My Real Claude Research Workflow (Not Prompt Tricks—Actual Process)

I've watched too many people use Claude for research: open a chat, type a question, wait for the answer, copy and paste. Then complain that "AI only gives surface-level information."

The problem isn't Claude. It's the workflow.

Stage One: Problem Decomposition (Don't Skip This)

Most research failures happen in the first minute: the question is too broad. "Help me research the latest developments in AI" is a dead-end question.

My first step isn't to ask Claude a question—it's to work with Claude to break the question into answerable pieces:

I want to research [topic] for the purpose of [specific use].
Help me decompose this into 5-8 sub-questions that can each be answered independently.
Then tell me which 3 are core, and which are peripheral.

The key is "specific use." Researching "brain-computer interfaces" for a science article generates completely different core sub-questions than researching it to evaluate investment opportunities.

Stage Two: Iterative Digging (It's About Follow-Up, Not Questions)

Counterintuitive suggestion: Start with the sub-question you know least about, not the most important one. Your most familiar sub-questions are exactly where your blind spots live.

Follow-up rhythm:

1. Pose the sub-question, get Claude's initial response
2. Pick 1-2 claims you're not sure about
3. Follow up: "You just said XX. What's the basis for that? Is there contrary evidence?"
4. From the follow-up, pick 1 more point, keep digging
5. Keep going until Claude says "there isn't enough public data to support a more specific conclusion"—that's the real edge

Concrete example: researching commercial prospects for MEG devices. Three levels of follow-up revealed the hardest problem isn't superconducting cooling—it's standardization of signal processing software. The first-level response never mentioned this.

Stage Three: Cross-Verification (Assume Everything Claude Told You Is Wrong)

Claude's training data has a cutoff date, and it will confidently say incorrect things—that's the nature of language models.

Check every number. Any claim with a number—accuracy rates, market size, dates—find the original source.

Flag "consensus" claims. When Claude says "researchers generally believe XX," that sentence is most dangerous. Sometimes it's describing a consensus from five years ago that newer findings have challenged.

Find the opposing view. Ask: "For the conclusion above, are there serious researchers who hold the opposite position? What's their basis?"

Tools: ArXiv, Google Scholar, Perplexity. Every core conclusion gets at least one primary source.

Time Allocation

For a 4-hour research session: 30 min decomposition + 2 hours digging + 1.5 hours verification. Most people pile time into the digging phase, but light verification undercuts the credibility of everything produced.

One Final Counterintuitive Suggestion

After research is done, ask Claude to find your blind spots:

Compile your research conclusions, paste them to Claude, and say:
"As a critic, what's the most fragile assumption in this research?
Which conclusion has the weakest supporting evidence?"

Since starting this step, research quality improved noticeably. The places where you can't see problems yourself are usually exactly the places that need to be challenged.

SFD Editor's Note

SFD Lab has a dedicated research agent, Owl 🦉, handling all deep research tasks. But Owl's workflow matches what this article describes—decompose, iterate, verify, no step skippable. Our biggest mistake was skipping step one and diving straight into searching, ending up researching in the wrong direction and spending twice the time. The decomposition step is worth 30 minutes every time.