Do you want to check the credibility of an opinion?

I. Core Directives & Guardrails

1. Scope & Triage: This is for statements of opinion or mixed fact and opinion. First, apply fact_vs_opinion_triage skill. If the input is a pure provable fact, redirect to fact-check workflow. Otherwise, proceed.
2. Active Tool Use: You are an active investigator. Do not rely on training weights. Use web search and browser tools to find primary data, comparable numbers, and actor motives.
3. Triage: Before writing anything, break the claim down into (a) factual assertions, (b) cause-and-effect interpretations, (c) unstated assumptions, and (d) timing — pin down when the claim appeared to search for what else was happening then. Causal-link preservation: If a claim contains a compound causal chain, do NOT split it into isolated nouns. Test the causal relationship itself as one unit.
   1. Example: "Capital goods imports are up because we're building factories." Do NOT split into "capital goods imports are up" + "factories are being built" — both are individually true, but the link is false. The capital goods surge was actually computer equipment for data centers, not factory machinery. Factories were built only minimally. Instead, test: "The increase in capital goods imports is specifically driven by factory construction." Only decompose further if first-pass evidence is too ambiguous.
   2. Document this decomposition as Section 0 ("Triage Decomposition") of the Synthesis & Verdict memo (04_synthesis_and_verdict.md). Write this section before beginning Phases 1-3, then return to complete the memo after Phase 3. As a final step, condense the triage section if it became too long during the investigation.
4. Separate Source from Substance: Strictly apply source_bias_disclosure skill. Beyond disclosure, you must actively weigh bias in your verdict (e.g., claims supported only by adversarial sources score lower in the Monte Carlo than those with independent primary data).
5. Citation Formatting (Strict Compliance):
   • Tagging: Prefix claims with **Fact:**, **Opinion:**, or **Unsure:**.
   • Citations as Sub-Bullets: Every factual claim's citation should be nested as a bullet point below it.
   • Format: *Citation:* (Author Last Name, "Title," Publication, Date, <https://full-url-in-plaintext.com>). Quote: "exact verbatim quote."
   • Link Rules: Do NOT hide URLs behind markdown syntax. Write raw URLs wrapped in angle brackets (e.g., <https://example.com>). Use accessed [date] for undated pages. If unverifiable, state so. Wrapping in angle brackets makes them clickable while keeping the raw URL visible.
   • Calculations: Cite inferences as: [CALCULATED] Based on data from (source). Raw values: X, Y, Z. Calculation: X + Y = Z.
6. The Echo-Chamber Guard: NO circular citations. Apply anti_circular_citation_guard skill.
7. Exhaustive Depth Requirement: Do not write thin memos. Provide multiple distinct pieces of evidence, historical context, and comprehensive elaboration. One or two bullet points is a failure. Keep digging until the argument is rigorous and multi-layered, but stop when additional evidence would be redundant.
8. Analytical Integrity: Before rendering judgment, apply analytical_integrity skill (premise de-anchoring, symmetrical search, scope fidelity, evidence engagement, bias symmetry).

II. Phase 0: Claim Identification (Gated Approval)

Goal: Ensure you evaluate the correct opinions.

1. Identify Opinions: If the user provides a large document, extract the most prominent opinions or mixed claims.
2. Stop and Ask: STOP and ask the user: "Here are the opinions I identified. Which should I credibility-check?"
3. Wait for Approval: Do NOT proceed until the user confirms which claim(s) to evaluate.

III. Phase 1: The "PRO" Advocacy Memo (01_claim_is_credible.md)

Goal: Act as a ruthless advocate proving the claim TRUE. Marshal every supporting fact under the most charitable interpretation. Do not fabricate.

MANDATORY MEMO STRUCTURE: (Do not skip headers. Execute targeted searches to fill them).

• ### 1. Authenticity & Strategic Distribution: Defend the source's genuineness. Invert the Absurdity: If it looks messy or amateurish, argue why that was deliberate — e.g., bypassing mainstream channels. If actions look self-destructive, search for the hidden rationale. (If none found, say so rather than inventing one.)
• ### 2. Hidden Catalysts & Actor Motives (Active Search Required): (Action: Run 3-5 targeted searches about the people/domain of THIS claim to find urgent pressures — deadlines, crises, disruptions — at the time it appeared. Some searches must target the actors' broader situation, not just the claim's keywords). What pressure made releasing this data rational right now?
• ### 3. Do the Numbers Actually Check Out?: Defend the math. Look for comparable numbers or hard limits that back the claim. Are there defunct companies, legacy contracts, or dead weight explaining why a number looks odd? If squeezed, search for what's sitting on part of the total.

IV. Phase 2: The "CON" Advocacy Memo (02_claim_is_not_credible.md)

Goal: Demolish the claim using exhaustive scrutiny of its source, logic, and structural reality.

MANDATORY MEMO STRUCTURE: (Do not skip headers. Execute targeted searches to fill them).

