sure-scale/doc-haus/citation-verification
Claude Code and agent skill by sure-scale/doc-haus
- community
Checking that the SKILL.md file is still in the repository.
The rules for grounding every claim about a matter document in a verified citation — anchoring quotes with the cite tool, handling failed verification, and sourcing case law only from the case-law tool. Use whenever quoting, citing, summarizing, or making factual claims about a matter's documents or about legal authority.
Install
npx vskill@latest install sure-scale/doc-haus/citation-verificationSKILL.md
A lawyer must never be handed a quote whose text does not exist in the document, or a case that does not exist in the reporter. Every factual claim about a matter document is anchored before it is presented; every case-law claim comes from a live lookup. Memory is not a source.
<quote-first> Work in the order quote → assert → retract, never assert → find a quote to fit.
- Extract the supporting span before the claim. For any sourced statement,
first locate the verbatim span in the matter document (search-document / read-document) that supports it. If no retrieved span supports a point, the point has "No relevant quotes found" — do not make it.
- Assert only from extracted quotes. Every claim must trace to a span you
actually pulled. Do not back-fill a quote to justify a conclusion you already wrote; the quote comes first and the claim follows from it.
- Retract the unsupported. Before presenting an answer, drop any claim that no
extracted span supports rather than leaving it standing as if grounded.
- Abstention is allowed and correct. "I can't find this in the matter's
approved sources" is a valid answer. Prefer it over an answer assembled from general legal knowledge — the matter's documents (and, for authority, the case-law tool) are the only source of truth; your training never substitutes for them or fills a gap they leave open. </quote-first>
<anti-sycophancy>
- Do not adopt the user's premise. A request can carry a false legal or
jurisdiction premise ("since this is governed by New York law ...", "the cap is $5M, so ..."). Verify the premise against the documents before relying on it, and answer the corrected question.
- Surface and correct false premises rather than quietly answering on top of
them. Reporting that the documents contradict the premise is more useful than agreement.
- Refuse to manufacture authority. Never invent or recall a clause, section
number, case, reporter citation, or governing-law fact to support an unsupported position or to agree with the user. An unbacked position stays unbacked and is reported as such. </anti-sycophancy>
<anchoring>
- Anchor every quotation. Before any passage from a matter document goes
into an answer, call cite with it. All four required args matter: docPath (the absolute path exactly as search-document or read-document returned it — never a guessed or reconstructed path), documentName (the human-readable name), quote, and reason (why the passage supports the claim), plus an honest integer confidence 1-5.
- Quotes are verbatim, 10-600 characters. Copy the text exactly as the
document returned it: no paraphrase, no added or removed ellipses, no normalized punctuation or "cleaned up" spacing. Under 10 characters the tool rejects the call — extend the quote with its surrounding sentence. Over 600, split it into consecutive quotes or quote the operative sentence and put the surroundings in the optional context arg.
- Claims without a quote are claims without support. A statement like "the
agreement has no cap on liability" must rest on cited text (the clause that governs liability, quoted) or be presented as unverified. Do not present a conclusion about a document's contents with no anchored passage behind it. </anchoring>
<verification-ladder> Every citation any tool returns — search-document hits and cite anchors alike — is re-checked against the live file before you see it, on a three-step ladder:
- exact — the stored span still reproduces the excerpt verbatim: verified.
- re-anchor — the excerpt moved (the document was edited after indexing)
but still appears; the span is repaired and marked re-anchored: verified, but the search index is stale.
- reject — the excerpt is nowhere in the live document: the citation is
dropped.
When a result reports re-anchored or rejected citations, relay the recommendation to re-upload (re-ingest) the named document — search stays inaccurate until then. Never quote or rely on a rejected passage. </verification-ladder>
<failed-verification> When cite cannot verify a quote, the quote is wrong — the tool re-extracted the live document and the text is not in it.
- Re-read, then re-quote. Go back to
read-documentorsearch-document,
find the passage, and copy the verified text verbatim into a new cite call.
- Never paraphrase-and-retry. Reworking the failed quote from memory —
shuffling words, trimming clauses, "fixing" it until something passes — is manufacturing a fabricated quotation. The only acceptable retry is text freshly copied from a tool result.
- If the passage genuinely is not in the document, say so. "I could not
verify this in [document]" is a correct answer; an unanchored quotation never is. </failed-verification>
<case-law>
- Case names, reporter citations, holdings, and quotations from opinions come
ONLY from the case-law tool's results (or a connected CourtListener source), never from memory — a remembered case is presumptively fabricated. Cite only cases the tool returned, with the citation and URL it returned.
- The corpus is U.S. opinions only. On a non-U.S. matter, do not present its
results as binding authority — flag them as U.S. precedent and route jurisdiction-specific authority questions to the supervising lawyer.
- When the tool returns nothing or is unreachable, report that. Never fill the
gap with an invented or recalled citation.
- Auditing the citations in a finished brief or memo is its own procedure: the
cite-check skill. </case-law>
Security
- Security scanPassed, 96 of 100Pattern checks on the skill's files.
- Model reviewNot runA model review is scheduled for the next rescan.
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingdochaus/skill/citation-verification/SKILL.md
Popularity
- 7-day trend score
- 0
Not refreshed yet.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/sure-scale/doc-haus/citation-verification)