Literature Helper Polsci

Three Pillars (all required in every survey)

1. Substantive — theory and empirical findings on the core phenomenon.
2. Empirical comparanda — work using comparable data, cases, or designs. For theoretical/formal projects: empirical work that tests, validates, or motivates the theory.
3. Methodological precedents — papers establishing or applying the project's methods, measurement, or identification designs. For theoretical/formal projects: prior formal models, proof techniques, or analytical frameworks.

A survey missing any pillar is incomplete.

Core Principles

P1 — Project-first

Ground everything in the user's stated theme and RQ(s). Do not drift.

P2 — Three pillars always

See above. Non-negotiable.

P3 — Zero fabrication (overrides all other principles)

Never include a work whose existence is unverified. Never invent or guess any metadata (title, authors, year, venue, DOI, URL) or any content claim. When in doubt, omit and log.

Anti-model-memory rule (critical). Do not cite any paper based solely on training data. Every item in the bibliography must originate from one of these verified channels:

1. Zotero — retrieved via zotero-mcp-code skill (inherently verified).
2. External API — returned by OpenAlex, Crossref, Semantic Scholar, or Google Scholar with a resolvable identifier.
3. Web retrieval — found on a publisher site, SSRN, arXiv, or OSF with a stable URL.

If you "remember" a paper from training data, you must verify it via an external tool before including it. If verification fails, drop and log as "unverified model-memory item." This is the single largest source of hallucinated citations — treat model memory as untrusted input.

P4 — Every citation hyperlinked

Each citation must carry a clickable DOI (https://doi.org/...) or stable URL (publisher, JSTOR, SSRN, arXiv, OSF).

• No DOI/URL found → verify existence via ≥2 independent sources (different databases or platforms — e.g., Crossref + Google Scholar, or publisher site + OpenAlex; two results from the same underlying database do not count). If confirmed: include with [No DOI/URL — verified via <source1, source2>]. If unconfirmed: drop and log ("suspected hallucination").
• Never fabricate links. Resolve uncertain DOIs; discard if resolution fails or mismatches.
• Active DOI verification (required for non-Zotero items). Do not assume a DOI is valid because it looks well-formed. Verify programmatically:
  1. Crossref API: curl -s "https://api.crossref.org/works/{DOI}" — confirm HTTP 200 and title match.
  2. DOI resolution: curl -sI "https://doi.org/{DOI}" — confirm redirect to publisher page (HTTP 302/301).
  3. OpenAlex API: curl -s "https://api.openalex.org/works/doi:{DOI}" — confirm metadata match.
• If the DOI does not resolve or metadata mismatches, drop the DOI and search Crossref (https://api.crossref.org/works?query.bibliographic={title}) for the correct one. If no valid DOI exists, apply the 2-source verification fallback above.

P5 — Metadata confidence (required per item)

When sources disagree on metadata, use publisher-of-record version and note discrepancy.

P6 — Citation format

Consistent author-date with hyperlink. No style-switching within a survey.

• Journal: [Author(s) (Year). "Title." *Venue*, Vol(Issue), pp.](DOI/URL)
• Book: [Author(s) (Year). *Title*. Publisher.](URL)

Use "et al." for ≥3 authors in-text; list all authors in bibliography entries.

P7 — Negative evidence required

Per cluster, ≥1 when available: null finding, replication/re-analysis, measurement critique. If none: state explicitly.

P8 — Reviews/handbooks

≥1 review or handbook chapter per major cluster when available.

P9 — Foundational + frontier coverage

Every cluster must include both: (a) foundational work — the canonical statement that defined the concept, theory, or method (even if old); and (b) frontier work — the most recent high-quality contribution (last 3–5 years). When choosing among works of comparable quality, prefer the more recent. Label foundational works as [Foundational] in the bibliography.

Source Rules

Include: peer-reviewed articles; books/chapters from major presses; review/handbook chapters; working papers passing the WP test.

Exclude: non-scholarly sources; predatory outlets; retracted papers (log if encountered).

Working Paper Test

Include only if ≥2 of: (a) high citations/downloads for age; (b) established author; (c) top venue presentation (note venue/year); (d) sole source for needed dataset/method; (e) verifiable forthcoming/accepted. For each included WP, state which ≥2 criteria it meets. Label as "Working paper / preprint." Prefer published version. No Confidence C WPs in must-read without explicit justification.

Quick Workflow Reference

The full 9-step workflow with substeps is in references/workflow-reference.md. Summary:

Tools and Fallback Modes

Try Mode 1 first. Fall back to Mode 2 only if Zotero MCP is unreachable. Log mode in Audit Log.

Zotero-first rule: Always search the user's Zotero library first before any external API or web search. Use the zotero-mcp-code skill (comprehensive search via SearchOrchestrator) for efficient, multi-strategy Zotero retrieval. Zotero results form the baseline; external sources then fill gaps. This ensures the user's existing collection is leveraged and not duplicated.

Deliverable Structure (Summary)

Full structure details in references/workflow-reference.md.

• A) Project-to-Literature Map (~1–2 pp): clusters, mechanisms, narrative synthesis, cross-cluster tensions
• A1) Project Positioning: gaps addressed, lines extended/challenged, contributions, scope conditions
• A2) Concept Dictionary: constructs, synonyms, operationalizations
• A3) Dataset & Measurement Alignment Table (if applicable)
• B) Annotated Bibliography (25–50 items): hyperlinked citations, pillar/confidence/screening tags, 3–6 bullets, connection notes
• C) Must-Read Shortlist (10–15): ≥80% Confidence A
• D) Coverage Report: sources, counts, confidence breakdown, limitations
• E) Audit Log: run date, queries, dropped/unlinked citations, retracted papers, validation pass log

Chat Response (after writing deliverable)

Report only: (1) file confirmation, (2) item count by pillar, (3) cluster count, (4) top 5 must-reads with DOI links, (5) confidence breakdown, (6) link coverage, (7) project positioning summary, (8) suggested next steps. Do not reproduce file content in chat.

Style

Concise, professional academic prose. Faithful, specific summaries. No hedging about existence ("likely argues…", "may exist…"). Every content claim grounded in verified text. Defer to user's definitions over generic domain knowledge.

Causal language discipline. Reserve causal verbs ("causes," "increases," "reduces") for findings from designs that warrant causal inference (experiments, RCTs, credible quasi-experiments). For observational/correlational work, use: "is associated with," "predicts," "correlates with." When summarizing a paper, match the strength of language to the design, not to the authors' own claims.

Disagreement language. State disagreements precisely: who claims what, on what evidence. Do not flatten into "the literature is mixed."

Accessibility. Write for a general political science audience. Define subfield jargon on first use. The survey should be usable as a draft lit review for a top generalist journal (APSR, AJPS, JOP).

Context Budget Management

A full run (Steps 0–8) can exceed context limits. Checkpoint intermediate state to _lit_review_workspace.md after Steps 2, 4, and 5. Process API responses immediately — extract needed fields and discard raw JSON. The deliverable file is the final checkpoint.

Cross-Project Reuse

Check memory for prior surveys before Step 0. Reuse verified items (tag [Prior survey: <filename>]). Re-verify DOI/URLs for carried-over items. Update published WPs. Separate file per project.

Memory

If persistent memory is available (e.g., agent memory directory or skill-level state), consult it at the start of every run and update at end of every run with: high-impact venues/authors; clusters and anchors; productive vs. unproductive queries; contested debates; WPs since published; cross-project links. Do not rely on the user to prompt a save.