- Home
- Skills
- Agents & Orchestration
- Evidence & Citation Guard: every claim cites a stored source
More screenshots
Works with the AI Agents you already use
Evidence & Citation Guard: every claim cites a stored source
Enforce verifiable evidence by fetching sources, matching quotes word-for-word, and ranking source authority.
$12
Evidence & Citation Guard: every claim cites a stored source
Example session with this skill installed
Check the draft in research_note.md for citation errors and verify all quotes against the registry. Use strict mode.
- Read your context and instructions
- Compiled the evidence & citation
I have performed a strict evidence check on research_note.md. The script verified all 4 citations against the .evidence/ registry. Two errors were found regarding quote accuracy and source ranking.
Evidence Check Results
| Severity | Code | Location | Description |
|---|---|---|---|
| Error | E02 | research_note.md:14 | Quote mismatch in [S02]. Registry has "8,000 msg/s", draft has "8,500 msg/s". |
| Error | E05 | research_note.md:28 | Rank violation. Fact cited from S04 (Rank 4), but fact_max_rank is 3. |
| Warning | W01 | research_note.md:42 | Sentence contains a specific date (2024-05-12) but lacks a citation. |
| Pass | - | research_note.md:08 | [S01 §Intro "v2.1 is stable"] matches registry perfectly. |
Source Registry Status
| ID | Source | Rank | Status |
|---|---|---|---|
| S01 | docs/internal/v2-spec.md | 1 | Stored |
| S02 | https://api.provider.com/limits | 2 | Stored |
| S03 | notes/interview-jdoe.md | 3 | Stored |
| S04 | https://blog.industry-news.io/rumors | 4 | Stored |
Next steps
- Correct the throughput limit in line 14 to match the stored source text.
- Find a higher-rank source for the claim on line 28 or rephrase it as an unverified report.
- Run
python scripts/evidence.py check research_note.md --strictagain to confirm the fix.
Connects securely to your tools. The creator never sees your data.
About this skill
Trusting citations is a gamble when sources are never verified or hallucinated. This skill makes evidence checking mechanical by fetching and storing source text, enforcing word-for-word quote matching, and ranking sources by authority. It prevents weak or missing evidence from leaking into research, specs, or technical recommendations.
A citation proves nothing unless someone checks it. This skill makes checking mechanical. Sources are fetched and stored by script, every claim cites one with a short word-for-word quote, and every quote is matched against the stored copy. Your source policy decides what may back a fact: a blog can’t prove a product limit.
The check catches misquotes, invented source IDs, weak sources, over-long quotes, and pages that changed since you cited them. It also lists sentences with numbers and no citation. Then comes the step no script can do. A claims table puts each claim beside its quote, so you or a second agent can mark whether it really supports it (“up to 8,000” is not “8,000”).
For research summaries, vendor comparisons, specs, due-diligence notes and agent-generated reports. Python 3.10+, nothing to install, safe and polite fetching. Includes a worked example, guides for the citation rules, source policy and support review, and 29 tests.
What it does
- Source Verification fetches and caches the exact text of URLs or local files to prevent link rot or memory-based citations.
- Strict Quote Matching ensures every citation contains a word-for-word quote of 25 words or fewer that exists in the stored source.
- Authority Ranking applies a custom source policy to ensure critical facts are backed by high-rank sources, not hearsay or low-quality leads.
- Conflict Management identifies when sources disagree and forces the agent to document the discrepancy rather than picking a winner silently.
- Audit Table Generation produces a claims table for human or agent review to verify if a quote actually supports the claim it follows.
How it works
- Initialize the evidence registry and define your source policy in
policy.jsonto set authority ranks. - Collect sources using
add-urloradd-fileto build a local cache of the evidence before writing begins. - Write content using the specific citation format
[S01 §Section "Quote"]immediately following each claim. - Validate the draft using the
checkcommand to find misquotes, missing citations, or rank violations. - Audit the logic by generating a claims table to confirm the quotes logically support the assertions.
Frameworks & tools
The core logic is a standalone Python 3.10 script with no external dependencies for text processing. It integrates with markitdown or pdftotext for PDF handling.
Why this beats prompting it yourself
Prompting an agent to "be accurate" or "cite sources" often results in broken links or quotes that don't actually exist in the text. This skill uses a local file-system registry and a Python-driven validation loop that refuses to pass a document unless every single quote is found byte-for-byte in the original source.
Use cases
- Vendor Comparisons where pricing and feature claims must be backed by current documentation.
- Technical Specifications that must cite internal requirement docs and architectural decisions.
- Due Diligence Notes where every fact about a company or project needs a verifiable trail.
- Agent-Generated Reports that require a final "verification pass" to eliminate hallucinations.
Known limitations
Does not support JavaScript-heavy sites or scanned PDFs without external OCR tools. Limited to 25-word quotes to prevent copyright issues and maintain citation density.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
Requires Python 3.10+. Cannot process JavaScript-rendered sites or scanned PDFs natively. Maximum quote length is 25 words. Registry and cache must be managed via the provided script.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 11 days ago
- Software
Python