Deep Research mode in Claude Opus 4.7

# Deep Research mode in Claude Opus 4.7 You built a research prompt in the last workshop. Now you're going to run it. This workshop walks you through Claude's Research mode — how to enable it, what actually happens when it runs, and how to read the output you get back. The whole thing takes about 20 minutes because Claude is doing real work while you wait. ## What Research mode actually does Regular Claude chat gives you an answer from its training data. Research mode is different: Claude actively searches the web, reads sources, synthesizes across multiple documents, and produces a cited report. The output is longer, slower, and more verifiable. You can click through to the sources and check whether Claude's summary accurately represents what's there. That's the important capability — not that it searched the web, but that you can audit the reasoning. This isn't perfect. Claude can misread a source. It can weight a low-quality source too heavily. It can miss recent developments that aren't well-indexed. But for building an initial picture of a question you don't know well, it's substantially more reliable than asking a model to recall from training. ## Enable Research mode Go to [claude.ai](https://claude.ai) and open a new conversation. In the input area, look for the tool attachment icons — the paperclip, the web search toggle, or a "Research" option depending on the current interface. Click to enable Research mode. It may be labeled "Deep Research" or have a search-icon toggle. If you don't see it, check that you're on a plan that includes research features — the free tier may not have it, and Claude's interface evolves regularly. The core workflow here applies whenever you have access to it. ## Paste your prompt Take the research prompt you built in Workshop 2 — or write a fresh one if you're starting here — and paste it into the input box with Research mode enabled. Before you hit send, scan it one more time: - Is the question specific enough that Claude knows what counts as a good answer? - Is the date anchor present? - Does it specify what format you want? - Did you ask for citations? If yes to all four, send it. ## What happens next Claude will show you that it's working — usually a running display of the sources it's consulting and the queries it's running. This is worth watching for a minute. You'll see it searching, reading, and looping back to gather more evidence when an initial source is thin. Expect the run to take 3–8 minutes for a well-formed prompt. Don't close the tab. The session can time out if the browser loses focus for too long on some setups. ## Reading the output The report Claude produces will have several layers: **The main text.** This is the synthesis — Claude's answer to your research question, organized by theme or section. Read this first for the overall picture. **Inline citations.** Look for numbered references or linked footnotes in the text. These are the specific claims Claude is supporting with a source. **The sources list.** At the end (or in a sidebar) you'll find the list of sources Claude consulted. Scan this. Notice the types of sources — academic papers, news articles, company blog posts, government data. A good research output draws from multiple source types. **What to verify.** Pick 2–3 claims that matter most to your actual question and click through to the source. Does the source actually say what Claude says it says? Does the broader article context change the meaning of the claim? This 5-minute verification pass will save you from acting on a misread. ## Before vs. after a structured prompt The quality gap between a vague prompt and a structured one is visible in the output: **Vague:** "What's happening with AI in healthcare?" → Output: a broad overview of obvious trends, sparse citations, generic conclusions **Structured:** "What are the regulatory and reimbursement barriers that have slowed AI diagnostic tool adoption in US outpatient settings between 2022 and 2025? Provide a structured report with an executive summary, key findings, and a limitations section. Cite primary sources including FDA guidance documents where relevant." → Output: a focused, section-organized report with traceable citations and a usable answer to a specific question The structure you built in Workshop 2 is doing the work here. The model is capable of either output. Your prompt determines which one you get. ## Save the output Before you close the tab, copy the research report and save it somewhere — a Notion page, a Google Doc, a plain text file. You'll use it in Workshops 7 and 8 to synthesize it into a brief and a spec. Treat it as raw material, not a finished document. It's what you'll work with next. --- Now you can apply this to: - Evaluating a new market or space before committing development resources - Building a knowledge base on a topic where you're starting from near-zero - Producing a first draft of any research-heavy document (competitive analysis, due diligence, technology evaluation)