# Legal Research — practical answers and attorney-prep briefs
Shared ChiefOS skill conventions live in `references/skill-operating-standards.md`; Tree D placement is owned by `references/output-paths.md`, and citation discipline follows the source rules in this skill plus `references/client-deliverable-standards.md` section 4.
This skill has two jobs: answer an ordinary legal-risk question usefully in chat, and prepare a deep attorney-prep brief when Patrick explicitly requests one. In both modes, ground decision-changing legal claims in current primary law rather than blog paraphrases.
What this skill is — and is not
**It is not legal advice, and it is not a substitute for a lawyer.** It may still give a direct risk-management recommendation such as "do not process live personal data until the contract gap is fixed." Label that as the practical course, not a definitive legal conclusion. Say what is known, what is a reasonable inference, and what remains a question for counsel.
**What makes it better than a generic summary** is the thing to protect at all costs: it rests on **primary law**. A generic answer paraphrases a law-firm alert that paraphrased a statute; every hop drops a nuance. This skill goes to the actual code section, reads it, quotes the operative language, and links the free official text. Secondary sources (blogs, firm alerts, legal-info sites, AI-generated "legal" content) are leads and plain-English explainers only — they never stand in for the primary source. If a decision-changing claim cites only blogs, you are not done: find the statute, regulation, or case the secondary source is summarizing.
Choose the mode
Decision mode — default
Use decision mode unless Patrick explicitly asks to **create or save a research brief, memo, or durable artifact**, or uses `--brief`. Asking for a brief answer, asking to review an input document, or merely mentioning a memo/document stays in decision mode. Ambiguous requests stay in decision mode. A serious issue changes the urgency and recommendation; it does not silently authorize a document, Drive, or git write.
Research only the authority needed to support the practical answer. Ask a clarifying question only when the missing fact could materially reverse the recommended action. Otherwise state the assumption and proceed.
Reply in chat, normally in 250–400 words and never as a long research memo. Use this order:
1. **Bottom line** — the honest practical read in one or two sentences. 2. **What I would do now** — the lowest-regret next action, including a safe interim option. 3. **Real exposure and tradeoffs** — distinguish the likely operational or contractual consequence from lower-probability statutory, regulatory, or litigation exposure. Do not dump every theoretically relevant law on the reader. 4. **What could change the answer** — two to four concrete facts, not a generic caveat list. 5. **Question for counsel** — at most one tightly framed question, and only when professional review would materially improve the decision.
Include a compact "not legal advice" sentence. Cite primary authority inline next to each decision-changing legal claim and re-check each URL against the claim before replying. A short answer may have only one or two citations; citation volume is not a proxy for usefulness.
Decision mode does **not** pick a slug, create a file, commit, or offer a Drive mirror. If there is an active dispute, regulator contact, imminent deadline, threatened claim, or other urgent matter, say plainly that counsel should be contacted promptly and give the safest immediate operational step. Offer a full attorney-prep brief, but do not start one without an explicit request.
Brief mode — explicit request only
Use the workflow below only for an explicit brief request. Preserve the full primary-law research, citation-verification, Tree D artifact, disclaimer, and attorney-question payload.
Brief mode: load context
Read these first so you research against real conventions, not guesses:
1. **`references/legal-source-hierarchy.md`** (in this skill) — the authority hierarchy, the free primary-source databases by jurisdiction, the citation format, and **the citation-verification protocol**. This is the core reference; read it before Step 3. 2. **`references/brief-template.md`** (in this skill) — the exact output structure, including the verbatim disclaimer block. Read it before Step 5. 3. **`references/output-paths.md`** (repo root `references/`) § Tree D — confirms the output location and Drive-mirror behavior. 4. **`references/org/identity.md`** (repo root `references/org/`) — Conversint's home jurisdiction (Texas) and entity facts (Texas LLC), so jurisdiction defaults to **Texas + federal** unless the question says otherwise. If a session asserts an org fact that contradicts the bundle, pause and drift-check per AGENTS.md before proceeding.
