Claude Opus 4 Deep Research: The Right Way

Last March, a German couple contacted me about buying a quinta in the interior of Madeira. Before I could even schedule a call, I needed to understand the micro-market around Câmara de Lobos — price trends, planning restrictions, comparative sales data, foreign buyer regulations, and the specific legal quirks around rural agricultural land. That research used to take me a full day. I did it in under two hours using a Claude Opus 4 deep research workflow I’d spent three weeks building. That workflow is exactly what I’m walking you through here.

This isn’t a feature overview. It’s a step-by-step tutorial showing you the exact process, prompts, and settings I use in my one-person real estate consulting business in Madeira. If you run any kind of solo knowledge business — consulting, research, analysis, content — this workflow applies directly to you.

What You’ll Build in This Tutorial

By the end of this tutorial, you’ll have a repeatable Claude Opus 4 deep research workflow that:

• Takes a complex research topic and breaks it into structured sub-questions automatically
• Uses Claude’s extended thinking to reason through contradictions and gaps in source material
• Produces a structured, citation-ready research brief you can use directly with clients or in reports
• Can be templated and reused for different property types, client profiles, or market segments

Time investment to set it up: about 90 minutes the first time. Ongoing time per research session: 45 minutes to 2 hours, depending on complexity — down from a full day of manual work.

Prerequisites Before You Start

You need the following before running this workflow:

1. Claude Pro or Claude Team subscription — Opus 4 is not available on the free tier. As of 2026, Claude Pro costs $20/month and gives you access to Opus 4 with extended thinking enabled. Without extended thinking, you lose roughly 60% of the workflow’s value.
2. A clear research objective — Vague inputs produce vague outputs. Before you open Claude, write one sentence that describes exactly what decision this research supports.
3. Source documents ready to upload (optional but recommended) — PDFs, market reports, planning documents, legal texts. Claude Opus 4 handles files well. I typically upload 3–6 documents per session.
4. A text editor or Notion page open — You’ll be copying outputs and building a research brief as you go. Don’t rely on the chat history alone.

Step-by-Step: The Claude Opus 4 Deep Research Workflow

Step 1 — Write a One-Sentence Research Brief

This is the step most people skip, and it’s why their AI research comes back generic. Before you type anything into Claude, write this sentence by hand:

“I need to understand [topic] so that I can [specific decision or action].”

My example for the Câmara de Lobos quinta: “I need to understand current market pricing and legal restrictions on rural agricultural land in Câmara de Lobos municipality so that I can advise a German couple on realistic purchase price and timeline.”

That sentence does three things: it defines the topic, anchors the purpose, and sets the audience context. All three matter when you’re prompting Claude.

Step 2 — Open Claude Opus 4 and Enable Extended Thinking

Go to claude.ai and select Opus 4 from the model selector. Click the “Extended thinking” toggle — it appears below the message input box when Opus 4 is selected. Extended thinking tells Claude to reason through the problem in a scratchpad before writing its response. For complex research, this is not optional. The quality difference is significant.

Start a new conversation. Do not carry over threads from other topics. Fresh context keeps the model focused.

Step 3 — Run the Research Decomposition Prompt

Paste in the following prompt, replacing the bracketed sections with your actual research brief:

I'm conducting research on the following topic:

[PASTE YOUR ONE-SENTENCE RESEARCH BRIEF HERE]

Before giving me any answers, I want you to do the following:

1. Break this research topic into 5–8 specific sub-questions that, if answered well, would fully address my brief.
2. For each sub-question, tell me: (a) what kind of source would best answer it, (b) what assumptions you'd need to flag if no source is available, and (c) how confident you are that you can answer it from your training data alone (scale: high / medium / low).
3. Identify any tensions or contradictions you'd expect to find across different sources on this topic.

Do not answer the sub-questions yet. Just give me the decomposition.

This prompt does something critical: it makes Claude map the territory before entering it. You get a structured list of sub-questions, a confidence map, and a heads-up on where the research might have gaps or contradictions. I use this output to decide which documents I need to upload in the next step.

