Quote Evidence Guard

The Problem

AI-generated drafts often look sourced before they are actually supported.

A paragraph may include a URL that does not prove the claim. A quoted line may be a paraphrase. A market number may be stale. A confident recommendation may depend on a source with no date, no author, or no authority for the topic. Once that draft becomes a report, marketing page, newsletter, policy note, or client brief, the reader cannot easily tell which claims are solid and which are decoration.

Quote Evidence Guard gives the agent a practical evidence ledger before the text is published or relied on.

How It Works

The Skill forces the agent to create a QUOTE_EVIDENCE_REPORT.

It separates material factual claims from style text, then maps each important claim to a source URL or file path, source date, authority label, and quote boundary. It checks whether a line is an exact quote, paraphrase, summary, inference, or unsupported claim.

It also looks for stale evidence, date-unknown sources, citation-to-claim mismatches, unsupported numbers, vague expert attribution, and source conflicts.

Before / After

Before:

An agent writes a polished research summary with links, but the links only mention the general topic. One quote is not actually in the source, the statistics are from an old report, and a recommendation is stated more strongly than the evidence supports.

After:

The agent creates a claim ledger, marks the statistics as stale, downgrades the quote to an unverified paraphrase, labels the recommendation as weak, identifies the unsupported claims, and rewrites the draft so every material factual statement matches the available evidence.

Hard Rules

1. A citation is not support unless the cited text actually supports the claim.
2. Exact quotes require visible source wording and matching text.
3. Every material claim gets a source, date, authority label, quote/paraphrase boundary, and verdict.
4. Current facts, prices, laws, product specs, rankings, and market claims require current evidence.
5. Conflicting sources must be reported as conflicts, not averaged into confidence.
6. Unsupported high-stakes claims are blocked until primary/current evidence is supplied.

Compatible With

Claude Code, Codex, Antigravity, Cursor, GitHub Copilot, OpenCode, Windsurf, Kiro, and any agent environment that supports SKILL.md style skills.

Best For

• Research briefs and analyst notes
• Marketing copy with factual claims
• Newsletter and article drafts
• Citation audits before publishing
• Quote verification and paraphrase cleanup
• Reports using pasted sources, web links, PDFs, or local reference files

Important Disclaimer

This Skill improves evidence discipline and citation hygiene. It does not guarantee legal, medical, financial, scientific, or regulatory correctness. High-stakes claims still require current primary sources and qualified human review.