This is an internal admin skill, so it does **not** load the full org-strategy bundle or the client-deliverable brand styling — `identity.md` for jurisdiction is the only bundle file it needs.
Brief mode workflow
Step 1 — Frame the question and scope the jurisdiction
Restate the legal question in your own words and confirm scope before researching. If the ask is ambiguous in a way that changes the research — *which* jurisdiction, *whose* obligations, contract vs. statute vs. tort — ask. Use `AskUserQuestion` for genuine forks; don't interrogate when the default is obvious.
- **Jurisdiction.** Default to **Texas + federal** (Conversint is a Texas LLC). If the facts reach another state's law (e.g., a client or its customers are elsewhere), name it and say why. A `--jurisdiction` flag overrides the default. - **Whose obligations.** Conversint's? A client's? Both? This changes which code applies. - **Pick the topic-slug.** Short kebab-case, reused across related admin docs (e.g., `website-services-contracting`, `independent-contractor-classification`). A `--slug` flag pins it. This becomes the Tree D folder name.
Then print one line: `Researching <question> · jurisdiction: <…> · slug: <topic-slug>`.
Step 2 — Decompose into discrete legal issues
Break the question into the separate legal issues a lawyer would spot. "Can we run websites for clients on a handshake?" is really: contract formation and enforceability; IP/copyright ownership of deliverables; liability for defects; accessibility (ADA) exposure; data-privacy obligations; insurance. Each becomes an issue subsection. Getting the decomposition right is most of the value — a missed issue is a missed question for the lawyer.
Step 3 — Research each issue against PRIMARY sources
For each issue, find the **actual controlling law** and read it. Work from `references/legal-source-hierarchy.md`:
- Go to the **free official source** (uscode.house.gov / eCFR / Cornell LII for federal; statutes.capitol.texas.gov / Texas Admin Code for Texas; CourtListener / Google Scholar / Justia for cases). Read the actual section or opinion — don't stop at a summary of it. - **Capture the precise citation and the URL of the exact text you read, as you go.** Section level for statutes/regs, opinion level for cases. Reconstructing citations at the end is how the wrong URL lands next to the right claim. - Rank every source (primary-binding / primary-persuasive / secondary) and note federal-vs- state and binding-vs-persuasive as you collect. - Secondary sources are fine **as a map** to the primary source and for plain-English framing — but chase every one down to the code section it summarizes and cite that.
For a multi-issue question this is parallelizable: dispatch a Sonnet subagent per issue (via the `Agent` tool) to gather and verify primary sources for that issue — each briefed to fetch the official text, capture exact citations + official URLs, apply the verification protocol (`references/legal-source-hierarchy.md` §5), and report what it could *not* confirm. Synthesis stays with you on Opus. The subagents do the per-cite verification at gather time; Step 4 is then your orchestrator-level gate — spot-check the most consequential citations and confirm nothing unverified slipped into the synthesis. Don't blindly re-fetch every cite a subagent already confirmed; do re-check the ones the brief leans on hardest. Brief each subagent with the issue, the jurisdiction, and a pointer to `references/legal-source-hierarchy.md`.
Step 4 — Verify every citation (the safety gate)
**Never write a statute, regulation, or case citation you have not confirmed against the actual source text in this session.** Fabricated case citations are a documented, sanctionable LLM failure — and a fake `123 F.3d 456` repeated to a real lawyer destroys the entire value of this tool. That single rule is why this skill is safe to ship.
Run the full citation-verification protocol — `references/legal-source-hierarchy.md` §5 is the authoritative steps (fetch the real text → confirm it stands for the proposition, not just that the case exists → cut anything you can't confirm and route the gap to the attorney → carry the verified URL to Sources → re-verify URL↔claim pairings at finalize). Read §5 and follow it; don't work from memory of it. Honest gaps beat confident fakes, every time. If a whole sub-area can't be grounded in verifiable primary law, say so plainly rather than papering over it.