Step 4 — Upload Your Source Documents

Look at the sub-questions Claude generated. For any question marked “low” confidence or flagged as needing current data, you need to provide documents. For real estate in Madeira, my standard upload set includes:

• The relevant municipal PDR (Plano Diretor Regional) excerpts for the area
• Any recent APEMIP market data I have on file
• Notarial transaction records or CMI price data from the Finanças portal, exported as PDF
• Client-supplied documents (existing survey reports, property registers)

Upload these directly in the same Claude conversation thread. Then run this follow-up prompt:

I've uploaded [NUMBER] documents. Before we proceed to the full research, please:

1. Confirm you can read each document and briefly describe what each one contains (1 sentence per document).
2. Tell me which of the sub-questions from your previous decomposition are now answerable using these documents.
3. Flag any sub-questions that are still under-resourced.

This verification step takes 2 minutes and prevents a frustrating situation where Claude hallucinates document content it can’t actually read. I learned this the hard way after a planning document upload failed silently and Claude proceeded to answer from training data without flagging it.

Step 5 — Run the Deep Research Prompt Per Sub-Question

Now you answer each sub-question systematically. Do them one at a time, not all at once. This keeps responses focused and lets you redirect if something goes off track.

Now answer sub-question [NUMBER]: [PASTE THE SUB-QUESTION TEXT]

Instructions:
- Use the uploaded documents as your primary source. Quote directly where relevant, including page or section references.
- Where you're drawing on training data rather than the documents, say so explicitly.
- If you find contradictions between sources, describe the contradiction rather than resolving it silently.
- End your answer with a confidence rating: High / Medium / Low, and one sentence explaining why.
- Keep your answer under 400 words unless complexity genuinely requires more.

Run this prompt for each of the 5–8 sub-questions. Copy each answer into your Notion page or text editor as you go.

Step 6 — Run the Synthesis Prompt

Once all sub-questions are answered, run this final synthesis prompt:

You've now answered all [NUMBER] sub-questions. Please synthesize the findings into a structured research brief with the following sections:

1. Executive Summary (3–5 sentences: the single most important finding and its implication for the decision stated in my brief)
2. Key Findings (bullet points, organized by sub-question theme)
3. Confidence Map (a simple table: sub-question | confidence level | primary source)
4. Gaps and Caveats (what we don't know, what needs verification from a local professional or current data source)
5. Recommended Next Steps (3–5 specific actions I should take based on this research)

Write this for a professional audience. Avoid hedging language unless it reflects a genuine uncertainty captured in the research.

This gives you a document you can actually use — send to a client, use in a meeting, or build a full report from. The confidence map alone has saved me from presenting uncertain findings as facts more than once.

Step 7 — Save the Workflow as a Reusable Template

Save all six prompts — the decomposition, the document verification, the per-sub-question deep dive, and the synthesis — as a single template in Notion or a plain text file. The only parts that change each time are your one-sentence research brief and the number of sub-questions. Everything else stays identical.

I keep mine in a Notion database called “Research Workflows” with a separate entry for each property type I work with: residential, rural/agricultural, commercial, and holiday rental. Each entry has the base prompts plus notes on which documents to typically upload for that category.

My Real-World Experience Using This Workflow in Madeira

I’ve run this workflow 23 times since I finalized the template in January 2026. Here’s what actually happened with the Câmara de Lobos quinta case I mentioned at the start.

The German clients — a couple in their early fifties looking for a semi-rural retreat with rental income potential — needed me to give them a realistic picture fast. They’d been burned by an agent in the Algarve who promised them a rural property was legal for tourist rentals, only to discover 3 months in that the planning classification made that impossible.

My one-sentence brief: “I need to understand current market pricing, legal classification, and Alojamento Local licensing feasibility for rural agricultural land and quintas in the Câmara de Lobos municipality so that I can advise German buyers on realistic purchase price, holding costs, and rental income potential.”

Claude’s decomposition returned 7 sub-questions. Three were flagged as low confidence without documents, so I uploaded the Câmara de Lobos PDR excerpts (agricultural zoning section), a CMI price printout I’d pulled for rural properties sold in that municipality in the previous 18 months, and a summary document I’d written after consulting a local lawyer on Alojamento Local rules for rural properties in 2026.

Total research session: 1 hour 48 minutes. The synthesis brief ran to 4 pages. I sent it to the clients the same afternoon, formatted as a PDF with a cover note. They told me it was the most thorough pre-purchase briefing they’d ever received from a real estate consultant. They made an offer on a property 6 weeks later.

Before I built this workflow, that same level of research took me 6–8 hours spread across two days — pulling data manually, cross-referencing planning documents, writing notes, and structuring them into something client-presentable. The workflow cut that to under 2 hours. Over 23 sessions, that’s roughly 100 hours recovered since January 2026. At my consulting day rate, that’s significant. More importantly, I can now take on more complex buyer briefs without the research backlog killing my schedule.

