Fact check

Skill: Fact-Check

Automatically verify numerical claims, data-driven statements, and factual assertions in documents before they are finalized.

When to Activate

This skill MUST be invoked when any of the following conditions are met:

• The user explicitly asks to "fact-check", "verify data", "validate claims", or similar
• A document, report, or presentation containing numerical claims is about to be finalized
• The user asks to "finalize", "publish", or "send" a document that includes data-driven statements
• A report has been generated from source data and is being reviewed

Definitions

Numerical claim: Any statement containing a number, percentage, count, dollar amount, ratio, date-based statistic, or quantitative comparison (e.g., "revenue grew 34%", "we processed 1,200 claims", "average turnaround is 3.2 days").

Factual statement: Any assertion presented as fact that can be checked against source data (e.g., "our largest client is X", "the most common denial reason is Y").

Source data: The original dataset from which claims are derived. This may be a database (via SQL), CSV files, JSON files, Excel/Google Sheets, API responses, or any structured data the user can point to.

Verification Process

Follow these steps in order. Do NOT skip any step.

Step 1: Read the Document

Read the full document or report content. If the document is being generated in-session, capture all text that will be included in the final output.

Step 2: Extract All Claims

Systematically go through the document and extract every:

• Numerical value (counts, amounts, durations, dates)
• Percentage or ratio
• Comparative statement ("increased by", "more than", "largest", "fastest")
• Aggregate or summary statistic (averages, totals, medians)
• Factual assertion that can be validated against data

Create a numbered list of every claim found. Do not skip claims that seem obviously correct -- verify everything.

Step 3: Identify Source Data

Determine where the source data lives. Check for:

• Files referenced in the conversation (CSV, JSON, XLSX, etc.)
• Database connections (Supabase, PostgreSQL, etc.)
• API endpoints that were queried
• Google Sheets or other cloud data sources

If the source data is unclear, ASK the user before proceeding. Do not guess.

Step 4: Write and Run Verification Scripts

For each claim, write and execute a verification script (Python or Node.js) that:

• Queries or reads the source data
• Computes the correct value independently
• Compares the claimed value against the computed value
• Calculates the percentage deviation

Use the Bash tool to run these scripts. Prefer Python with pandas for CSV/data analysis. For database queries, use the appropriate MCP tools (e.g., Supabase execute_sql).

Step 5: Classify Each Claim

Assign one of these statuses to every claim:

Status	Criteria	Action
CONFIRMED	Exact match or deviation <= 5%	No action needed
CLOSE	Deviation between 5% and 15%	Flag for user review; recommend correction
WRONG	Deviation > 15% or factually incorrect	Must be corrected before finalizing
CANNOT VERIFY	No source data available or accessible	Explicitly note in document or remove claim

Step 6: Check Internal Consistency

Beyond individual claim verification, check for:

• Totals: Do sub-items add up to the stated total?
• Percentages: Do percentage breakdowns sum to 100% (or close)?
• Contradictions: Does any claim contradict another claim in the same document?
• Methodology alignment: Does the described methodology match how the numbers were actually computed?
• Time period consistency: Are all claims referring to the same time period when compared?

Flag any inconsistencies found.

Step 7: Present the Fact-Check Report

Present findings in a clear table format:

## Fact-Check Report

Document: [document name/description]
Source data: [source files/databases used]
Claims verified: [count]

| # | Claim | Claimed Value | Verified Value | Deviation | Status |
|---|-------|---------------|----------------|-----------|--------|
| 1 | ...   | ...           | ...            | ...       | ...    |

### Internal Consistency Checks
- [List any inconsistencies found, or "All consistency checks passed"]

### Recommended Corrections
- [Specific corrections for WRONG and CLOSE claims]

### Unverifiable Claims
- [List claims marked CANNOT VERIFY with recommendation to source or remove]

Step 8: Apply Corrections and Re-verify

If any claims are WRONG or CLOSE:

1. Recommend the exact corrected text to the user
2. Wait for user approval before making changes
3. After corrections are applied, re-run verification on changed values to confirm

Step 9: Finalize Only After Verification

Only proceed with finalizing the document (saving, sending, publishing) AFTER:

• All claims are CONFIRMED or user-approved
• All internal consistency checks pass
• The user has acknowledged any CANNOT VERIFY claims

Important Rules

1. NEVER skip fact-checking when generating or finalizing reports that contain data claims. This is a blocking requirement.
2. Always show verification results to the user before finalizing. Never silently pass a document through.
3. Verify round numbers too. A claim of "approximately 1,000" should still be checked -- the actual value might be 800 or 1,300.
4. Check the math, not just the data. If a document says "34% of 500 is 180", verify both the percentage AND the arithmetic.
5. Preserve precision context. If source data shows 33.7%, and the document says "about 34%", that is CONFIRMED. But if source data shows 28.1% and the document says "about 34%", that is WRONG.
6. Document the verification. The fact-check report should be included or referenced so the user has an audit trail.
7. When in doubt, flag it. It is better to flag a correct claim for review than to let an incorrect claim pass.
8. Cross-reference multiple sources when available. If two data sources disagree, flag the discrepancy.

Example Invocations

• /fact-check -- Run fact-check on the current document or most recent report
• "Verify the data in this report before we send it"
• "Fact-check these numbers against the database"
• "Can you validate the claims in this investor update?"
• Automatically triggered when a report with data claims is about to be finalized