sure-scale/doc-haus/citation-verification

Claude Code and agent skill by sure-scale/doc-haus

Trust tier 2
  • community
LLM review pending
Source check pending

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-verification

SKILL.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:

  1. exact — the stored span still reproduces the excerpt verbatim: verified.
  2. 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.

  1. 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-document or search-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

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
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/sure-scale/doc-haus/citation-verification/badge)](https://verified-skill.com/skills/sure-scale/doc-haus/citation-verification)

Something wrong with this skill or its page? Tell us.