科研 Skill

深度法律研究备忘录

legal-research

Fans out multiple agents across free US legal sources, verifies every authority against CourtListener/Justia/Google Scholar, and returns a polished memo. Use for US case-law research.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Legal Research

Approximates a deep-research run, specialized for law, and tuned to be efficient by default. It runs a retrieval pipeline over free / open-web legal sources only (no Westlaw/Lexis); the stage-by-stage shape, caps, and model map live in step 4, next to the invocation. The design keeps token use down — the dominant cost in naive fan-out is a frontier model reading every fetched page — while still preventing hallucinated citations. Add deep (e.g. "do a deep pass") for high-stakes work where thoroughness beats speed.

Sources (open web)

Case law: CourtListener, Justia, Google Scholar case law, Cornell LII, Caselaw Access Project, official court sites. Scholarship: SSRN, Google Scholar, law-school digital-commons repositories, HeinOnline open content, journal sites, faculty pages. Scholarship modes also search the user's own Zotero library via the Zotero MCP server (skipped gracefully if unavailable).

How to run

1. Parse the request

Extract from the user's prompt:

  • question — the legal question (required).
  • mode — caselaw, scholarship, or both. Default both. Infer from wording ("what do courts hold" → caselaw; "state of the literature / scholarly debate" → scholarship).
  • outputStyle — practitioner (Question Presented / Brief Answer / IRAC) or academic (intro / landscape / analysis / footnotes). Default: practitioner for caselaw, academic for scholarship and for both. Honor an explicit override.
  • jurisdiction, asOfDate (default today), and any notes (facts, posture, scope).

2. Clarify only if genuinely blocking

If the question is too vague to research well, ask 2–3 targeted questions before launching. For caselaw, jurisdiction and procedural posture matter most; for scholarship, the specific thesis or sub-question and any time window. Don't over-ask — pick sensible defaults and note them.

Enumeration questions are a different job (2026-11-16). A question that asks for every member of a set ("every opinion Justice X has written on Y", "all the cases citing Z", "each circuit's rule") cannot be answered by this pipeline as asked: it triages to a fixed number of full reads (8 lean, 16 deep) however many candidates the search finds, so the memo comes back describing a sample. On the first live run of this kind the search found 42 candidates, 8 were read, and the findings about frequency and trend were artifacts of the 8. Before launching, tell the user so in a sentence or two and settle which of these they want:

  • List first, then read each (the real answer). Build the list from a source that can enumerate (a CourtListener author or citation filter, a database or lead list the user already has), show the count, and read every item with one agent per item outside this workflow, giving a cost estimate before launching. Then tabulate.
  • A narrower doctrinal question this pipeline can answer from leading cases ("how does he use the common-law canon"), with the memo titled and scoped as leading cases, not a survey. If the user wants the pipeline run anyway, pass the lead list in notes and say in the delivery message, first, how many of how many were read.

3. Pre-flight: confirm web tools and the CourtListener token (do this BEFORE launching)

The research subagents call WebSearch and WebFetch heavily. If those tools are not pre-allowed in the current agent, each call can raise an interactive permission prompt — and an interrupted prompt can kill a research agent mid-run (a real failure mode: a prior run lost 2 of 4 vectors this way).

Before calling Workflow, make the run self-contained:

  1. Confirm WebFetch/WebSearch are available in the current agent without an interactive prompt on every call. Never edit the agent's own configuration or permission files yourself; if web access would prompt on every call, tell the user and let them decide whether to pre-approve it through the agent's normal permission controls.
  2. If the current agent still raises a per-call prompt mid-run and that can't be pre-approved, tell the user up front that they'll need to accept a prompt each time, rather than silently risking a lost run.
  3. Optional but worth a glance: deterministic case-cite verification needs a CourtListener API token. Test it, don't go looking for a file — run python3 ../legal-legal-research/check_cite.py "347 U.S. 483"; exit 0 with "found": true means the token is working and there is nothing to report. check_cite.py resolves the token from COURTLISTENER_API_TOKEN, then COURTLISTENER_TOKEN, then the file ~/.config/courtlistener/token — in that order, and setting one of those two environment variables in the current agent's own environment configuration (its env/config mechanism — in Claude Code, the env block of its settings) is the normal install, so the absence of the file proves nothing. If the probe actually fails, don't block the run — verify agents fall back to web checking — but mention it once in the deliver step (token page: https://www.courtlistener.com/profile/api-token/).

