Legal Research

One skill for finding and verifying legal authority , case law, statutes, and regulations. It produces a reliable, citation-anchored body of authority you can build an analysis or a document on, having personally confirmed it is still good law. Research here is thorough by default: assume the user wants the full picture, not a single snippet. Keep this file in context; pull in a references/ file for the method you need.

This skill runs on the DingDuff research tools (CourtListener + code databases): opinion_search, courtlistener_full_search, fetch_opinion_file, opinion_view, show_citing_opinions, show_related_opinions for cases; codes_browse, codes_search, codes_view, metadata_view for statutes/regulations.

CRITICAL: Retrieve before you assert (anti-hallucination)

Never state a holding, quote any language, or recite statutory text from a search snippet. Snippets only help you decide what to retrieve , they are not citable substance.

• Cases: retrieve the actual opinion via fetch_opinion_file (preferred; needs network egress) or opinion_view, then cite by reporter citation + pinpoint page (Miranda v. Arizona, 384 U.S. 436, 444 (1966)). Cluster IDs are internal retrieval handles , keep them in your notes, never in user-facing citations.
• Statutes: retrieve the full section via codes_view, then cite by proper statutory citation (Tex. Bus. Orgs. Code § 21.401(a); 18 U.S.C. § 1001(a)(2)). statute_id stays internal.
• If you cannot retrieve a source's text, do not discuss its substance , note the gap and move on.

CRITICAL: Verify it's still good law

The databases do not flag overturned, abrogated, or limited cases, and do not guarantee real-time tracking of statutory amendments. You must check, yourself, for every authority you will rely on , overturn/treatment checks for cases, currency/amendment checks for statutes. This is the single discipline that separates real research from a snippet dump. Do it before anything you find becomes part of an answer.

Case validation has two tiers, and choosing between them per case is your judgment call:

• Lightweight , a heuristic screen you run inline, a few tool calls per case. Cheap and fast. It catches the loud failures, but a clean screen is not evidence of validity. Right for background cites, redundant supports, and quick orientation.
• Heavyweight , the dingduff-validity-check skill, delegated to a subagent. Answers whether the case is good law for the specific proposition you're citing it for, and states its own coverage limits. Slow and sequential, so spend it where it counts.

The test: does the answer change if this case is bad? If yes, heavyweight , the anchor of a rule you assert, anything headed into a filing or client advice, adverse authority you're distinguishing, anything the screen flagged. Do not run heavyweight on every case in a network; screen broadly, escalate narrowly.

Then tell the user which cases got which tier. A screened case and a verified case look identical in a finished answer, so the distinction only exists if you state it. Labels and the reporting convention are in references/validity.md, along with both procedures and the delegation contract.

Research and analysis interleave (the recursive loop)

Research and analysis are one intertwined process, not two phases. Analysis guides research (you research the specific issues, elements, and ambiguous terms the analysis has framed) and research guides analysis (what you find reshapes the issues, the rule, and the framing , a new exception, a circuit split, a controlling definition, an issue that turns out to be moot).

So do not treat research as a single up-front pass. Loop:

1. Start from the framed issues (from dingduff-legal-analysis; if none exist yet, frame them lightly first).
2. Research them (cases and/or statutes, below).
3. Hand findings to analysis , let it synthesize the rule, analyze the facts, structure the proof.
4. Return to research whenever analysis surfaces a new question , an ambiguous controlling term needing its judicial gloss, a newly-spotted sub-issue, a missing element, a needed counter-authority, an adverse case to confirm.
5. Repeat until both the research network is saturated and the analysis has no unresearched gaps.

dingduff-legal-analysis escalates back to research when it hits a hole; this skill closes the loop by sending you forward into analysis and back again. Neither is done until both are done.

Calibrate scope to the question

Thorough by default , but set the breadth to the question. A regime-mapping, bet-the-matter, or appellate question gets the full sweep (deep citation network; complete statutory map). A tightly-bounded question (one well-defined hook, one defined term) gets a focused pass , don't expand a one-section lookup into a 100-case network. If a network grows past ~75–100 cases or ~50 statute sections, the issue was scoped too broadly: narrow it (to an element, jurisdiction, or fact pattern) and re-anchor. Rank what you find by weight of authority (binding > persuasive; higher court > lower; more recent and on-point > older and tangential) and spend your effort there.

Local storage: saved cases / saved statutes

When the user has connected a folder, create archive subfolders in it and save every retrieved opinion and section as markdown , a durable record for the user and a way for you to re-read without spending more tool calls.

mkdir -p "<user_folder>/saved cases"
mkdir -p "<user_folder>/saved statutes"

Naming: cases <cluster_id>_<short_case_name>_<reporter_cite>.md (e.g., 107252_miranda_v_arizona_384_US_436.md); statutes <jurisdiction>_<code>_<section>.md (e.g., us_18_usc_1001.md). Sanitize: lowercase, spaces → underscores, strip punctuation except underscores/dots, "v." → "v". If no folder is available, fetch_opinion_file still works , save into the working directory and tell the user no archive was produced. Fall back to inline retrieval (opinion_view / codes_view) only without network egress, which the returned download URLs require.

Workflow

1. Take the issues , from dingduff-legal-analysis, or frame them lightly. Decide whether the question is statute-driven, case-law-driven, or both.
2. Statutes & regulations (when a statutory or regulatory scheme governs) , map the code bidirectionally, pull definitions, flag phrases needing interpretation. → references/statutes.md
3. Case law , search broadly, then map the citation network (back to the anchor, forward to current applications, plus a topical sweep). → references/case-law.md
4. Validate everything you'll rely on , lightweight screen on every case, heavyweight (delegated dingduff-validity-check) on the load-bearing ones, currency checks for statutes. Report the tier each case received. → references/validity.md
5. Interleave with analysis (the loop above) , feed findings to dingduff-legal-analysis; return here for each new question it surfaces. Continue until saturated and gap-free.
6. Synthesize , present the authority with its current-validity status. Output formats (doctrinal genealogy + network diagram + case table for cases; statutory tree + roadmap + section table for statutes) are in the domain references.

References map

• references/case-law.md , finding cases (search strategies, tool patterns), mapping the citation network (ancestors / descendants / topical sweep), the working set, and the case-law output format (genealogy, network diagram, case table).
• references/statutes.md , jurisdiction/scope check, bidirectional code mapping (top-down hierarchy + bottom-up keyword), definition discovery, flagging phrases that need interpretation, targeted interpreting case law, and the statutory output format (tree, roadmap, section table).
• references/validity.md , the two validation tiers and how to choose between them; the lightweight screen; the delegation contract for heavyweight dingduff-validity-check; statutory currency and amendment checks; the verification pass and how to report tiers to the user.

Related skills

Interleaves with dingduff-legal-analysis (the reasoning it feeds and takes direction from). Heavyweight validation is delegated to dingduff-validity-check , always in a subagent, never loaded here. QC: dingduff-legal-citation-format, dingduff-citation-check.