Chapter 9: Document Intelligence

10 min read

What It Is

Claude Code can read, analyze, synthesize, and create documents across multiple formats. PDFs, spreadsheets, Word documents, presentations, text files, and even unconventional formats like Parquet data files — Claude can process them all. More importantly, it can work with multiple documents simultaneously, cross-referencing between them, aggregating data across files, and producing structured outputs.

This goes far beyond “summarize this PDF.” Claude Code can extract specific data points from a stack of invoices, compare findings across multiple research reports, convert between formats, and create new documents synthesized from multiple inputs.

Why It Matters for Non-Technical Professionals

Every knowledge worker drowns in documents. Contracts, reports, invoices, proposals, slide decks, spreadsheets, memos, research papers. The time spent reading, extracting, comparing, and synthesizing information from documents is a significant portion of any professional’s week.

Claude Code transforms documents from something you read into something you query. Instead of reading sixty pages of a research report, you ask Claude to extract the three most surprising findings. Instead of manually tabulating ten months of invoices, you ask Claude to aggregate the line items and break down spend by category. Instead of re-reading five competitive analyses, you ask Claude to synthesize the common themes.

What You Need

Plugins: - document-skills plugin — For creating PDFs, DOCX, XLSX, and PPTX files - react-pdf plugin — For creating custom-designed PDF documents

No MCPs required. Document processing uses Claude Code’s core file-reading capabilities.

The Workflows

Workflow 1: Invoice Analysis Across Multiple PDFs

This is one of the most immediately practical workflows in the entire playbook. Any professional who manages expense reports, vendor invoices, or reimbursement processing will recognize the pain this solves.

As a Claude prompt (simple version):

Go through the invoices folder at:
/Users/krishna/Desktop/Finance/Agency-Invoices/

Scan all PDF invoices from January to October. For each invoice:
1. Identify the month and date
2. Extract all line items with amounts
3. Classify each line item as either "Retainer" or "Production"

Then give me:
- A summary table showing monthly retainer vs. production costs
- The cumulative annual spend by category
- Flag any months where invoices are missing
- Identify the top 5 biggest production cost line items across
  all months

As a Claude prompt (mixed-currency reimbursement version):

I have 22 PDF invoices in this folder:
/Users/krishna/Desktop/Finance/Jan-Reimbursements/

These are office reimbursement invoices. Some are straightforward
INR payments, but others are forex (USD) transactions that include:
- The base rupee debit amount
- A forex markup fee
- IGST (GST) on the markup fee

For each invoice, extract:
1. Invoice name/description
2. Dollar value (if applicable)
3. Total INR amount (combining base + markup + IGST for forex
   transactions)

Create a table saved as a file. Then add a breakdown column showing
the component figures for forex transactions (e.g.,
"₹15,000 + ₹450 + ₹81") with addition symbols between them.
Leave this column blank for single-amount INR-only payments.

What Claude does: 1. Lists all PDF files in the folder 2. Reads each invoice, extracting amounts and identifying currency type 3. For forex invoices: separately extracts the base debit, markup fee, and IGST 4. