When I realized AI can get stuck on specialist problems
Last week in Hangzhou, I was revising a client proposal. I kept working until 10 p.m. and still missed a key assumption. I broke the problem apart and asked AI, which gave me a polished answer with one point that collapsed under follow-up questions.
Around the same time, I saw Anthropic describe Claude helping with black hole physics. It wasn't announcing answers for scholars; it was helping them investigate questions that had remained open for years. That changed my focus: could we borrow this workflow to avoid taking the “looks right” route?
Its value isn't an answer; it's helping you chase the question down
Anthropic's research team used Claude on two kinds of open questions: when black holes “evaporate,” and how information enters phenomena that remain only partly explained. Researcher Pablo Hayden used Claude to generate and screen candidate explanations, then checked each path with his own professional judgment. The work didn't stop at “one-click paper.” AI sat inside a loop of gathering evidence, proposing hypotheses, finding contradictions, and asking the next question.
I got this wrong too. At first, I asked Claude for the final conclusion. The prose looked great; the next day, I found it had blended two premises together. I switched to asking it to list the known conditions, the unknowns, and possible counterarguments before checking each one. Rework dropped after that.
For me, Claude was more like a patient research assistant than the expert making the final call.
Try it today: near-zero cost, but keep verification in the loop
Money: The Claude web or app version usually starts with a free allowance, so there is no need to buy a complicated plan first.
Time: The first hands-on try takes about 20 minutes. To adapt the workflow to your own business problem, I would run a small test for one hour.
Technical barrier: No coding is required, but you do need to judge whether your sources are reliable.
First step: Open Claude, start a conversation, enter “The work problem I'm currently trying to solve,” then paste what you know and what information is still missing into two separate columns.
If you try a specific project such as Anthropic Research, you don't need to copy the black hole terminology. Rewrite your problem into three parts: what facts are established, what explanations are possible, and what evidence would overturn each explanation. Let AI list candidate paths, then return to the original sources and client feedback to verify them.
There's no pressure to try it immediately; this workflow isn't for everyone. If independent research is already your strength, it may simply add a little polish.
At different stages, I give it different jobs
Just starting: While I'm still looking for a product or customer direction, I ask Claude to handle one small question at a time, such as organizing interview notes or listing assumptions that need verification, without rushing it for business advice.
With one or two clients: Once I have real cases, I give Claude both the successful process and the failure retrospective, ask it to look for contradictions, and then decide which problems deserve priority by checking them myself.
Scaling up: Once I'm building a team, I use it to organize research tasks, record evidence sources, and flag points waiting for human judgment. I keep the verification history so a colleague can pick up the work.
The most important thing I took from this wasn't “AI can solve impossibly hard questions.” It was the reminder that good answers often come from persistent follow-up and human verification, not from one confident-looking summary.