Us Legal Research

US Legal Research

미국 법령 연구를 위한 스킬이야. claude-for-legal 레포의 방법론을 Hermes Agent 포맷으로 이식했어.

Available Data Sources

| Source | What | Auth | Cost | How to use | |--------|------|------|------|------------| | Federal Register API | US federal regulations, proposed rules, notices | None — open API, no key required | Free | curl with URL encoding (see Workflow 1) | | Web search (subagent) | Case law, statutes, secondary sources | None* | Free | delegate_task(toolsets=["web"]) — spawns a subagent with web_search | | r.jina.ai reader | Full-page markdown from any URL | None | Free (rate-limited) | Prefix URL with https://r.jina.ai/ — preferred for article/statute text | | Browser tools (Chromium) | JS-rendered pages, login-gated content, complex layouts | None | Free | browser_navigate + browser_snapshot — fallback for hard-to-reach pages | | Manual URLs | GovInfo, Congress.gov, court websites | Varies | Free | Direct browser navigation

> \* web_search is free in most Hermes configurations. However, web_extract (for reading full page content) requires a configured backend (firecrawl, tavily, exa) — see Optional API Integrations below for the r.jina.ai zero-cost workaround.

---

Optional API Integrations

The skill works out of the box with zero API keys. For enhanced research capabilities, these optional integrations can be configured:

r.jina.ai — Free content extraction fallback

Hermes Agent's built-in web_extract tool requires a configured backend (firecrawl, tavily, or exa) with potential API costs. As a free zero-config alternative, prefix any URL with https://r.jina.ai/:

# Instead of web_extract which may not work:
# Use r.jina.ai directly:
curl -sL "https://r.jina.ai/https://www.courtlistener.com/opinion/12345"

This returns clean markdown with no API key needed. Use this when web_extract returns empty results or errors.

> For Hermes Agent users: To make web_extract work natively, set extract_backend in your ~/.hermes/config.yaml: > ``yaml > web: > extract_backend: firecrawl # or tavily, exa, parallel > ` > Then set the corresponding API key in ~/.hermes/.env`.

CourtListener API — Verified case law citations

Benefit: CourtListener provides free, structured access to federal and state court opinions with verified citations. When connected, the skill can tag citations as [CourtListener — verified] instead of [web search — verify].

Setup:

1. Register for a free API token at https://www.courtlistener.com/api/register/
2. Set the token in your Hermes environment:

# Add to ~/.hermes/.env
COURTLISTENER_API_KEY="your_token_here"

1. Then query via curl with source tagging:

# Search for AI copyright cases
curl -s "https://www.courtlistener.com/api/rest/v4/opinions/?q=AI+copyright&court__id=cafc" \
  -H "Authorization: Token $COURTLISTENER_API_KEY"

Limitations: Free tier has rate limits. Paid tiers remove limits. Some state courts may not be fully indexed.

Hermes Config: Where API Keys Live

| File | Purpose | Example | |------|---------|---------| | ~/.hermes/.env | API keys and secrets | COURTLISTENER_API_KEY=..." | | ~/.hermes/config.yaml | Tool and service configuration | web.extract_backend: firecrawl | | ~/.hermes/skills/research/us-legal-research/ | This skill (SKILL.md + references) | Install via README instructions |

Environment variables from .env are automatically loaded by Hermes Agent and available in all sessions. No restart needed — run /reload in session or start a new session.

---

Core Research Principles

1. Source attribution on every citation — never make a claim without tagging where it came from
2. Separate observations from inferences — mark what the source says vs. what you're concluding
3. No silent supplement — thin evidence = mark as needs-verification, never extrapolate
4. Citation verification before reliance — AI-generated citations can be fabricated
5. State uncertainty explicitly — every output is a draft, not a legal conclusion

Source Tags

| Tag | Meaning | |-----|---------| | [Federal Register] | Retrieved from official FR API | | [web search — verify] | Found via web search — higher fabrication risk | | [model knowledge — verify] | From my training data — verify before relying | | [user provided] | You gave me this info | | [secondary source] | Commentary/aggregator, not primary source |

---

Workflow 1: Federal Regulation Research

Search the Federal Register for federal agency rules, proposed rules, and notices.

Search by keyword

# Note: URL-encode brackets as %5B and %5D in shell, or use --data-urlencode
curl -s "https://www.federalregister.gov/api/v1/documents.json?\
per_page=20&\
order=relevant&\
conditions%5Bterm%5D=${KEYWORD}&\
conditions%5Bagency_ids%5D%5B%5D=${AGENCY_ID}"

Search by agency

Use the agency's numeric ID (not slug). Find IDs via the agencies endpoint:

curl -s "https://www.federalregister.gov/api/v1/agencies?per_page=200" | \
  python3 -c "import sys,json; [print(f'{a[\"id\"]}: {a[\"name\"]}') for a in json.load(sys.stdin).get('results',[]) if 'trade' in a['name'].lower()]"

Common agency IDs:

• FTC: 192
• SEC: 466
• CFPB: 573
• DOL: 271
• HHS: 221
• FCC: 161
• EPA: 145
• DOJ: 268

Search by date range

curl -s "https://www.federalregister.gov/api/v1/documents.json?\
per_page=20&\
order=newest&\
conditions[publication_date][gte]=2026-01-01&\
conditions[publication_date][lte]=2026-05-16"

Output format for a regulation digest

When processing Federal Register API results, the document object's html_url field already contains the full URL:

full_url = doc['html_url']
# Already: "https://www.federalregister.gov/documents/2026/05/18/2026-09943/..."
# Do NOT prefix with base URL again

