Tell Claude to Destroy Your Idea: The Prove-It-or-Destroy-It Pattern

Ask any capable agent 'I think X is true — check it for me' and watch what happens: it goes hunting for evidence that X is true. It finds some, sounds confident, and hands you a tidy answer that flatters your hunch. That feels great and is quietly dangerous, because you asked it to confirm, not to test. The fix isn't a setting — it's a framing you impose in the prompt. You explicitly tell Claude its job is to try to destroy your idea, and to only let through the conclusion that survives a fight. This is a red-team pattern borrowed from security and intelligence work, applied to any high-stakes decision. The walkthrough that inspired this guide (by Samin Yasar) used a stock hunch as the example, so we'll use that too — but the pattern is about rigour, not trading. Read the safety note before you treat any of it as a recommendation.

When you give Claude (the Fable 5 model in the walkthrough) a research goal and tell it not to stop until the job is done, it can spin up several sub-agents that each work in parallel — instead of doing one search and quitting. Each sub-agent owns a different angle and a different source, so they don't all read the same blog and nod along. Then comes the part that makes the answer trustworthy: they challenge each other's findings before a single conclusion is written. The agent that pulled the price history argues with the agent that pulled the analyst notes. What survives that argument is your answer.

• Agent A — the historian: pulls the raw, primary data (e.g. the last two years of price history) and reconstructs what actually happened, not what you remember happening.
• Agent B — the back-tester: runs your hunch against that data point by point. 'Every time it dipped to X it bounced' becomes a testable claim with a hit-rate, not a vibe.
• Agent C — the outside check: looks at what bigger players are doing — public disclosure feeds, analyst and fund commentary — to see if the smart money agrees with you.
• Agent D — the scraper: gathers sources the others can't reach directly (this is where Firecrawl comes in, below).
• The debate: the agents surface their disagreements to each other first. You don't get the cheerful first draft — you get the conclusion left standing after they've tried to knock each other down.

Agent D's job is to scrape sources — disclosure sites, news, anywhere the data lives. The problem: bots get blocked or rate-limited the moment they try. Firecrawl is a scraping/search service built to get through that and return clean, structured text instead of a wall of HTML. You connect it to Claude as a custom MCP connector once, and every sub-agent can use it. MCP (Model Context Protocol) is the open standard, created by Anthropic, that lets an agent call an external tool — a 'connector' is one such tool wired into Claude. Here's the exact, verified setup.

1. Create a free Firecrawl account at firecrawl.dev and open the dashboard. The free tier is enough to try this.
2. Copy your API key from the dashboard. You'll paste it into the connector URL — keep it private, it's a secret.
3. In Claude, open Settings → Connectors (Pro/Max), or + under the message box → Connectors → Manage connectors.
4. Click Add custom connector. If you don't see that option, enable Developer Mode in Settings first — it's hidden otherwise.
5. Paste Firecrawl's remote MCP URL with your key in the path: https://mcp.firecrawl.dev/{YOUR_API_KEY}/v2/mcp — replace {YOUR_API_KEY} with the key you copied. Name it 'Firecrawl' and click Add. (If Claude offers a separate API-key / Authorization field instead, paste the key there and use the base URL https://mcp.firecrawl.dev/v2/mcp.)
6. Set Firecrawl's tools to Always allow so the agents aren't stopped for a permission click on every scrape, then refresh the tools list. You should now see Firecrawl in the connector menu — that's the green light.

This is the load-bearing part of the guide. Set the model to its highest reasoning effort, make sure Firecrawl shows as connected, then paste a prompt shaped exactly like this and swap in your own hypothesis. The structure matters more than the words: (1) state the hunch plainly, (2) order it to prove OR destroy — not confirm, (3) name the specific sources/checks you want, (4) demand the agents argue before answering, (5) ask for a downloadable report with your hypothesis at the top.