Step 5 — Write the brief
Follow `references/brief-template.md` exactly: verbatim disclaimer block at the very top → **Executive summary** (most important findings, the decisions they force, key takeaways) → the question(s) → issue-by-issue analysis with inline primary citations, **each issue closing with a plain-language "So what — the business decision"** → optional cross-cutting "how this maps to Conversint" → **Questions for your attorney** (the payload) → "What I could not verify" → Sources (primary first, then secondary).
The brief has two readers in one: the part that has to be precise (the law, with verified citations) and the part that has to be *actionable* (what to decide). The Executive summary and the per-section "So what" lines carry the second job — they translate the law into the business call Patrick faces, in plain English with no citations. A section that explains a statute but never says what to do about it makes the reader redo the translation the brief exists to save. So write every issue's "So what" as the decision it forces ("decide now," "draft a clause," or "ask the lawyer before acting" — or honestly "nothing to decide, just be aware"), and lead the Executive summary with the highest-stakes one.
Write the prose in plain English a thoughtful non-lawyer reads without a glossary; keep the citations precise. Then **re-verify every inline URL↔claim pairing** — walk each cite and confirm the linked source backs the specific claim next to it. Citations drift during editing; this catches it (it's also the repo's standing source-citation finalization rule).
Make the **Questions for your attorney** section the tightest part of the document. Each question names the issue and the code section and asks something a generalist couldn't answer from the brief alone. That section is what Patrick actually carries into the meeting.
Step 6 — Save, commit to main, return the link
Write the brief to Tree D (see below), then commit it to `main` so prior versions live in `git log` regardless of the active branch:
```bash scripts/commit-md-to-main.sh \ "<path to the brief .md>" \ "Add legal research brief: <topic-slug>" ```
Surface the `GitHub URL:` line the helper prints as the shareable link. **Do not auto-invoke `/gdrive`** — tell Patrick he can run it himself to mirror the brief into Drive's `Legal & Admin › <Topic>` folder.
Output location & filename
Tree D — Conversint admin / operational docs:
``` conversint-admin/<topic-slug>/<Topic> - Legal Research Brief.md ```
- `<topic-slug>` is the kebab-case folder (e.g., `website-services-contracting/`); related admin docs co-locate there. A topic folder may already hold other admin artifacts — write alongside them, don't overwrite. - `<Topic>` in the filename is the title-cased subject (e.g., `Website Services Contracting - Legal Research Brief.md`). - No date/version suffix in the filename — overwrite in place on a re-run; `git log` is the history. (Matches Tree D / `references/output-paths.md`.)
Brief-mode citation discipline (inline)
Every non-trivial legal claim in the saved brief carries an inline hyperlink to a verified primary source, at every restatement (bottom line, issue analysis, questions section) — not just first mention. Use standard legal citation format (see `references/legal-source-hierarchy.md` §4) so a lawyer recognizes it instantly: `42 U.S.C. § 12182(a)`, `28 C.F.R. § 36.302`, `Tex. Bus. & Com. Code § 541.107`, `Robles v. Domino's Pizza LLC, 913 F.3d 898 (9th Cir. 2019)`. Reproduce all sources at the end, primary first.
Guardrails
- **No false legal certainty.** Give direct risk-management recommendations in decision mode, but do not state a contested or fact-dependent legal conclusion as settled. Brief mode explains the law, open questions, and attorney asks rather than replacing counsel. - **Primary over secondary, always.** An issue grounded only in blogs is unfinished. - **Verify or cut.** No unverified citation reaches the brief. - **Jurisdiction discipline.** Don't import another state's law into a Texas question unless the facts put it in play; when they do, say so. - **Currency.** Note effective dates and flag possible amendments; laws change. - **Internal only.** This is a Tree D admin doc, not a client deliverable — no Conversint client-facing brand styling, no auto-Drive.