4. Launch the workflow

Call the Workflow tool with the bundled script and the parsed args:

Workflow({
  scriptPath: "../legal-legal-research/legal-research.workflow.js",
  args: {
    skillsDir: "..",      // where the bundled helper scripts live
    question: "<the legal question>",
    mode: "caselaw" | "scholarship" | "both",
    outputStyle: "practitioner" | "academic",
    jurisdiction: "<e.g. federal / 9th Cir. / Missouri / U.S. Supreme Court>",
    asOfDate: "<YYYY-MM-DD or 'today'>",
    notes: "<facts, posture, scope, anything the user emphasized>",
    deep: false,  // true only when the user asks for a deep/thorough/high-stakes pass
    finalizeModel: "opus",  // ONLY when this session's model ranks above Opus (Fable, Mythos); otherwise OMIT the line
    projectDir: "<the current working directory — enables the local-corpus survey vector;
                  omit only if the cwd is clearly unrelated to the question>"
  }
})

Lean by default: ~4 search-only survey vectors (plus a Zotero-library sweep in scholarship modes and a local-corpus sweep of projectDir — prior memos and downloaded sources are maps and leads, never citable authority). Scope also classifies the question (doctrinal vs. scholarly; historical or not) — that selects an analytical-method brief injected into Draft (synthesis discipline, holding/dicta hygiene, era-spanning taxonomy for doctrinal-history questions) and, for historical questions, spreads the search vectors across time layers so coverage doesn't cluster on the modern canon. Direct-answer coverage cell (scholarship/both, added 2026-07-16 after a Codex comparison run beat a 64-agent Claude run on the actual question): Scope must dedicate at least one vector to scholarship treating the question's NAMED PHENOMENON as its own object — built from the specialist title terms an expert expects and including a citation-chase of the most on-point piece — kept distinct from the adjacent literatures the phenomenon merely belongs to. Survey flags such central candidates; triage promotes them to full reads ahead of broad adjacent surveys. The failure this fixes: mapping five adjacent literatures (extrajudicial speech, recusal, celebrity) while missing the handful of works squarely about the phenomenon (e.g. judicial-autobiography studies, "justices' book deals" ethics sections), then mistaking that corpus gap for a finding that "no direct scholarship exists." Absence/gap claims are now gated by a central-source omission audit (in the scholarly method brief): no "no scholarship directly addresses X" unless a phenomenon-targeted search AND a both-directions citation-chase actually ran — otherwise the claim is scoped to "in the sources reviewed" and the un-run search is flagged. Fan-out is organized around coverage cells, and the agent cap is a ceiling, not a quota — nonoverlapping lanes beat many overlapping generalist searches. Pipeline: survey vectors → dedup → ≤8 unique full reads total (not per-vector; readers fall back to open-access mirrors — the journal's own site, author pages, CourtListener/LII — before giving up, test every candidate URL in one deterministic call first (probe_sources.py: HTTP status, known-host rules for SSRN/bepress/Hein/sign-in walls, Wayback snapshot), route PDFs through a hard size-capped fetcher, fetch_page.py, and extract posture/limitation/separate-opinions/significance alongside the proposition and quote) → one session-model draft (guided by the method brief) → a bounded gap-fill (≤3 targeted retrievals plus one retry wave, only if the draft flags missing authorities) → chunked cite-checks of just the cited authorities (8 per agent; case cites confirmed deterministically via CourtListener's citation-lookup API, check_cite.py, plus a systematic citing-opinions treatment scan via cl_case_tools.py citing — a relevance-ranked treatment-language probe over the case's actual citing opinions by TWO routes, the citation graph and a full-text search for the case name, deduplicated to distinct cases, with a liveness date; a cite that will not resolve goes through cl_case_tools.py find (caption search + opinions citing the string) before