--- name: brycewang-stanford/aer-literature source: https://app.decimal.ai/s/brycewang-stanford-aer-literature@1/SKILL.md source_sha256: 13b09f5db939 --- # AER Literature ## Overview This skill does two jobs that AER referees grade separately: 1. **Positioning** — finding the closest predecessors, mapping what each did and missed, and choosing the value-added claim the paper can actually defend. A mispositioned paper draws the fatal report sentence "the authors appear unaware of [paper]." 2. **Citation integrity** — guaranteeing that every reference exists, says what the manuscript claims it says, and is cited to the right version. The second job is **non-negotiable for AI-assisted manuscripts**. Language models produce plausible-looking citations that do not exist, attach real authors to invented titles, and attribute findings to papers that show the opposite. One hallucinated reference, found by a referee, destroys the credibility of every other sentence in the paper. The working rule: > **No citation is written from memory. Every entry is verified against a > fetched source before it enters the bibliography, and every claim about a > paper is verified against that paper's actual text.** ## When to Use - During `aer-topic-selection`, to run the novelty scan before committing - Before `aer-introduction`, to build the antecedents paragraph - When a referee names a missed paper and the positioning needs repair - Before submission, as the bibliography verification pass (pairs with `aer-consistency`) ## Search Protocol Run all five channels — each finds papers the others miss: 1. **Working-paper series**: NBER, CEPR, IZA, SSRN, BFI. Top-5 competitors live here for 2-4 years before publication; the novelty threat is almost always an unpublished working paper, not a published article. 2. **RePEc / IDEAS and Google Scholar**, searching by question ("effect of X on Y"), by method-in-setting, and by dataset name. Dataset search catches the most dangerous neighbor: someone using the same data. 3. **Forward citations** of the two or three obvious antecedents — everyone who cited them since, sorted by recency. 4. **Recent tables of contents**: AER, QJE, JPE, Econometrica, ReStud, AEJ family, ReStat, JEEA — last 3-5 years, scanning titles in the subfield. 5. **Handbook chapters and *Journal of Economic Perspectives* / *Annual Review of Economics* surveys** for the field's own map of itself. Log every search (channel, query, date, hits) — the log answers the referee who asks "did you know about X?" and feeds the novelty audit in `aer-topic-selection`. ## The Antecedents Map Reduce the search to the **3-6 closest papers** and fill this table honestly — it is the skeleton of the introduction's antecedents paragraph and the evidence base for the contribution claim: | Paper | Question | Data / setting | Identification | Headline finding | What it does not do | |---|---|---|---|---|---| | A (2019) | ... | ... | ... | ... | ... | Rules: - "What it does not do" must be a fact about the paper, not a slight. The introduction will say it in neutral language; referees include these authors. - If the closest paper answers the same question with the same data and a defensible design, the contribution claim must change — return to `aer-topic-selection` rather than writing around the conflict. - Verify each row against the paper's **published version** — abstracts and early working papers often differ from the final result. ## Positioning Moves The value-added paragraph claims one primary move (plus at most two supporting ones): | Move | Defensible when | |---|---| | New question | No antecedent asks it — and it matters that nobody has | | New identification | Antecedents are correlational or use a now-rejected design | | New data | Resolution, coverage, or linkage the literature lacked | | New mechanism | The *why* was open even though the *whether* was settled | | Opposite / corrected finding | A credibly identified sign flip or magnitude revision | | Unification | Two literatures explained by one framework or fact | Match the claim to the map: a "new identification" claim dies instantly if row 3 of the antecedents table already used the same design. ## Citation Coverage Norms - A full-length empirical AER paper typically carries **50-110 references**; *AER: Insights* far fewer. Padding is as suspicious as thinness. - Cite: every antecedent in the map; the primary source for every estimator and diagnostic (keys in `docs/methods-reference.md`, entries in `references.bib`); the source for every institutional fact and every benchmark magnitude used in Results. - Do not cite: textbooks in place of original methods papers; surveys as stand-ins for the