deep-literature-review

name

deep-literature-review

description

Systematic literature search and organization for academic papers. Use when starting a new paper's related work, expanding an existing Related Work section, surveying a research area, building a reading list, or when the user asks to "find papers on X", "what's the related work for X", "literature review", "survey this area", or "who else has worked on X". Also use when the user provides a draft and wants to strengthen its citations or find missing references to important prior work.

Deep Literature Review

Build a structured, thematic literature map for a research topic. The goal is not just to find papers but to understand the landscape — who did what, how ideas evolved, where the gaps are, and how the user's work fits in.

When to Use This vs. research-paper-writing

This skill handles the research phase — finding, reading, and organizing papers. The research-paper-writing skill handles the writing phase — turning that organized knowledge into polished Related Work prose. Use this skill first, then hand off to research-paper-writing for the actual drafting.

Core Workflow

1. Scope the Search

Before searching, establish boundaries:

• Research question or topic — What is the user's paper about?
• Key terms and synonyms — The same idea often goes by different names across subcommunities (e.g., "few-shot learning" vs "low-resource learning" vs "meta-learning")
• Time range — Default to the last 5–7 years for fast-moving fields (ML/AI), longer for established fields
• Venue preferences — Does the user care about specific conferences or journals?
• Scope limits — How exhaustive should this be? A workshop paper needs fewer references than a survey

Ask the user to confirm the scope before diving in. A well-scoped search saves significant time.

2. Search Systematically

Use multiple search strategies to avoid blind spots:

Keyword search — Start with the core terms. Use web search to find papers on Semantic Scholar, arXiv, Google Scholar, and ACL Anthology. Vary the keywords: try synonyms, broader terms, and narrower terms.

Citation chaining — When you find a highly relevant paper:

• Forward: Who cites this paper? (Find via Semantic Scholar or Google Scholar)
• Backward: What does this paper cite? (Read its references)

This is how you find papers that keyword search misses.

Access limitations. In practice, you can reliably access abstracts, metadata, and publicly available content (arXiv, Semantic Scholar API, open-access papers). Most published papers behind paywalls cannot be read in full. Be upfront about this: when a paper seems relevant but you can only see the abstract, say so and flag it for the user to read. Do not fabricate details about papers you haven't fully read — summarize only what you can actually see.

Anchor papers — Ask the user: "Are there 2–3 papers you consider most relevant to your work?" Start citation chaining from these. The user's own prior work is also an anchor.

Venue scanning — For the user's target venue, skim recent proceedings for relevant titles. Conference papers cluster by topic; finding one often reveals several nearby.

3. Read and Extract

For each paper found, extract a structured summary based on what you can access (full text, abstract-only, or metadata-only — be explicit about which):

**[Short Label]** Author et al., Year. "Title." Venue.
- **Access level**: full text / abstract only / metadata only
- **Problem**: What problem does it address?
- **Approach**: What's the key idea or method?
- **Results**: What are the main findings?
- **Relation to our work**: How does it connect — precursor, alternative approach,
  complementary, addresses different aspect?
- **BibTeX key**: suggested_bibtex_key

Don't try to read every paper in depth. Triage:

• Must-read (3–8 papers): Directly related, same problem or method
• Should-cite (10–20 papers): Related but different angle, establishes context
• Awareness (unbounded): Good to know, cite only if space permits

4. Organize Thematically

Group papers into themes, not chronologically. A good Related Work section tells a story about the research landscape, not a timeline.

Common organizational patterns:

• By problem aspect — "Prior work addresses X through three angles: A, B, and C"
• By methodology — "Approaches fall into model-based, data-based, and hybrid categories"
• By progression — "Early work focused on A, which led to B, and recent work explores C"

For each theme, write a 2–3 sentence summary of the group's contribution and how the user's work relates to or differs from it.

Produce an outline like:

## Related Work Themes

### 1. [Theme Name]
Summary: ...
Papers: [Label1], [Label2], [Label3]
Gap/difference from our work: ...

### 2. [Theme Name]
Summary: ...
Papers: [Label4], [Label5]
Gap/difference from our work: ...

5. Identify Gaps

After mapping the landscape, explicitly state:

• What has not been done (the gap the user's paper fills)
• What assumptions prior work makes that the user's work relaxes
• What combinations haven't been tried
• Where results are contradictory or inconclusive

This directly feeds into the Introduction's motivation and the Related Work's positioning.

6. Generate BibTeX

Produce a clean .bib file with all recommended citations. For each entry:

• Use a consistent key format: authorYear (e.g., vaswani2017attention)
• Include: author, title, year, booktitle/journal, and DOI/URL when available
• Prefer the published venue version over arXiv preprints when both exist
• Flag any entries where you're uncertain about correctness — the user should verify

Output Deliverables

When the literature review is complete, provide:

1. Literature map — Thematic organization with paper summaries
2. Gap analysis — What's missing that the user's work addresses
3. BibTeX file — Ready to paste into the user's .bib
4. Recommended reading order — For the must-read papers, suggest which to read first
5. Related Work outline — Suggested section structure for the paper

Search Quality Checks

Before declaring the search complete:

• Checked at least 3 different keyword formulations
• Followed citation chains from the 2–3 most relevant papers
• Included work from the last 2 years (field moves fast)
• Covered the major relevant venues for this topic
• No obvious subcommunity or methodology is unrepresented
• User's own prior work is included if relevant

Pitfalls to Avoid

Recency bias — Don't only cite recent work. Foundational papers matter, even if they're from 2015.

Popularity bias — High-citation papers aren't always the most relevant. A niche paper that directly addresses the user's setting may be more important than a famous one that's tangentially related.

Confirmation bias — Include papers that challenge or complicate the user's narrative. Reviewers will know about them, and it's better to acknowledge them proactively.

Over-collection — Know when to stop. The goal is a well-organized 20–40 paper bibliography, not an exhaustive database. For a conference paper, 30–50 references is typical; for a journal survey, more is expected.