The workflow also forced me to be more honest about gaps. The confidence map in the synthesis step regularly surfaces things I’d previously glossed over — moments where I was presenting “likely” findings as “confirmed” ones. That transparency has made my client relationships more solid, not less.

What This Workflow Does NOT Do Well

I want to be direct about three real limitations I’ve run into.

Real-time data is a hard wall. Claude Opus 4’s training has a cutoff, and no amount of prompting gets you current transaction prices, live mortgage rate data, or last week’s planning decisions. For anything where recency matters — and in real estate, that’s often — you still need to manually pull current data and upload it. The workflow helps you process that data faster, but it doesn’t source it for you.

Portuguese-language documents perform worse than English ones. I’ve uploaded Portuguese legal documents and planning texts dozens of times. Claude handles them, but the precision of its extraction drops noticeably compared to English documents of similar complexity. I now add a step where I ask Claude to confirm its reading of key Portuguese legal phrases before relying on them. Adds 10 minutes but prevents expensive misreadings.

Extended thinking sessions eat tokens fast. A full 7-sub-question deep research session with extended thinking enabled will hit context limits if you’re not careful. I’ve had two sessions cut off mid-synthesis. The fix is to start the synthesis in a fresh conversation, pasting in a condensed summary of the sub-question answers rather than the full chat history.

Comparing Research Approaches: Manual vs. This Workflow vs. Perplexity

Approach	Time per Session	Source Handling	Output Quality	Real-Time Data	Monthly Cost
Manual research	6–8 hours	Excellent	Excellent	Yes	$0 (but time cost is high)
Claude Opus 4 workflow (this tutorial)	1.5–2 hours	Good (with uploads)	Very good	No (you supply data)	$20 (Claude Pro)
Perplexity Pro	30–60 minutes	Moderate (web only)	Good for overviews	Yes	$20
ChatGPT Deep Research	45–90 minutes	Good (web + uploads)	Good	Yes (with browsing)	$20–$200

I use Perplexity for quick market overviews and current-events context before a client call. I use this Claude workflow for anything that requires reasoning across private documents or nuanced regulatory analysis. They complement each other.

Troubleshooting the Most Common Problems

Claude ignores the uploaded documents and answers from training data. First, verify the upload actually succeeded — refresh and try again. If it did succeed, add this line to your prompt: “Do not answer this question from your training data if the answer can be found in the uploaded documents. If you cannot find it in the documents, say so explicitly before drawing on training data.”

The sub-questions are too broad or too generic. Your one-sentence research brief is probably too vague. Add the specific geography, property type, buyer profile, and decision at stake. The more context you put in the brief, the more specific the decomposition.

Extended thinking produces very long, repetitive reasoning. This happens with poorly scoped prompts. Add a word limit to your prompt: “Keep your total response including reasoning under 600 words.” The reasoning scratchpad isn’t shown to you by default anyway — what you see is the final output.

The synthesis brief is too long to use with clients. Add to the synthesis prompt: “The executive summary should be suitable for sending directly to a non-technical client. Use plain language and avoid jargon.” You can also ask Claude to produce a one-paragraph client-facing summary separately from the full internal brief.

Context window cuts off mid-session. As mentioned above — run the synthesis in a fresh conversation. Paste in your sub-question answers as a numbered list, not the full chat. Keep each sub-question answer under 300 words when you copy it over.

My Rating and Practical Summary

Rating: 4.5/5 — The workflow consistently produces client-ready research briefs in under 2 hours that previously took me a full day, which at my consulting rate makes the $20/month subscription cost irrelevant; the half-point deduction is for the real-time data gap, which still requires manual sourcing on my end.

Here’s what to remember:

• Start with a one-sentence research brief. Non-negotiable.
• Use the decomposition prompt before asking Claude to answer anything. This single step is where most of the quality comes from.
• Upload your own documents for any low-confidence sub-questions. Don’t let Claude guess at current data.
• Run sub-questions one at a time, not in batch.
• Save the whole workflow as a template you reuse. The setup cost is 90 minutes. The ongoing savings are 4–6 hours per research session.
• Run the synthesis in a fresh conversation if your session is long.
  

  Robson Penassi

  Real estate consultant in Madeira, Portugal. Solopreneur since 2012. Testing AI tools since 2023 to automate his one-person business. Writes about what actually works — and what does not.

  More articles by Robson →