← Back to all workflows
Verification

Do you want to verify the AI's claims?

This workflow verifies the AI's own prior claims by finding independent evidence for every factual assertion, applying anti-circular-citation guards, and producing an HTML evidence packet with screenshot proof.

Under the Hood

This is the actual text of the workflow

Evidence Packet Workflow (for Any Claims)

Use this workflow when the user asks you to prove, cite, or build an evidence document for data referenced in a text, transcript, article, or previous answer.

[!IMPORTANT] Step 0 — Triage: Before doing anything else, read and apply fact_vs_opinion_triage skill. When scanning the document, extract only verifiable facts — skip pure opinions. If the document quotes someone's opinion (e.g., "Expert X says Y will happen"), verify that the person actually said it, but do NOT try to prove or disprove the opinion itself.

[!IMPORTANT] Analytical Integrity: When searching for and evaluating evidence, read and strictly apply analytical_integrity skill.

I. Steps

  1. Identify all factual claims. Review the target text/answers and list every specific number, percentage, or data point that was cited. For each one, record:

    • The exact claim (e.g., "Native-born unemployment rate rose from 4.3% to 4.7%")
    • The exact number(s) involved
    • The source URL where the number can be found

  2. Verify and Capture screenshots. For each source URL:

    • Open the visible browser using the visible_browser MCP tool and strictly follow the policies in web_browsing_policy skill
    • Navigate to the URL and read the section containing the cited number.
    • Crucial Anti-Hallucination Guard: You must strictly apply the logic from anti_circular_citation_guard skill to ensure the source is genuinely independent and not merely quoting or echoing the original claim.
    • Crucial Verification Step: If the source is a circular echo, or if the source data does NOT match your cited number, or if the source URL is dead/non-existent, immediately flag this claim as a FAILURE. Do not force an unrelated screenshot.
    • If the data is correct and passes the circular-citation guard, capture a screenshot showing the number clearly visible and save it to the working directory with a descriptive filename like evidence_[topic]_[date].png.

  3. Build the HTML evidence file. Create a self-contained HTML file in the working directory. Structure:

    • An executive summary at the top
    • One "evidence card" per valid claim, containing:
      • A numbered badge (EVIDENCE 1, EVIDENCE 2, etc.)
      • The claim text in large, bold font
      • A data table showing the exact numbers side-by-side (with red/green highlighting for bad/good changes)
      • The source URL as a clickable link
      • The screenshot embedded directly below the link via <img> tag
      • A caption under the screenshot describing what to look for in the image
    • For any FAILED/Hallucinated claims, build a bright red RETRACTION CARD instead, containing:
      • A red numbered badge (e.g., RETRACTION 1)
      • The original, hallucinated claim in large, bold font
      • A clear admission that the statistic was hallucinated, unable to be found, or contradicted
      • If contradicted, a data table comparing the fake/wrong claim alongside the actual source data
      • A link to the source (if it existed) demonstrating where the logic misread the information
    • Use a dark theme with clean, modern typography (Inter font)
    • All screenshots must be embedded as local file references (relative paths), NOT base64
    • The HTML file must be self-contained (all CSS inline in a <style> block, no external dependencies except Google Fonts)

  4. Preview and verify. Open the HTML file in the visible browser using the visible_browser MCP tool to confirm:

    • All screenshots load correctly
    • All data tables are accurate
    • All source links are clickable and correct
    • The document looks professional enough to send in an email

  5. Naming convention. Name the HTML file [topic]_evidence_packet.html. Name the screenshots evidence_[source]_[date].png.

II. Important Notes

  • Always capture screenshots on the same day the evidence packet is built so the screenshots match the live data.
  • If the BLS or any source is blocking automated access, use the visible browser using the visible_browser MCP tool.

The Output

When you run this workflow, the AI agents will generate the following folders and files:

  • [topic]_evidence_packet.html — A self-contained HTML file containing an executive summary and individual evidence cards for each claim, complete with data tables, source links, and embedded screenshot proofs.

    • Example: economy_evidence_packet.html

  • evidence_[source]_[date].png — Screenshot proof files showing the exact cited numbers clearly visible on the independent source websites.

    • Example: evidence_bls_gov_2023-10-26.png

How to Set This Up

Option 1: The Easy Way (Automated)

Just point Antigravity to this webpage and ask it to figure it out for you. Antigravity can read this documentation, copy the workflow script, and automatically generate all the required skill files in the correct directories on your machine.

Option 2: The Hard Way (Manual Copy & Paste)

If Antigravity fails to set this up automatically, you will need to manually copy the scripts into your local directories:

  1. Copy the raw workflow script from the "Under the Hood" section above.
  2. Save it as C:\Users\[Your Name]\.gemini\antigravity\global_workflows\find-evidence-for-claims-in-ai-answer.md.
  3. You must also click every hyperlinked skill file and save its contents into your skills directory. You must do this for every single skill file linked in the workflow. For example, the legal_analysis_partner skill must be saved to C:\Users\[Your Name]\.gemini\antigravity\skills\legal_analysis_partner\SKILL.md.
  4. Once all files are saved, open Antigravity and type /find-evidence-for-claims-in-ai-answer in the chat to run it.

New to Antigravity? Read the Master Installation Guide first.