Claude Sonnet 4.6: Hidden Features Guide 2026

Most people open Claude, type a question, read the answer, and close the tab. That’s it. They’re using maybe 20% of what the model can actually do — and the 80% they’re missing is where the real productivity difference lives. I know because I did exactly the same thing for the first three months after switching from ChatGPT. Then I started actually reading the Claude.ai interface instead of just using it, and I found features that changed how I run my entire solo real estate operation in Madeira.

Claude Sonnet 4.6 is Anthropic’s workhorse model — not the flashiest release, but the one I run my business on because it’s fast, handles long documents cleanly, and costs a fraction of Opus for daily work. What most users miss isn’t in any changelog. It’s baked into the interface and the API in ways that are easy to overlook if you’re just here for quick answers.

Let me walk you through the features I actually use, what they do, and where they fall short.

What “Hidden Features” Actually Means Here

To be clear: I’m not talking about jailbreaks or secret modes. I’m talking about documented, officially supported capabilities that the average user never discovers because Anthropic’s UI doesn’t shout about them and most tutorial content online still covers Claude like it’s a basic chatbot.

These features fall into three categories:

• Interface features — things inside Claude.ai that most users skip (Projects, custom instructions, file upload workflows)
• Model behaviors — capabilities of Sonnet 4.6 itself that most prompting habits don’t activate (extended thinking, multi-step reasoning, structured output)
• API-level tools — tool use, web search integration, and document analysis that solo operators can access without being developers

Feature 1: Projects With Custom Instructions That Persist Across Sessions

This is the single most underused feature in Claude.ai Pro. Projects let you create a persistent workspace where Claude remembers context, files, and your custom instructions across every conversation inside that project. It’s not memory in the ChatGPT sense — it’s more structured than that.

Here’s what that looks like in practice for me. I have a Project called “Madeira Listings.” Inside it, I’ve uploaded my standard listing template, a style guide I wrote for property descriptions (tone, word count, what to emphasize for international buyers), and a one-page document about the Madeira real estate market I update monthly. Every time I start a new conversation inside that project, Claude already has all of that context. I don’t re-explain what I do or paste my template every single time.

The custom instructions field inside a Project is where this gets powerful. You can tell Claude to always write descriptions in a specific tone, always include certain legal disclaimers, always format outputs a certain way. Set it once, it applies to every conversation in that project forever.

Most users either don’t know Projects exist or they use them as simple conversation folders without setting up the instructions. That’s leaving 80% of the value on the table.

Feature 2: Extended Thinking Mode for Complex Analysis

Extended thinking is available on certain Claude models including Sonnet 4.6 via the API, and it’s beginning to surface in the Claude.ai interface for Pro users. When you activate it, Claude takes more time before responding — it runs through a chain of reasoning steps that are visible to you as a kind of thinking log before the final answer appears.

For quick tasks, you don’t want this. It’s slow. But for complex analysis — comparing multiple offers on a property, thinking through a pricing strategy, reviewing a contract for potential issues — it produces noticeably more careful output than standard prompting.

The practical way to trigger more deliberate reasoning without full API access is to ask Claude explicitly: “Think through this step by step before answering, and show me your reasoning.” That’s not a trick — it’s how the model was trained to allocate more of its reasoning capacity. Combined with Sonnet 4.6’s longer context window, you can paste in a 20-page market report and ask for a structured analysis with actual reasoning, not just a summary.

What it doesn’t do well: extended thinking is not a substitute for real-time data. Claude still has a knowledge cutoff, and no amount of careful reasoning will conjure up a property sale that happened last week. I’ll come back to this limitation in detail later.

Feature 3: Multi-Document Cross-Referencing Inside a Single Conversation

Claude Sonnet 4.6 can hold multiple uploaded documents in context simultaneously and reason across all of them at once. This isn’t new in principle, but most users upload one document, ask one question, and move on. The real power is uploading three or four documents and asking Claude to compare, reconcile, or synthesize across them.

For a real estate consultant, this is genuinely useful. I’ve uploaded a property’s purchase contract, the building inspection report, and a PDF of comparable sales in the same parish — then asked Claude to identify any clauses in the contract that become relevant given the inspection findings, and to flag pricing relative to the comparables. That’s three documents, one question, one coherent answer. Doing that manually takes an hour minimum.

The limit is roughly 200,000 tokens total context — which sounds enormous, but dense PDFs with tables and images can eat that up faster than you’d expect. And Claude cannot read scanned image-based PDFs without OCR pre-processing. If your documents are scanned paper files (common with older Portuguese property records), you need to run them through an OCR tool first.