it is called unconfirmed; load-bearing lower-court cases also get cl_case_tools.py history, the RECAP docket and later same-litigation opinions; sources a reader already opened this run skip existence re-checks) → light finalize (returned as edits the script applies; see the Finalize model note below), whose Table of Authorities carries a Validation column per authority (3-voter refutation panel / checked by existence + treatment probe / not validated) and one sentence naming which authorities got the panel, because a screened case and a verified case look identical in prose. Unselected candidates AND the authorities each read source itself relies on are carried into the draft as hedged unread leads, so breadth stays wide while fetches stay few. Every run is spend-guarded: a hardwired soft ceiling (~250k output tokens lean, ~1M deep — raised from 600k on 2026-07-15 after a deep historical run bound mid-ladder and left 35 of 43 citations unchecked) arms the same graceful-degradation ladder as an explicit +500k budget — trim reads → skip gap-fill → verify only load-bearing cites — so a runaway run degrades instead of spiraling. Read, gap-fill, and verify agents all carry an access-discipline block (added same date): hard attempt caps, WebSearch/WebFetch only, and an explicit ban on browser automation, library proxies, and SSO/login flows — a paywalled source is an honest opened=false, not a reason to grind. Models: Haiku only for the search-snippet survey; Sonnet for triage, full reads, and verification; Scope and Draft inherit the session model (Fable when Fable is on), and so does Finalize unless you pass finalizeModel. Finalize model (2026-11-16). Nothing is hard-coded in the script: with no finalizeModel a session on Sonnet or Opus runs every session-model stage on that model. Pass finalizeModel: "opus" only when your own model ranks above Opus (Fable or Mythos, whose use is capped); never pass a model that ranks above the one you are running on. This is safe because no stage after Draft retypes the memo any more: the deep calibration pass, Finalize, and the call-in below each return find-and-replace edits that the script applies, so text no edit names stays the Draft's own words, and the Table of Authorities and Limits sections come back as separate fields. With finalizeModel set, the changes that alter what the memo can claim are not left to the cheaper model: a load-bearing authority that failed verification (or that no checker examined), negative treatment of one, an authority verified over a minority objection, and every gap-fill authority (woven in as a full member of the memo: a table row, updated counts and title, a footnote; one the cite-check did not confirm is woven in with that stated, never described as unread) go to one call on the session model first (the call-in, finalize:revise). It runs only when the cite-check results contain such an item; escalate: false turns it off, and if it fails or is squeezed out by the budget the Finalize pass applies those items itself. The result's finalize block says which model ran Finalize, whether the call-in ran and for which authorities, and lists any edit that could not be placed (unappliedEdits). Effort is pinned per call in the script (low for search, reads and cite-checks; medium for Scope, the refutation voters, the call-in and Finalize; high for Draft and calibration), so nothing inherits the session's effort setting. deep:true (or a +800k-style token budget) widens to up to 7 vectors, raises the read cap to 16, and gives each load-bearing citation (capped at 8) a 3-vote adversarial refutation pass (re-opens sources to confirm quotes verbatim; the three voters lead with different routes — the record, the passage in context, the strongest adverse reading — meant to reduce correlated omissions; all three still run on the same model, so this is not independence, and a refutation the majority outvotes is carried into the memo as a disclosed objection rather than erased). The workflow runs in the background and notifies you on completion.

5. Deliver

The workflow returns { memo, exitState, degradation, counts, finalize, ... }.