• ### 1. Where Did This Actually Come From?: Expose echo chambers, astroturfing, or amplification hiding the true origin. Trace identical phrasing to its source. (Don't confuse messenger with message — legitimate info is often first published by unreliable outlets.)
• ### 2. Who Benefits From This Being Believed?: Explain how Phase 1's urgent pressures could equally motivate fabrication or astroturfing. (Motive explains WHY the claim exists, not whether it's true.)
• ### 3. Do the Numbers Actually Work? (Active Search Required): (Action: Run 2-3 searches for comparable numbers, hard limits, or legacy data for THIS claim's figures. Include sideways searches beyond the claim's own terms). Attack the methodology. Find factors making the numbers impossible or misleading. Check if the anomaly is a mundane mistake before inventing conspiracies.

V. Phase 3: The "Public Misconception" Memo (03_other.md)

Goal: Identify and debunk arguments the public thinks are relevant but are actually irrelevant, debunked, or categorically different.

MANDATORY MEMO STRUCTURE:

• ### 1. The Distraction: Identify the primary myth or irrelevant tangent the public associates with this claim.
• ### 2. Why It Is Irrelevant: Explain mechanically why this fear does not apply to the claim being evaluated.

VI. Phase 4: Synthesis & Verdict (04_synthesis_and_verdict.md)

Goal: Act as an impartial judge weighing the preceding memos to find the truth across different semantic interpretations of the claim.

0. Triage Decomposition: (Already written before Phases 1-3 per Directive 3. Review and condense if needed.)
1. The Claim & Interpretations: At the very top, state the claim whose credibility is being evaluated. If the claim contains subjective or ambiguous terms (e.g., "popular," "successful," "dangerous"), you MUST explicitly define 2-3 distinct semantic interpretations of what the claim could mean in context (e.g., "Interpretation A: Sweeping National Majority" vs. "Interpretation B: Structural Critical Mass").
2. Memo Summaries: Provide a comprehensive bulleted summary. You MUST extract and include every single idea, argument, risk, and debunked misconception from the PRO, CON, and Public Misconception memos. Do not leave anything out.
3. The Monte Carlo Histograms (Per Interpretation):
   • For EACH interpretation defined in Step 1, simulate 1,000 highly intelligent evaluators reading the memos.
   • Generate a separate markdown table histogram for EACH interpretation (using ████████ blocks) with these five exact buckets: Definitely True, Mostly True, Unknown / Evidence Too Weak, Mostly False, Definitely False. The total for each histogram MUST sum to exactly 1,000.
   • State the specific Interpretation directly above its respective histogram.
   • Cohort Narratives: Under each histogram, use bullet points for each bucket. Explain in plain language exactly which specific piece of evidence, deadline, or conflicting fact swayed that cohort for that specific interpretation. What real-world messiness did they weigh against the clean version of events?
4. Executive Verdict: Provide a 2-3 sentence plain-English summary synthesizing the results across all interpretations. Name the single most important fact that swung the majority. If the claim is technically false under one interpretation but directionally right under another, evaluate both angles. If it's probably true but can't be properly proven with available data, declare it "PROBABLY TRUE but CAN'T BE PROVEN YET." List specific upcoming events (named trials, rulings, earnings reports) that would change the answer.

VII. Phase 5: External Adversarial Audit

Goal: Stress-test the analysis using external tools and force the analyst to defend against findings.

1. Submit to External Tools: Submit the claim and your interpretation to Deep Research and Deep Think (in parallel, per the Gemini workflows). Do NOT bake your own conclusions into the prompt — let the external tool evaluate independently.
2. Extract and Save Results: Save as deep_research_result.md and deep_think_result.md in the session folder.
3. Run Adversarial Audit: For each external source, execute credibility-check-audit workflow. Complete ALL audits before proceeding to v2.
4. Create v2 Memos: Based on audit recommendations, create v2 copies of updated memos (01_claim_is_credible_v2.md, 02_claim_is_not_credible_v2.md, 04_synthesis_and_verdict_v2.md). Do NOT overwrite originals.

VIII. Pre-Delivery Audit & Hand-Off

1. Silent Audit: Review your files silently. Correct:
   • Factual claims lacking a direct-quote citation with a plaintext URL.
   • Generic text in "Active Search Required" sections (did you actually search or just guess?).
   • A histogram that fails to sum to 1,000.
   • Unflagged circular echo-chamber citations.
2. Next Steps: Present the verdict. Offer verify-links-in-document workflow.
3. Follow-Up Protocol: Answer questions conversationally. Do NOT edit memos unless explicitly instructed.

• README.md — Triage breakdown of the claim into factual assertions, interpretations, assumptions, and timing.

  • Example: 001_credibility_check_claim_name\​README.md

• 01_claim_is_credible.md — The PRO advocacy memo: the strongest possible case that the claim is true.

  • Example: 001_credibility_check_claim_name\​01_claim_is_credible.md

• 02_claim_is_not_credible.md — The CON advocacy memo: the strongest possible case that the claim is false.

  • Example: 001_credibility_check_claim_name\​02_claim_is_not_credible.md

• 03_other.md — The public misconception memo: debunks irrelevant arguments that the public associates with the claim.

  • Example: 001_credibility_check_claim_name\​03_other.md

• 04_synthesis_and_verdict.md — The final verdict with Monte Carlo histograms (1,000 simulated evaluators) across multiple interpretations of the claim.

  • Example: 001_credibility_check_claim_name\​04_synthesis_and_verdict.md