Claude: Deep Dive · Professional Workflows
Long-Document Mastery
Reason across long documents
Long-Document Mastery
You've made it to the final lesson of this course. Big congrats! You've come a long way — from running your first connector prompt to orchestrating multi-surface workflows.
This last lesson focuses on reasoning across long and complex documents. Claude shines when the task requires extracting structured data, comparing versions side by side, and finding contradictions.
Let's get started!
The Capability Claim
When you bring a large document to a chat with Claude, you're working with large context window. The whole document is in scope at once. Claude can reference page 87 while answering a question about page 12 without losing the context.
This is what makes long-document work feel different. The whole document and your questions about it live in the same conversation — no need to break it apart in chunks and feed it piece by piece.
Three Ways to Work With Long Documents
Working with long documents makes three kinds of tasks easier:
- Structured extraction: pulling obligations, dollar amounts, and deadlines, not a narrative summary
- Version comparison: finding material changes between two versions of the same document, with risk flags
- Contradiction detection: finding contradicting clauses, the kind of error that hides in plain sight when you read sequentially
These are the skills you'll work through in this lesson.
Scenario: Analyzing a Complex Contract
Imagine this: you've been offered a senior role at a new company. The offer landed in your inbox last night. The package is 95 pages across five files:
- The offer letter itself
- The equity plan document
- The company handbook
- The benefits summary
- The indemnification agreement
You have a week to respond. Before you sign anything, you need to know exactly what you're agreeing to:
- What gets paid when
- What's guaranteed vs. conditional
- What happens if you leave (voluntarily, involuntarily, after acquisition)
- Where the fine print might contradict the headline terms
Getting the Documents Into the Chat
You upload all five documents into a single chat — the same way you learned for File creation before (drag-and-drop or the attachment menu). Claude reads them as one combined context.
For best results on long-document work, start a fresh chat dedicated to this analysis. Don't add the documents into an existing conversation — keep the context window focused on the work at hand.
Summary vs. Structured Extraction
With the documents loaded, the temptation is to prompt as you would for any document: "Summarize the key terms of this offer."
This produces a readable summary. It is useful if you want to understand what's in the document at a high level, but not really when you're making a data-informed decision.
A different kind of prompt — a structured extraction prompt — produces actionable data. The difference shows up the moment you try to act on the output.
The Two Prompts, Side by Side
Here are the two prompts for the same offer package:
Prompt A — Summary: "Summarize the key terms of this offer letter and equity plan."
Prompt B — Structured Extraction: "For every financial commitment in this offer package, extract: the dollar amount, when it pays, whether it's guaranteed or conditional, and what conditions would cancel the payment. Output as a table."
Both prompts cover the same material, but they produce different kinds of output: a narrative summary and structured data.
Select all that apply
What do you think makes Prompt B drive the better output?
The Extraction Pattern
The pattern Prompt B uses works for long-document extraction when the goal is to turn prose into structured data:
- Scope: what set of items you want extracted (every obligation, financial commitment, and deadline)
- Fields: which attributes to extract for each item (named explicitly)
- Format: how the output should be structured (table, JSON, structured list)
When all three slots are filled, the output is easier to review, compare, and use.
practice preview
Interactive practice
Fill in the blank
Write the prompt for extracting all post-termination obligations from the offer package — anything that binds you after you leave the company.
Version Comparison
With the extraction in hand, you knew what you wanted to push back on — the obligations that were too restrictive and the conditions that needed to be tightened. You shared the extraction with your lawyer and gave them a list of changes to negotiate. A week later, a revised version comes back — same five documents, but with redlines and counter-proposals from your side.
Now you need to know exactly what changed between the two versions: the actual differences — every moved term, added or removed clause, and shifted number.
This is where the context window earns its place. Both versions, all five documents each, all loaded at once.
The Comparison Prompt
Comparison prompts have the same structural pattern as extraction prompts, with one addition: you name both versions, and you specify what counts as a meaningful change.
The trick: don't ask Claude to find "all differences." That returns formatting noise alongside substantive changes. Instead, name the categories of change that matter and ask for risk flags for changes that contradict your stated priorities.
practice preview
Interactive practice
Fill in the blank
Complete the prompt to compare the original document (V1) and the revised version (V2). Both are loaded as separate files.
Finding Contradictions
Once you trust the comparison, you can use Claude's full-document view to do something harder: find contradictions within the document itself.