  • Check exitState BEFORE presenting. complete → deliver. complete_with_limits → deliver, surfacing the documented gaps (listed in degradation and the memo's Limits section). Since 2026-11-15 those gaps include agent failures that used to leave no trace: a search vector that did not run (counts.surveyLanesFailed; say which, and treat any absence claim that depended on it as untested), a selected source nobody read (counts.readsFailed), a citation correction only a minority of the panel proposed (counts.panelDisputedCorrections; not applied, listed in Limits), and panel voters who reported different treatment for the same case (counts.panelTreatmentSplits; every report is in the memo, and the case needs a citator check). budget_exhausted → the budget guard cut actual coverage (trimmed reads, skipped gap-fill, unchecked cites — see degradation for specifics); the memo is not yet the deliverable. The proven rescue pattern (July 2026 pre-1960 run): launch targeted read-and-verify Agent passes — one per uncovered area, each given the specific cases/quotes to fetch and confirm against official reporter text (CAP static.case.law / LOC usrep scans) — then write the final memo yourself from the workflow's verified core plus the fill passes. Do not hand the user a memo that says "Confidence: insufficient" for half its scope without either filling the gaps or asking. Rescue agents need the same access discipline as workflow agents (lesson from a July 2026 scholarship run, where an uncapped rescue agent burned 30+ minutes and ~245k tokens escalating to browser automation against a paywalled 1962 article): write into every rescue prompt a hard attempt cap (~3 per source), WebSearch/WebFetch only, an explicit ban on browser automation / library proxies / SSO flows, and the instruction that "paywalled — here is the strongest basis I reached" is a fully acceptable answer. If a rescue agent grinds anyway, SendMessage it to wrap up and report rather than waiting it out.
  • Save the memo with the script; never retype it (2026-11-16). The memo is the primary deliverable, and it is delivered as a file. The completion notification names an <output-file> holding the workflow's result; copy the memo out of it: python3 ../legal-legal-research/save_memo.py "<output-file>" "<working dir>/<Title>.md" --header "<Topic/Date/By block>" --footer "<Production details block>". The script writes the memo byte for byte, puts the header under the title and the footer at the very end after a --- rule (see the header/footer rule below), refuses to overwrite an existing file, and prints a summary without the memo's text. If the draft wrote its own By: (or Topic: / Date:) line under the title, the header's line replaces it and the summary lists it under replacedHeaderLines; nothing else in the memo is touched. Do NOT paste the memo into chat and do NOT write it out with the Write tool: on a metered session model, retyping a finished memo is the largest avoidable cost of delivering it, and a retyped memo can drift from the one that was cite-checked. In chat, give the path, the bottom line in a few sentences, the verification counts and the limits. If the script exits 1 (no memo found in the output file), fall back to writing the memo from the notification's inline result. Later changes to the saved memo are made with Edit, never by rewriting the file.
  • Check the memo's quotations by script (2026-11-16). In lean mode the cite-check confirms that each authority exists and is cited correctly; it does not re-open an opinion to confirm the words quoted from it or their page. After saving, run python3 ../legal-legal-research/check_memo_quotes.py "<saved memo>" --sources "<project dir>" --fetch --append. No model reads anything: the script takes each quotation and the citation it hangs on out of the memo, finds a copy of the source (a local PDF or text file whose name carries the reporter cite, else a public copy: a URL in the memo's footnote, the Caselaw Access Project page for a U.S. Reports cite, CourtListener's API text), runs the cite-check skill's deterministic verifier, and writes an addendum table into the memo ahead of the Production details. Omit --sources when the project has no local copies and --fetch to stay offline. Then act on its JSON summary, with Edit, never by retyping the memo:
    • pinDiffers: the quotation is in the source on a different page from the memo's pin. Correct the pin and say in the footnote where the page comes from. A page 'counted' from a local PDF is right only if the PDF starts on the report's first page; CourtListener text of a recent opinion usually has no reporter pages at all, and then no page is reported.
