tiers: [X] Foundational / [Y] Established / [Z] Emerging
## Strands:
1. [Strand 1] — [N papers]
2. [Strand 2] — [N papers]
...
## Key gaps identified:
- [Gap 1]
- [Gap 2]
## Key conflicts identified:
- [Conflict 1]
## Output files:
- Review: [path]
- Reading list: [path]
- Bibliography: [path]

What this skill does NOT do

• Position a specific paper's contribution against the literature (use literature-positioning-map)
• Write a related-literature section for a manuscript in progress (use research-paper-writer)
• Generate research ideas from identified gaps (use research-idea-generator, though the gaps from this review are excellent inputs)
• Screen or evaluate a specific research idea (use finance-idea-screening)

This skill can feed into all of the above. A natural workflow is: literature-review to map the field, then research-idea-generator to brainstorm from the gaps, then literature-positioning-map to position a chosen idea.

Guardrails

• Search before claiming. Every assertion about what the literature does or does not contain must be grounded in the Corbis search results. Do not rely on parametric knowledge alone.
• Do not pad. If only 35 papers are genuinely relevant, do not stretch to 50. Quality over count.
• Do not fabricate. If a paper's details are uncertain, use get_paper_details or get_paper_details_batch to verify before including it. If unavailable, flag it for the user.
• Conflicts are features, not bugs. When papers disagree, present both sides. Do not smooth over genuine disagreements.
• Gaps are claims. Saying "the literature has not studied X" is a strong claim. Support it with search evidence or soften the language.
• Respect the checkpoint. Do not skip Phase 2 approval. The strand structure determines the quality of the review.
• No topic-label reviews. "This relates to three literatures" with no internal structure is the failure mode this skill exists to prevent.

Example prompts

• "Write a comprehensive literature review on climate risk and commercial real estate pricing."
• "Survey the literature on bank capital regulation and lending, focused scope, output as markdown."
• "I need a literature review on mortgage default and neighborhood spillovers for my dissertation. LaTeX standalone."
• "Add a literature review section on algorithmic trading and market quality to my paper at paper/main.tex."
• "What does the literature say about corporate cash holdings and investment? Comprehensive review, last 20 years."
• "Review the literature on housing supply elasticity. I know Saiz (2010) and Gyourko, Saiz, and Summers (2008) should be included."