Literature Review

---
name: literature-review
description: "Build or audit a literature review. Use for evidence maps, gap analysis, contribution checks, source verification, and synthesis planning."
---

# Literature Review Evidence Mapper

## Heritage and scope

This is an original Open Science Skills workflow for experimental and computational social science. It remixes high-level ideas from Cheng-I Wu's *Academic Research Skills for Claude Code* (CC BY-NC 4.0), especially evidence mapping, source verification, and mode separation between narrative literature review and formal systematic review. It is not a full ARS pipeline and should not copy ARS prose.

## Instructions

### 1. Classify the review task

Decide what the user needs:

- **Narrative/theory review:** organize concepts, mechanisms, and debates for an introduction.
- **Design precedent review:** identify prior treatments, measures, samples, estimands, or analysis strategies.
- **Contribution audit:** test whether the claimed gap survives contact with the closest prior work.
- **Evidence map:** summarize what each study establishes, where it applies, and what remains unresolved.
- **Systematic-review escalation:** when the user needs exhaustive search, screening, risk-of-bias, and PRISMA reporting.

Default to a narrative/evidence-map review unless the user explicitly asks for a systematic review, meta-analysis, or PRISMA-compliant output.

If the user does want a systematic review, be explicit about what this skill can and cannot do: steps 2-7 below produce the **protocol scaffold** -- question, boundaries, search strings, inclusion/exclusion criteria, and the evidence-map structure -- but exhaustive multi-database screening, dual-coder risk-of-bias assessment, and PRISMA flow accounting require dedicated tooling and human coders. Deliver the protocol scaffold, recommend registering it (PROSPERO or OSF), and point the user at screening tools (e.g., Covidence, ASReview) for the systematic phase.

### 2. Define the question and boundaries

Before summarizing papers, specify:

- Research question or review question.
- Population, setting, outcome, treatment/exposure, and mechanism scope.
- Disciplines and literatures that must be included.
- Time window and language restrictions, if any.
- Inclusion/exclusion logic for sources.
- What counts as "closest prior work."

If the user only gives a broad topic, first produce a short scoping memo with 2-4 possible review boundaries rather than writing a generic review.

### 3. Build the source base

Use the user's supplied sources first. Then identify obvious missing source classes:

- Seminal theoretical anchors.
- Most recent directly related empirical work.
- Meta-analyses, registered reports, study registries, or working-paper series.
- Methods papers that justify the research design.
- Null findings, failed replications, or unpublished registered studies when visible.
- Adjacent literatures that use different vocabulary for the same construct.

Run `citation-check` when the source list is large, messy, DOI-heavy, or likely to contain stale working papers.

### 4. Extract evidence, not summaries

For each important source, record:

- **Claim actually supported:** one sentence, no inflation.
- **Design and identification:** sample, setting, treatment/exposure, comparison, outcome, estimand, and key limitations.
- **Observable implication:** what the source lets a reader expect in the user's setting.
- **Boundary condition:** where the finding may fail.
- **Use in the user's paper:** background, theory, design precedent, measurement precedent, competing explanation, or gap support.

Do not produce chronological "Author A says X, Author B says Y" prose unless chronology is theoretically important.

### 5. Cluster the literature

Organize sources into 3-6 clusters. Prefer conceptual or mechanism clusters over method-only clusters:

- Mechanism families.
- Competing theories.
- Measurement traditions.
- Empirical settings or populations.
- Identification strategies.
- Evidence-quality tiers.

For each cluster, state what is settled, what is contested, and what would change the interpretation.

### 6. Test the claimed gap

Write a gap verdict:

- **Holds:** no close prior work answers the same question with the same population/mechanism/design.
- **Partly holds:** prior work answers part of it; the contribution must be narrowed.
- **Does not hold:** the claimed contribution is already established; reframe as replication, extension, boundary test, or synthesis.
- **Cannot assess:** missing sources or inaccessible source text prevent judgment.

When the gap is weak, propose a better contribution frame rather than only criticizing it.

### 7. Compose with sibling skills

- Use `narrative-building` after the evidence map exists to turn the review into the "Why-to-If-Then" funnel.
- Use `hypothesis-building` when the review implies falsifiable expectations and estimands.
- Use `pre-registration-writing` when the review supports confirmatory hypotheses.
- Use `methods-reporting` when reviewing how prior studies report designs, sample flow, and transparency.
- Use `journal-review` when auditing someone else's manuscript for novelty and placement.

## Output

Produce a `Literature Review Evidence Map`:

```
# Literature Review Evidence Map

Review question:
Scope and exclusions:
Search/source base:
Gap verdict: Holds / Partly holds / Does not hold / Cannot assess

## Closest Prior Work
| Source | What it actually establishes | Boundary | Relation to user's claim |

## Evidence Clusters
### Cluster 1: <name>
Settled:
Contested:
Missing:
Key sources:

## Contribution Diagnosis
Claimed gap:
Verdict:
Better contribution frame:

## Literature Review Architecture
1. <section purpose>
2. <section purpose>
3. <section purpose>

## Sentences the Review Must Earn
- <sentence-level claim that needs source support>

## Sources Needing Verification
| Source | Why |
```

## Quality checks

- [ ] The review question and scope are explicit.
- [ ] Closest prior work is identified before novelty is judged.
- [ ] Each key source is tied to an observable implication or boundary condition.
- [ ] The output distinguishes settled evidence from contested evidence.
- [ ] The gap verdict is calibrated, not inflated.
- [ ] Publication bias, registered studies, and null findings were considered when relevant.
- [ ] `citation-check` was invoked or recommended when source integrity was uncertain.
- [ ] The synthesis plan can feed directly into `narrative-building`.