danepps/legal-research-skills

Find Cite

A lightweight sibling of legal-research for filling out a footnote. The deliverable is a short, ranked list of 2–4 verified authorities for one proposition — not a memo. Run it inline (no Workflow): a handful of searches, roughly ≤6 page fetches total, minutes not tens of minutes. If the request turns out to need real synthesis across many authorities or an answer to an open question, say so and offer the full legal-research skill instead.

1. Parse the request

• proposition — the exact sentence or claim needing support (required). If the user pasted draft text with a footnote call, the proposition is the sentence the footnote hangs on, not the whole paragraph.
• authority type — read it off the kind of proposition, not off a default. The ideal answer is a few of each when the proposition genuinely supports both; don't force a category that doesn't fit.
  • Pure doctrine ("the Court has held…", black-letter statements, vote lineups) → cases; add a scholarly cite only if a well-known article states the rule crisply or is the standard synthesis.
  • Claims about the literature, theory, or empirics ("scholars have argued…", "the conventional account is…") → scholarship; a case belongs only if it's the doctrinal anchor the literature is reacting to.
  • Characterizations of doctrine or the Court ("the Court has grown skeptical of…", "this doctrine is famously incoherent") → both: cases as the primary evidence, plus a scholar making the same observation.
  • Framing/background claims → both, lightly — one or two of each. If the user says which they want, that controls. When the mix is a judgment call, note in the answer why you leaned the way you did.
• constraints — era, "the canonical cite," "something recent," a particular justice or scholar, jurisdiction other than SCOTUS. Honor them.

Don't ask clarifying questions unless the proposition itself is ambiguous in a way that changes what you'd search for. Sensible defaults, noted in the answer, beat a round-trip.

2. Search (lean)

Cases. Prefer the authority that controls in the proposition's jurisdiction and supports it directly. For federal constitutional and SCOTUS-focused work that is usually the Supreme Court; a state-law or circuit-specific proposition gets its own controlling authority first, not a looser Supreme Court analogy. For well-worn propositions there is usually a canonical cite (the case everyone cites); find that one first — WebSearch phrasing like "<proposition keywords>" Supreme Court held or searching the proposition's signature quote works well — then one or two reinforcing or more recent cites. Sources: CourtListener, Google Scholar case law, Justia, Cornell LII, official slip opinions.

Scholarship. Search the Zotero library first (mcp__zotero__zotero_search_items) — the user may already own the right article, and a library record gives reliable bibliographic metadata. It is still only a candidate: open the text for any pincite or quote. Then Google Scholar, SSRN, and law-school digital commons. Ranking:

1. Recognized scholars in flagship general law reviews (Harvard, Yale, Stanford, Columbia, Chicago, NYU, Michigan, Penn, Virginia, California, Georgetown, Duke, Northwestern, Cornell, Texas, Vanderbilt) and top peer-edited journals (Supreme Court Review, Journal of Legal Studies, Journal of Legal Analysis).
2. Well-cited work in strong secondary or specialty journals.
3. Student notes — last resort only. Cite one only when nothing better exists, and label it as a Note in the answer so the user can decide.

Fit comes first: a piece squarely on the proposition outranks a famous one that is merely adjacent. Among sources that genuinely fit, prefer the more influential author and venue; Google Scholar citation counts are a fine proxy when unsure.

Parallelize when it helps. For both mode, two concurrent general-purpose Agent sweeps (one cases, one scholarship, each told to return candidates with URLs and the supporting language) beat serial searching — but keep each agent lean and capped. For a single-type request, just search inline.

3. Verify (non-negotiable)

An unverified cite never ships — that rule is identical to the full skill's.

• Cases: run every candidate through the deterministic lookup: python3 ${CLAUDE_PLUGIN_ROOT}/skills/legal-research/check_cite.py "<vol> U.S. <page>" … (falls back to web verification if no CourtListener token). Then open the opinion and confirm the passage actually supports the proposition, at the pincite, in the majority (not a dissent/concurrence unless flagged). Pincite to official U.S. Reports pagination where available (CourtListener, CAP static.case.law, LOC usrep scans).
• Scholarship: confirm the article exists and get the real journal/volume/first-page from the actual source (Zotero record, journal page, SSRN, digital commons, HeinOnline citation page). If quoting or giving a pincite, open the text and copy verbatim. If only the article's existence and general thesis can be confirmed — not the specific page — deliver it without a pincite and say so.
• Anything that fails verification is dropped or explicitly flagged, never smoothed over.

4. Deliver

Short and footnote-ready, best cite first:

• Full Bluebook cite with pincite for each authority.
• A clickable URL to the source consulted (CourtListener/CAP opinion, SSRN/repository, journal page) after each cite — per the author, sources cited without links are "not helpful".
• A parenthetical or verbatim quote (quotes only if actually copied from the source — paraphrases go unquoted).
• One line on fit (canonical cite / recent reaffirmation / leading article / student note — only authority found).
• Verification basis in a word or two, honest about how thin it is: opened at the pincite is full verification; CourtListener + snippet or Zotero + journal page confirms existence and metadata only — label such a cite partially verified and give it no pincite you did not see.

Then be honest about gaps: if no SCOTUS case squarely supports the proposition, say what the closest authority is and how it falls short — a candid "the strong version of this claim isn't citable" is a useful answer. Deliver the cites as paste-ready text; never edit the user's draft unless they explicitly ask you to insert them.

Guardrails

• All global authority rules apply: never invent a case, cite, quote, holding, or authorship; quotes are verbatim or marked as paraphrase; uncertainty is flagged plainly.
• Open-web coverage only (no Westlaw/Lexis); no authoritative KeyCite/Shepard's — for a case doing load-bearing work, glance for negative treatment and say if you didn't.
• Suggest signals (see, see, e.g., cf.) only tentatively — signal choice is the user's call.