• Line 1 — the hunch, stated as a claim to be attacked: 'I have a hunch and I want you to either PROVE it or DESTROY it. My hunch is: [your specific, falsifiable claim].'
• Line 2 — the standing order: 'Run deep research on this and do not stop until the job is done. I do not want your first answer — I want the answer that survives the arguments.'
• Line 3 — the source list (your Agent A/B/C): 'Pull the last two years of real primary data and back-test the hypothesis against every instance. Separately, check [public disclosure source] for what large players actually did. Separately again, gather what independent analysts and funds are saying.'
• Line 4 — the debate, made mandatory: 'Have the agents challenge each other's findings BEFORE you give me a conclusion. Surface where they disagree and resolve it with evidence.'
• Line 5 — the deliverable: 'Then produce a polished, downloadable report: my hypothesis at the top, then the bull case vs the bear case side by side, the back-test result, the data each agent found, and a clear verdict with the reasoning.'

The win condition for this pattern is NOT 'Claude agreed with me.' It's a report you can defend to a sceptic. In the walkthrough, the agent came back and told Samin his hunch was bad — it showed that the dip-buying strategy could have 'won' most trades and still lost money overall versus simply holding. That's the pattern working: it told the truth even though the truth was inconvenient. A good output has these traits.

The pattern is powerful precisely because it can be wrong in fewer places than your gut — but it is not an oracle. Treat these as the guardrails that keep it research rigour instead of false confidence.

• It can only argue over the data it found. If Firecrawl got blocked on a key source, or a feed was stale, the debate is rigorous but under-informed. Check what sources actually got pulled.
• Garbage hypothesis in, confident report out. Make your hunch specific and falsifiable ('bottoms near $X then bounces'), not mushy ('it feels strong'). A vague claim can't be destroyed, only waffled around.
• The debate is only as adversarial as your prompt. Drop the 'destroy it' framing and you're back to a flattering yes-machine. The wording is the mechanism.
• Past data is not the future. A back-test that 'held' for two years is evidence, not a guarantee. The report tells you what was true, not what will be.
• You are the decision-maker. The artifact exists to help you judge faster with more evidence — it does not make the call for you.

We used a trading hunch because the source walkthrough did and it's the clearest example of an idea that wants to be tested. But nothing here is financial, investment, or trading advice, and Claude is not a financial adviser. The value of the prove-it-or-destroy-it pattern is the method — forcing your own idea through an adversarial review before you act — and that method is most useful well away from markets.

• Treat the report as a thinking aid, not a signal. It organises evidence; it does not tell you to buy, sell, or hold anything.
• Markets carry real risk of loss. A surviving hypothesis can still lose money. If you're making actual financial decisions, talk to a licensed professional.
• The pattern shines on non-financial calls too: 'Should we adopt this vendor?', 'Is this the right architecture?', 'Does this hiring plan hold up?' — anywhere you'd otherwise just vibe it out.
• Use it to de-risk decisions, not to manufacture certainty. The goal is fewer confident mistakes, not a crystal ball.

This is Part 4 of a five-part series on putting Claude to work. Each part is a standalone pattern, and they compound. Part 1 set the goal and workflows so the agent runs without babysitting; Part 2 turned it into a hands-on operator (with a hard human-permission guardrail on anything it cancels or sends); Part 3 made it a creative engine; this part makes it a sceptic that tells you the truth. Part 5 is the payoff — turning the whole skill set into work people pay you for.

• Part 1 — Give Claude a goal: guides.kno2gether.com/fable5-payroll-goal — goals + workflows so the agent self-runs.
• Part 2 — The autonomous bookkeeper: guides.kno2gether.com/fable5-payroll-bookkeeper — connect your data and let it work, with a review-before-it-acts guardrail (it must never cancel, pay, or send without your explicit OK).
• Part 3 — The ad maker: guides.kno2gether.com/fable5-payroll-admaker — Claude as a creative engine that generates and self-sorts output.
• Part 4 — Prove it or destroy it (you are here): guides.kno2gether.com/fable5-payroll-redteam — adversarial research that survives the arguments.
• Part 5 — Turn it into income: guides.kno2gether.com/fable5-payroll-clients — package these patterns into a paid service, including outreach where Claude RESEARCHES and DRAFTS but a human reviews and sends every message.