FIELD NOTE 05 / TWO MODELS. ONE BETTER DECISION.
Get a second opinion before it matters.
Find the disagreement your first answer did not show you.
An answer can sound finished before it has been challenged. Give the same problem and evidence to two models independently, then compare the claims that would change your decision.
“This is the one I would keep if I could only keep one.”
Paul’s working notes · September 2026
What it looks like
“Are you sure?” in the same conversation, after sharing your preferred answer.
Same question + same evidence ↙ ↘ Claude, fresh chat ChatGPT, fresh chat ↘ ↙ Compare claims. Check disagreements. Decide.
Put it to work
- 01
Prepare two prompts from current guidance.
Ask one model to read the latest official Anthropic and OpenAI prompting documentation, including guidance for the exact models you intend to use. Have it write two tailored prompts and one synthesis prompt. Keep the question, source material, and success criteria identical.
- 02
Run the research independently.
Use fresh conversations. Give each model the same starting evidence, without the other model’s answer. Ask for sources, uncertainty, and evidence that would change the recommendation. Different models can still make the same mistake; this is a useful challenge, not independent proof.
- 03
Synthesize, then check the decision.
Paste both labeled answers into a third conversation. Ask for a comparison of claims, missing evidence, and disagreements. Verify the important sources yourself or with a tool. For code, keep the reviewer read-only, ask for concrete failure scenarios, let the author adjudicate each finding, and rerun the affected checks.
TRY THE WORKFLOW
One question.
Three useful prompts.
Use the setup prompt to have your model research the latest guidance and write the full set. The other tabs are adaptable starting templates, not live research results.
Your topic stays in this page. Nothing is sent or saved.
Help me prepare two independent research prompts and one synthesis prompt. Question: [YOUR QUESTION OR DECISION] Context and evidence: [paste the same relevant facts and source material into both research prompts] Constraints: [budget, timing, audience, and what cannot change] Success means: [the decision this research needs to support] Before writing them, read the current official prompting guidance from Anthropic and OpenAI, including the model-specific guidance for the exact models I plan to use. Ask me which models if I have not specified them. Use these starting points and follow current official links: Anthropic: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices OpenAI: https://developers.openai.com/api/docs/guides/prompt-engineering If browsing is unavailable, say so and ask for the documentation. Do not claim to have checked guidance you could not read. Cite what you read, give the access date, and briefly explain the adaptations. Produce three clearly labeled, copyable prompts: 1. A Claude research prompt, tailored to the selected Claude model. 2. A ChatGPT research prompt, tailored to the selected OpenAI model. 3. A Claude synthesis prompt to use after I supply both completed answers. Keep the question, evidence, constraints, and success criteria identical across the research prompts. Each researcher must work independently, cite primary sources, separate evidence from inference, report uncertainty, and seek evidence against its own recommendation. Do not disclose one answer to the other researcher. The synthesis prompt must compare claims, expose disagreements and missing evidence, and recommend what to verify. It must not treat agreement as proof. Provide concise reasons and evidence, not hidden chain-of-thought.
Two agreeing answers can share one bad source. Count evidence, not votes. For consequential decisions, use the appropriate qualified reviewer and independent checks; adding a model does not replace them.
Independent answers first. Evidence-based synthesis second.
Sources & further reading
Personal counts are from Paul’s prompt-history analysis. Tool guidance was checked against these sources in September 2026. Confirm current behavior in your installed version.