Feature 4: Tool Use and Web Search via the API

Claude’s tool use capability lets the model call external tools — web search, custom APIs, calculators, your own data sources — as part of answering a question. In the Claude.ai interface, this surfaces as the web search toggle (available to Pro subscribers). In the API, developers can define their own tools and Claude will decide when to call them.

For solo operators who aren’t developers, the web search toggle in Claude.ai is the accessible version of this. Turn it on and Claude can pull current information — today’s listings on Idealista, current exchange rates for UK buyers in Madeira, recent news about Portuguese property law changes. This makes a significant difference for anything market-related.

What most users miss is that you can combine web search with a specific research prompt structure to get something closer to a mini market report. Instead of “What’s the property market like in Madeira?”, try: “Search for current residential property prices in Funchal, Madeira, and compare them to the historical average from 2022–2024. Present the data in a table and flag any recent regulatory changes that affect foreign buyers.” Claude will use web search, synthesize across sources, and produce a structured output — not just a list of links.

Feature 5: Structured Output Formatting You Control

Claude will produce JSON, markdown tables, HTML, CSV-formatted text, XML — whatever structured format you need — if you tell it to. Most users don’t ask. They get a prose paragraph and copy-paste it somewhere, losing hours a month to reformatting.

Via the API, you can specify structured output schemas and Claude will conform to them reliably. In the standard interface, explicit instructions in the prompt work well. “Return your answer as a JSON object with the keys: property_address, asking_price, key_features (array), buyer_profile, suggested_listing_description.” It follows the schema cleanly.

I use this constantly for lead intake processing. When a new inquiry comes in with a long email from a prospective buyer, I paste it into Claude with a structured output prompt and get back a formatted client profile I can drop straight into my CRM. Thirty seconds versus five minutes of manual data entry.

Feature 6: System Prompt Behavior in Multi-Turn Conversations

This one is more nuanced. Claude maintains a consistent persona and set of constraints throughout a multi-turn conversation better than most models — but only if you establish those constraints clearly at the start. Most users just start typing. If you open a conversation with a clear framing — “You are acting as a Portuguese real estate market analyst. Your outputs should be formal, accurate, and avoid speculative claims. When you don’t have data, say so clearly.” — Claude holds that frame across the entire conversation, including follow-up questions.

This is most useful for client-facing work. I’ll set up a conversation with a specific persona framing, generate a full market report draft, then ask follow-up questions to refine specific sections — and Claude stays in the same analytical register throughout. It doesn’t drift into casual language or suddenly start hedging everything with “as an AI.”

Combined with Projects (Feature 1), you can bake the persona framing into your project instructions so it’s always active without re-prompting.

Hidden Feature Comparison: What’s Available Where

Feature	Claude.ai Free	Claude.ai Pro ($20/mo)	API Access
Projects + Custom Instructions	Limited	✅ Full access	✅ Via system prompts
Multi-document upload	Limited	✅ Up to ~200k tokens	✅ Full context window
Web search (tool use)	❌	✅ Toggle available	✅ Custom tool definitions
Extended thinking mode	❌	Partial (select models)	✅ Full control
Structured output (JSON/HTML)	✅ Via prompting	✅ Via prompting	✅ Enforced via schema
Persistent persona framing	✅ Per conversation	✅ Per project	✅ Via system prompt

My Real-World Experience: 90 Days Using These Features in Madeira Real Estate

I want to be specific here, because vague testimonials are useless. So here’s exactly what happened over the last quarter.

In January 2026 I had a brutal month — 17 active listings, three buyers in due diligence simultaneously, and two market reports to produce for investor clients. My previous workflow involved writing listing descriptions one at a time, starting each Claude conversation from scratch, re-pasting my style guide every single time. I was spending roughly 3.5 hours a week just on property description drafts alone.

I set up the “Madeira Listings” Project I mentioned earlier — uploaded my template, my style guide, and a one-page brief on what international buyers (primarily British, German, and Scandinavian) actually look for in Madeira properties. Added custom instructions telling Claude to always produce descriptions in English first, flag any features that carry specific legal notes in Portugal (like non-licensed annexes), and keep descriptions to 180–220 words. Took me about 40 minutes to set up properly.

