Anti-Fabrication

When invoked, enforce source-traceability on every metric and factual claim in the work. No unsourced numbers. No invented statistics. No "probably around X" placeholders.

Approach

1. Before generating any metric, confirm the source:
   • Database query (show the query)
   • MCP tool call (show the call + result)
   • File on disk (show the file path + relevant excerpt)
   • User-provided data (confirm with user first)
2. If the source isn't available, use a clearly-marked placeholder like [TBD: pull from Google Ads] instead of a number.
3. For every claim in the deliverable, tag it with [source] or [unsourced] inline so the reviewer can verify before shipping.
4. At the end, list every metric and its source. Flag any [unsourced] items as blockers.

Output format

For reports/deliverables:

Revenue: $42,380 [source: shopify MCP - order_total aggregated over last 7 days]
ROAS: 3.2x [source: calculated from Google Ads revenue / spend, both from pipeboard MCP]
Conversion rate: [TBD: pull from GA4]

...

Source trace (end of report):
1. Revenue → shopify MCP → order_total query → last 7 days
2. ROAS → calculated → Google Ads pipeboard MCP data
3. Conversion rate → NOT YET SOURCED - do not publish until fixed

Rules

• Refuse to fabricate. If a source isn't available, say so and use a placeholder. Never invent plausible-sounding numbers.
• "Industry benchmark" claims need citations. If you can't cite the source, don't use the claim.
• Estimates are OK if marked as estimates. "~30% of sessions" is acceptable if clearly marked. A specific "31.4%" without source is not.
• Calculated metrics must show the calculation. "ROAS = 3.2" is a claim; "ROAS = revenue $42,380 / spend $13,243 = 3.2" is a trace.
• Run at the end of every deliverable. The skill catches hallucinations before the user reviews, not after.
• If you catch yourself about to fabricate a number, stop. Mark it TBD. Tell the user. Move on to the next section.