Perform independent fact-checking on the target document using multiple models within Claude, and cross-check the results.
Procedure
Step 1: Identify the target
- Identify the target document from the conversation context
- If the target is ambiguous, confirm with Daiki
Step 2: Check by Opus (itself)
- Read through the target document and fact-check it from the following perspectives:
- Accuracy of technical facts (versions, APIs, configuration values, limit values, etc.)
- Consistency of numerical values (calculation results, totals, etc.)
- Internal contradictions (inconsistencies between sections)
- Accuracy of code examples
- Verify primary sources with WebSearch and clearly state the evidence
- Classify the results as "Correct / Needs correction / Improvement recommended"
Step 3: Independent check by a Sonnet subagent
- With the Task tool, Sonnet (
model: "sonnet") Launch a subagent - Pass only the path of the target document to the subagent, and do not share Opus's results (to ensure independence)
- Example prompt:
以下のドキュメントをファクトチェックしてください。
技術的事実の正確性、数値の整合性、内部矛盾、コード例の正確性を検証し、
「正確 / 要修正 / 改善推奨」に分類して報告してください。
WebSearch で一次ソースを確認し、根拠を明記すること。
対象: <ファイルパス>
Step 4: Comparison and discussion of results
- Cross-check the results from both
- Matching findings → Adopt as is
- Findings from only one side → Opus performs additional verification and makes a judgment
- Findings where opinions are divided → Perform additional verification with WebSearch and reach a conclusion based on evidence
Step 5: Output of the consensus report
- Report to Daiki in the following format:
## ファクトチェック結果
**対象**: <ドキュメント名>
**チェック者**: Claude Opus 4.6 + Claude Sonnet 4.5
### 要修正
| # | セクション | 内容 | 根拠 |
|---|-----------|------|------|
### 改善推奨
| # | セクション | 内容 | 根拠 |
|---|-----------|------|------|
### 検証済み(正確)
<主要な検証済み項目のリスト>