Judge

GPTClaudeGemini··▲ 922 copies·updated 2026-07-14

judge-4.prompt

# Prompt 3 — Blind Judge (v2)

Run on the **judge** model — a different family from the responder. Run every
pair through **two judge families** if you can; treat any axis where they
disagree as contested rather than averaging.

Two v2 changes: the judge now reports a **margin** per axis (so a decisive win
and a coin-flip are not counted the same), and is instructed to **ignore
surface register**. Note: you should already have normalized the answers
(stripped confidence labels, section headers, meta-framing) before this step —
the instruction is a backstop, not a substitute for normalization.

when to use it

Community prompt sourced from the open-source GitHub repo tonirissanen2nd/thinking-partner-mode (MIT). A "Judge" style prompt — adapt the placeholders and specifics to your task. Imported as-is and not independently retested here, so check the output before relying on it.