    • notFound: a prompt to look, not proof of a misquotation. Read the hint (the closest passage in the source). OCR debris in the copy and a footnote falling inside the quoted sentence both defeat the match; a changed word is a real finding. Fix the quotation, or tag it as unconfirmed, and add a line under the addendum saying which it was.
    • foundElsewhere: the words are in a different source the memo cites. For a table cell with no footnote that is the expected result; for a footnoted quotation it means a quotation inside a quotation or a wrong footnote, and the sentence needs a look.
    • noSource: say in the delivery message which cases could not be checked. Paraphrases and their pins are not checked by this. Give the counts in the delivery message.
  • Apply any unapplied edits before delivering. If the summary's unappliedEdits is not empty (counts.editsUnapplied > 0), the checking passes called for changes whose target text could not be located. Each entry has find, replace, why and the reason it failed. Open the saved file, make each change by hand with Edit, then remove the 'Changes not made automatically' note from the end of the Limits section. One of them may be the caveat on an authority that failed verification, so do not deliver with any outstanding.
  • Say how Finalize ran in Production details: the Finalize model, and whether the session-model call-in ran, was not needed, or failed (finalize.callIn). If it reads failed or skipped (budget) while finalize.callInAuthorities is not empty, the cheaper model made the changes for those authorities: re-read the memo's treatment of each one yourself before delivering, and say so.
  • Hyperlinks (required, per the author, 2026-08-29): every news story, blog post, or other web source in the memo must carry its working URL as a clickable link — never an outlet-name-only attribution like "(Politico)" or "(Roll Call, July 14, 2026)". The workflow's citation block enforces this in agent prompts; hold hand-written edits, rescue passes, and any ad-hoc (non-workflow) research memos to the same rule, and instruct any rescue/fill agents to return the URL for every source they cite (a claim with no URL is unverified).
  • Research memo header & production footer (required, per the author — global rule in ~/.claude/CLAUDE.md): the memo opens directly under the title with the structured header block (Topic: / Date: / By:); the Production details: block goes at the very end of the file (after all appendices, incl. "Limits of this research", set off by a --- rule) — never up top. Date is the local date/time the memo was delivered/written; By is Claude-<orchestrating session model> (exact id); Production details names the skill, the workflow-agent model(s) actually used (read from the run's per-agent progress entries — don't assume), the wall-clock duration, and the token count from the completion notification's <usage> block (labeled subagent output tokens; orchestrator overhead not included). If you add a billed total from the agent transcripts, count each response ONCE: a response is logged on one line per content block and every line repeats its usage, so group by message.id before summing (a per-line sum doubled the figures in the first memo that reported one, 2026-11-16). If rescue/fill passes ran after the workflow, add their agent counts and note the duration covers the workflow run only.
  • Surface the verification counts ("N authorities verified, M unverified") so they know how much to trust it, and point to the memo's "Limits of this research" appendix (which now discloses each authority's verification basis: direct-full-text / authoritative-reproduction / secondary-corroboration / citation-only).
  • The result also carries skeleton — the citation-tagged analytical outline the draft was written from (each load-bearing claim tagged to its finding, holes marked GAP). Don't present it by default; offer it if the user wants to audit the reasoning, and use it yourself when editing the memo by hand (a claim with no skeleton line needs an authority or a hedge before it grows).
  • The Markdown file is always saved (step above). Offer the other formats: Word via the bundled converter — python3 ../legal-legal-research/md2docx.py memo.md memo.docx "<Title>"; or Google Drive via the connector. If the user asked for .docx up front, just produce it.
  • Offer a follow-up pass on any weak/unverified point, or an output-style switch.