specific result being invoked; papers included only to flatter potential referees. - Cite the **published version** wherever one exists. Citing the 2019 NBER version of a 2021 AER paper signals a stale literature review. Cite a working paper only when it is genuinely unpublished — and re-check its status at submission time. ## Citation Integrity Protocol Run this pipeline for **every** reference. No step is skippable for entries the model "remembers." ### Step 1 — Existence Verify against an authoritative record: resolve the DOI (Crossref / doi.org), or match the exact title and author list on the journal page, RePEc, or the NBER/SSRN abstract page. A reference with no resolvable DOI and no locatable landing page does not go in the bibliography. Use a web search or fetch for every entry; if a Zotero or reference-manager connection is available, import from the verified record rather than typing fields. ### Step 2 — Field accuracy Check, character by character, against the verified record: full author list and order, spelling and diacritics, year, exact title, journal, volume, issue, pages, DOI. The most common AI corruption modes: swapped coauthors, off-by-one years (working paper vs. publication), a real author pair attached to an invented title, and the wrong journal in the right family (AEJ: Applied vs. AEJ: Policy). ### Step 3 — Claim accuracy For every in-text sentence of the form "[Author (year)] find/show/argue X," read at least the abstract — and the relevant table if a magnitude is quoted. The claim must match what the paper actually establishes, including sign, population, and design caveats. Misattribution to a *real* paper is more damaging than a fake one, because the referee may be its author. ### Step 4 — Status Check that the paper has not been retracted or corrected, and whether a cited working paper has since been published (then cite the journal version). For replication-contested results, cite the dispute or choose a different benchmark. ### Step 5 — Ledger Maintain a verification ledger alongside `references.bib`: ```text KEY DOI-OK FIELDS-OK CLAIM-OK VERSION CHECKED goodman_bacon_2021 yes yes yes published 2026-06-12 hjort_poulsen_2019 yes yes yes published 2026-06-12 ``` Any row that cannot reach all-yes is removed from the manuscript or explicitly flagged to the user as unverified. **Never silently keep an unverified citation.** ## BibTeX Hygiene - One key convention, applied everywhere: `author_author_year` lowercase (matching `references.bib` in this repository). - Every entry carries a DOI, or an explicit no-DOI note with a stable URL. - No duplicate entries under variant keys; no fields invented to fill gaps. - In-text citations compile against the bibliography with zero unresolved keys — `aer-consistency` runs this check as part of the full-manuscript audit. ## Common Failure Modes - A bibliography assembled from model memory and "cleaned up" afterward — verification is the assembly step, not a post-process - The novelty scan run only on Google Scholar, missing the NBER working paper that scoops the contribution - "X et al. (2020) show that..." where the paper *suggests* it in the discussion section - Antecedents described from their abstracts, then misrepresented in the introduction's value-added contrast - Citing five surveys to establish a fact one primary source established ## Repository Resources Bundled with the installed skill, no repository checkout needed --- read it before the repo resources below: - `references/citation-verification-protocol.md` --- fetch-before-cite protocol, claim-audit table, BibTeX hygiene When working from the AER-skills repository or plugin bundle, load only the relevant resource: - Verified estimator and diagnostic citations with BibTeX keys: `docs/methods-reference.md` and `references.bib` - Recent papers by subfield for positioning benchmarks: `examples/modern-aer-exemplars.md` - Classic contribution archetypes: `examples/aer-exemplars.md` - The novelty-audit questions this skill's search must answer: `skills/aer-topic-selection/SKILL.md` ## Handoff ```text CLOSEST PAPERS MAPPED: (target 3-6) SCOOP RISK: POSITIONING MOVE: REFERENCES VERIFIED: / (existence / fields / claim / version) UNVERIFIED ENTRIES: NEXT SKILL: ``` ## Anti-Patterns - Writing the antecedents paragraph before building the antecedents map - Treating citation verification as a formatting task to do "at the end" - Keeping a citation because it "is probably real" — probability is not the standard; a fetched record is - Softening "what it does not do" into vagueness, leaving the value-added claim uncontrasted - Citing the referee-bait paper without reading it — referees test exactly those sentences