## Regulation Digest — [date]

**Source:** [Federal Register API]

### 🔴 Always material
**[Agency] — [Title]**
- **Type:** [Final rule / Proposed rule / NPRM / Notice]
- **Summary:** [One-line]
- **Effective / Comment deadline:** [date]
- **Link:** `{doc.html_url}` (already a complete URL)
- **Relevance:** [why this matters]

### 🟡 Review-worthy
[Same format, for items needing assessment]

### 📝 FYI
[N items — titles only]

---

Workflow 2: Source-Verified Citation Pattern

When citing a legal source, follow this format:

For statutes

[Statute] — [Title] § [Section] (Year)
[web search — verify] or [user provided]

For regulations

[Regulation] — [Title] § [Section] ([Year])
[Federal Register] or [web search — verify]

For case law

[Case Name], [Volume] [Reporter] [Page] ([Court] [Year])
[web search — verify]

Reviewer Note (prepend to every research output)

> **⚠️ Reviewer note**
> - **Sources:** [tool used — verified / unverified]
> - **Flagged for your judgment:** [N items marked [verify] | none]
> - **Currency:** [searched for developments since [date] — found N updates | could not search]
> - **Before relying:** [the 1-2 things to check first]

---

Workflow 3: Legal Document Templates

Claim/Element Chart

For mapping legal claims or defenses to evidence:

| [#] | Element | Evidence supporting (pin-cited) | Evidence contradicting | Strength | State |
|-----|---------|-------------------------------|----------------------|----------|-------|
| 1 | [element] | [source, pin cite] | [contra] | strong/moderate/weak/none | supported/partial/disputed/gap |

Always end with:

• Gap list — elements with thin or no evidence (the priority output)
• Conclusion line: This chart is a draft, not a finding. Attorney must verify every cite.

Chronology/Timeline

| Date | Event | Significance | Sources |
|------|-------|-------------|---------|
| YYYY-MM-DD | [what happened, one sentence] | 🔴/🟡/⚪ | [source path] |

• 🔴 Key event — moves the analysis
• 🟡 Relevant context
• ⚪ Background only

Research Brief

## Briefing: [Topic]

**Status:** [current state of research]
**Last updated:** [date]

### Summary
[One-paragraph overview]

### Findings
[Organized by question or theme, with source tags]

### Open Questions
[What remains unresolved]

### Sources
[Numbered list with direct URLs and source tags]

---

Workflow 4: Legal Research Loop (with Goal Integration)

When conducting legal research as part of an active /goal:

1. Set the goal — /goal Research [specific legal question]
2. Gather sources — Federal Register API + web_search
3. Synthesize — extract relevant findings with source tags
4. Validate — check if the goal is achieved (have we answered the question?)
5. Iterate or deliver — if more sources needed, loop; if answered, present findings

After each research cycle, run /goal validate to check progress.

---

Critical: web_search Availability

web_search may not be a direct tool in all sessions (depends on the current provider or config). Do NOT treat this as a blocker. Fallback is reliable:

Fallback: delegate_task with web toolset

# When web_search is not directly available, delegate to a subagent:
delegate_task(
    goal="Research [legal question]: search for case law, statutes, commentary",
    toolsets=["web"]  # grants the subagent web_search access
)

The subagent gets its own web_search tool even when the parent session doesn't have one. This is the canonical workaround — always use this instead of browser navigation or bare terminal for legal research.

When to use which tool

• delegatetask(toolsets=["web"]) — FIRST CHOICE for case law, statute, and regulation research; spawns a subagent with websearch
• Federal Register API (curl via terminal) — for structured federal regulation data (document type, agency, effective date, comment deadline). Lower fabrication risk than web search.
• r.jina.ai reader — PREFERRED for reading specific articles, statutes, or regulatory text pages. Prefix URL with https://r.jina.ai/ to get clean markdown. Free, no API key needed.

> Note: Hermes Agent's built-in fetch_content / web_extract tools require a configured extract_backend (firecrawl, tavily, exa) in ~/.hermes/config.yaml. Without one, they fall back to DuckDuckGo (search only, not extraction). r.jina.ai works as a zero-config drop-in replacement.

• Browser tools (Chromium) — FALLBACK when r.jina.ai fails (JS-heavy SPAs, login-gated content, complex layouts). Uses headless Chrome/Safari/Edge to render and extract the full page.
• browser_navigate(url) → load the page
• browser_snapshot(full=true) → get full text content
• browser_vision() → screenshot + vision analysis for visual-only content

Output conventions

• Every research output gets a reviewer note at the top
• Every citation gets a source tag (what tool produced it)
• Every conclusion is marked [draft — attorney review required]
• Save to outputs/ directory with slug-based naming
• Write .provenance.md sidecar listing sources and verification status

Reference files in this skill

• references/open-source-trade-secret-cases.md — compiled US case law on open source code vs. trade secret (Salerno v. Scialli, SCO v. IBM, Kalda, Bunner, etc.). Research conducted via delegate_task(toolsets=["web"]) fallback.

Key caution patterns

• AI-generated legal citations are sometimes fabricated — always verify against primary source
• CourtListener requires API token (free registration) — note as limitation if not configured
• Federal statutes (US Code) not directly searchable via API without Congress.gov key — use web_search + govinfo.gov as fallback
• State laws vary significantly — always specify jurisdiction in research scope

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: charliehotel
• Source: charliehotel/oh-my-hermes-for-legal-researcher
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.