The following week I processed all 17 listing descriptions through that project. Total time: 52 minutes. That includes the time to review each one, make edits, and copy them into my listing platform. The week before the setup, the same volume of work took me 3 hours and 20 minutes. That’s not a rounding error — that’s 2.5 hours recovered in a single week, from one feature I’d been ignoring for four months.

The multi-document cross-referencing feature came into play for the investor market reports. I had three Portuguese government PDFs on property transaction statistics, a regional tourism impact report, and a bank mortgage rate summary — five documents total, about 140 pages combined. I uploaded all five to a single Claude conversation, asked for a structured analysis comparing residential yield trends in the eastern versus western parishes of Madeira over the past three years, with specific attention to short-term rental regulation impacts.

Claude produced a 900-word structured analysis in under two minutes. I spent another 45 minutes verifying figures against the source PDFs and adding my own commentary — but the bones of the report were done. Previously I would have spent 4–5 hours just extracting and organizing the data manually before writing a word.

I also tested the structured output feature heavily for lead processing. I receive inquiries through four channels — email, WhatsApp (forwarded to email), my website form, and one partner portal. Formats vary wildly. I now paste each new inquiry into a Claude conversation with a fixed structured output prompt and get back a standardized client profile in under 30 seconds. Over 90 days I processed 61 new leads this way. Estimated time saved versus manual CRM entry: about 4 hours total.

I rate Claude Sonnet 4.6 for solo real estate work at 8.5/10 — it earns that score specifically because the Projects + multi-document combination handles the research and drafting bottlenecks that were eating my Tuesdays, though the lack of real-time Portuguese property registry data keeps it from replacing the manual verification step I still have to do.

The Real Limitations I Hit After 90 Days of Daily Use

No tool review from me is complete without the honest part.

Real-time data is still a wall. Claude has a training cutoff. Even with web search enabled, it can’t pull live data from Portugal’s property registry (Registo Predial), the official notary transaction database, or IMI (municipal property tax) records. For market analysis that matters legally or financially, I still have to verify everything manually through official sources. Claude’s web search surfaced useful public data — news articles, published statistics, portal aggregations — but that’s not the same as official registry data. Any consultant using Claude for client-facing market reports needs to be clear about this distinction.

The Projects feature has a file storage limit that I hit unexpectedly. I was adding PDFs month over month and eventually started getting errors when adding new documents. You need to manage what’s in a project actively — archive or remove older files rather than treating it as infinite storage.

Extended thinking via the interface is inconsistent. I couldn’t always predict when the slower, more deliberate reasoning mode would activate versus standard response mode. Via the API it’s controllable, but in the UI it’s murkier. For important analyses I now use explicit step-by-step prompting rather than relying on the model to decide how much thinking to do.

Portuguese language performance is good but not perfect. For Madeira-specific property terminology and local legal phrasing, Claude occasionally produces technically correct Portuguese that sounds slightly non-native. I always have a Portuguese speaker do a final pass on anything client-facing in the local language. This is a real cost to factor in if you’re using Claude for multilingual real estate markets.

How to Start Using These Features This Week

You don’t need API access to get 80% of this value. Here’s the practical starting sequence for a solo operator on Claude.ai Pro:

1. Create one Project for your core business function. For me it’s listings. For you it might be client proposals, content creation, or financial analysis. Give it a name, write custom instructions that reflect your actual standards and style, and upload 2–3 reference documents you currently re-paste into every chat.
2. Test multi-document analysis by uploading three related documents and asking Claude to synthesize across them. The prompt structure that works best for me: “You have access to [Document A], [Document B], and [Document C]. Based on all three, answer the following: [specific question]. Flag any contradictions between the documents.”
3. Set up one structured output template for your most repetitive task. Write the prompt once, save it in a notes app, paste it every time. Don’t over-engineer this — a simple JSON or table structure is enough to save meaningful time.
4. Enable web search for any market research conversation and combine it with a specific research prompt structure rather than a generic question. Specificity drives quality.

If you want API access for extended thinking and full tool use control, Anthropic’s API pricing for Sonnet 4.6 is usage-based — at roughly $3 per million input tokens and $15 per million output tokens as of 2026, it’s affordable for low-to-moderate solo operator usage. You’ll need a developer wrapper or a no-code tool like Make.com to use it without writing code.

Practical Summary: The Features Worth Your Time

Six features. Three of them — Projects with custom instructions, multi-document cross-referencing, and structured output prompting — are available to every Pro subscriber and require no technical setup. They’re the ones that recovered 2.5 hours of my week in January alone.