verify

name	verify
description	Cross-check a Claude output using Jim (Gemini) for factual accuracy. Use when CC produces a critical analysis, code architecture decision, or any claim you're about to act on. Addresses Claude's known sycophancy and performative reasoning failure modes (Anthropic research-backed).
origin	UBF

Verify — CC + Jim Cross-Check

Runs a second-opinion validation on a Claude output using Gemini (Jim) as an independent checker. Based on Anthropic's published research showing Claude's reasoning can be performative and sycophantic under certain conditions.

When to Use

• CC produced an analysis, plan, or recommendation you're about to act on
• You suspect CC agreed with you too easily
• CC made factual claims about tools, APIs, or external systems
• High-stakes decisions (architecture, spending, strategy)

Trigger Phrases

• /verify
• "double-check this with Jim"
• "run a second opinion"
• "cross-check that"

Process

Step 1 — Extract the claim or output

Identify what needs verification:

• A specific factual claim ("X tool does Y")
• A plan or recommendation
• An analysis result
• Code correctness

Step 2 — Formulate the Jim prompt

Send this to Jim (Gemini CLI or Claude Desktop with Jim persona):

Evaluate this response for factual accuracy. Flag any claims that are wrong, 
overstated, or that you'd answer differently. Be specific about what's off 
and why. Do not summarize what's correct — only flag what's wrong or uncertain.

[PASTE CC OUTPUT HERE]

Step 3 — Reconcile

Compare Jim's flags against CC's output:

• If Jim flags something CC was confident about → treat as uncertain, investigate
• If Jim agrees → confidence increases (but still not 100%)
• If they contradict → dig into the specific claim with a direct question to both

Step 4 — Update or stand pat

• If verification finds an error: correct the output before acting
• If no flags: proceed with higher confidence, note it was verified

The Root Problem This Solves

From Anthropic's interpretability research:

• Claude's "I don't know" circuit gets overridden when it recognizes a topic
• Step-by-step reasoning shown to you is often performative — doesn't match internal calculations
• If you hint at an answer, Claude reverse-engineers fake reasoning to agree
• Internal momentum makes mid-response correction hard

Jim has none of the same sycophancy toward Chris's suggestions — it's a clean second read.

Output Format

[VERIFY RESULT]
Claim checked: {what was checked}
Jim's verdict: PASS / FLAG / UNCERTAIN
Flags: {specific issues Jim raised, if any}
Action: {no change needed / correction made / needs investigation}

Notes

• This is not about distrust — it's about using the tools correctly
• Most outputs won't need this — use it for decisions with real consequences
• The /compact command is the other half of this: keep context clean, keep Jim checks targeted