Long contracts contradict themselves more often than you can expect. A clause on page 3 grants you a right; a clause on page 58 quietly takes it back. Both clauses can be perfectly written on their own — and incompatible the moment you try to act on them together.
What This Looks Like in Practice
An example from your offer package: the offer letter says you can work on outside projects with approval; the IP assignment clause says all work product belongs to the company.
Finding these by reading sequentially is hard, but finding them with the document loaded as one context is what Claude can help you with.
practice preview
Interactive practice
Fill in the blank
Complete the prompt to find contradictions across the offer package.
The Attention Drift Problem
Three contradictions found across nearly a hundred pages. But before you trust this output completely, there's one risk worth knowing about: attention drift.
When Claude analyzes a long document, the first few sections can get more rigorous treatment than the later ones. The extraction is more thorough, the risk flags more numerous, and the structure more careful. By section 12, the work might be visibly thinner.
What Drift Looks Like
Specific indicators that attention drifted across the document:
- Missing categories: Fields that appear consistently in early sections but stop appearing in later ones.
- Shorter entries: The same kind of finding gets a one-line treatment late that would have gotten a paragraph early.
- Absence of risk flags: Flags appear regularly in the first half, then disappear in the second half (when the second half should be no less risky).
- Generic language: Specific clause references early, vague references late.
- Skipped subsections: Entire sections of the document not represented in the output.
If you see any two of these in the same analysis, the analysis drifted. The fix is structural.
Choose one
You extracted data from all 95 pages. The first 40 produced 15 detailed entries, including 9 risk flags, while the last 55 produced only 6 entries and no flags. What’s the most likely explanation?
Verification and Chunking
When drift is detected, you verify the analysis by rerunning it on a specific section and comparing the results.
The verification prompt is short: "Re-run the same extraction on pages 60 through 95 only. Use the same fields. Output as a table." Compare the new output against what the full-document run produced for those pages. The new output will typically catch entries that the first pass missed.
For documents long enough that drift is likely, you can chunk the work upfront — analyze sections sequentially, then merge. The trade-off: chunking can lose cross-sectional references (contradiction-finding becomes harder). So the decision of when to chunk matters.
When to Chunk vs. Process in Full
Document length is only part of the decision. The other part is cross-reference density — how much one section depends on another.
Three quick rules:
- If short and reference-dense (under 30 pages with heavy cross-references), process in full, no chunking.
- If long and reference-light (sections are mostly independent of each other), chunk by section.
- If long and reference-dense (over 50 pages, sections refer to each other constantly), process in full first, then verify by re-running on later sections.
practice preview
Interactive practice
Match pairs
Tap a left item, then a right item to match.
When Long-Document Mastery Earns Its Place
Not every document needs this treatment. Single-page memos don't. Documents you'll skim once and discard don't. Where this skill earns its place:
- Documents you're committing to: contracts, agreements, terms you'll be bound by
- Documents with money attached: financial commitments hidden across narrative prose
- Documents you'll need to reference later: extracting once, referring back many times
- Documents that come in versions: anything where the comparison work matters as much as the content
When any of these is true, the extraction–comparison–verification pattern is often more reliable than reading the document from start to finish.
- Structured prompts usually produce clearer, more focused results than general ones. Name the scope, the fields, and the format.
- Comparison work is its own skill. Name both versions, name what counts as a meaningful change, and ask for risk flags.
- Contradictions hide in cross-references. Clauses that look fine alone can be incompatible together; the context window lets Claude find them.
- Attention can drift across long documents. Missing categories, shorter entries, vanishing risk flags are the signals.
- Chunking trades cross-section visibility for per-section attention. Process in full when cross-references matter, chunk when sections are independent and you want each one analyzed thoroughly.
Look What You've Built!
You've now worked across every major Claude surface:
- Chat with Connectors, reasoning modes, skills, memory, and Artifacts
- Claude Code
- Claude Cowork
- Claude Design
- Claude for Small Business
And you've learned the meta-skills that make them work together: workflow design, multi-tool execution, and now long-document mastery.
That's the full toolkit! The work from here is bringing it to your own projects — the contracts you're about to sign, the launches you're planning, the analyses you've been putting off, the work that's been waiting for a better tool.
You have everything you need, now go build something ambitious! Good luck!