Claim calibration (v7)

Cite-verification and claim-calibration catch disjoint error classes: verification catches false statements about sources; calibration catches true findings assembled into unsupported generalizations. The pipeline now enforces both — the Draft prompt carries a claim-strength gate (frequency/priority/finality words like "first", "only", "overwhelmingly", "erased" require a search designed to support them, else sample-scoped language), readers tag each source's posture (including arguendo and nonprecedential dispositions) and bearing (process-only / assumes the premise / adjacent doctrine), historical runs get selection-bias, boundary-date, and harm/exclusion checks, and deep runs add a dedicated adversarial calibration pass on the draft. Two further rules ride in the same brief (added 2026-11-03): repetitions of the same underlying evidence are not independent corroboration and a head count is not a confidence level (confidence is assessed from authority, verification basis, independent contributions, contrary evidence and coverage); and for each material contested issue the counter-position is stated at full strength, as its proponent would put it, never as the conclusion re-hedged. Finalize then runs a consistency audit across sections — Brief Answer, Discussion, Conclusion and Limits must tell one story, and the summary sections are fixed to match the Discussion and the cite-check results, never the reverse. A third error class sits between them: a central-source omission — a true "we didn't find it" mistaken for "it isn't there." The direct-answer coverage cell and the omission audit (above) guard this: an absence claim about a FIELD requires a search designed to find the thing (specialist title terms + both-directions citation-chase), not the observation that adjacent abstracts stayed silent. When editing or extending a memo BY HAND, hold your own prose to the same gate — and before writing "no scholarship directly addresses X," run the phenomenon-named search yourself.

Prerequisites

  • Step 4 ("Launch the workflow") calls a Workflow tool with a bundled legal-research.workflow.js script — Claude Code's dynamic-workflow / multi-agent dispatch feature. In an agent without an equivalent tool, there is no drop-in substitute for this script; run the same stage shape (parse → clarify → pre-flight → search/screen/verify/compile, as the script's stages are summarized in this file) as ordinary sequential steps in the current session instead of expecting the Workflow call to succeed.
  • A CourtListener account/API token (free) for deterministic cite verification; see the setup skill. Without it, cite checks fall back to web search, which is slower and less certain.

Limitations

  • US jurisdiction only. Every source in "Sources (open web)" — CourtListener, Justia, Cornell LII, the Caselaw Access Project, SSRN, HeinOnline open content — is a US case-law or US-centric scholarship source. This skill has no coverage of non-US jurisdictions (including Chinese law; see this plugin's legal-issue-research skill for Chinese-law research instead) and does not state that scope anywhere in its own text.
  • Coverage is narrower than Westlaw/Lexis and citator flags (KeyCite/Shepard's) are only approximated — already stated under Guardrails, repeated here because it is the single most load-bearing caveat in the whole skill.
  • Not legal advice; final cite-checking and Bluebooking (a US-specific citation format) remain the user's responsibility.

Guardrails

  • No hallucinated law. Every authority the memo cites goes to a cite-check, and the Table of Authorities says what each one got. That is not the same as "every cite was independently confirmed": an authority a checker could not confirm stays only with an explicit caveat, one the budget left unchecked is tagged as such, and a source a reader opened this run is checked for citation form and treatment but its quote is not re-read (its basis is reported as a reproduction resting on the research read). In deep mode the load-bearing authorities also face the refutation panel. Preserve that discipline; never add cites from memory when relaying or editing.
  • Open-web only. Coverage is narrower than Westlaw/Lexis and treatment flags (KeyCite/Shepard's) are approximated, not authoritative. Always convey this caveat.
  • Opinions, articles and dockets that postdate the model's training data are real. A 2025 or 2026 decision nobody recognizes is not a hallucination signal; the fetched record is.
  • Say which tier each authority got. When editing a memo by hand, keep the Validation labels in the Table of Authorities and never let an unlabeled list read as "all verified".
  • Not legal advice. This is research support; final cite-checking and Bluebooking remain the user's responsibility. Recommend a paid-database pass before anything is filed or published.

示例提示词

帮我针对这个法律问题做一份多方来源验证的判例和学术综述备忘录。

所属插件

法学与知识产权面向法学研究与知识产权:中国法律检索、法学研究备忘录、引注查找与核验、案例准备,以及专利新颖性扫描与验证。

安装该插件会一并获得同领域的 13 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

深度法律研究备忘录能做什么?

Fans out multiple agents across free US legal sources, verifies every authority against CourtListener/Justia/Google Scholar, and returns a polished memo. Use for US case-law research。

如何安装深度法律研究备忘录?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「深度法律研究备忘录」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

深度法律研究备忘录需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

深度法律研究备忘录经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。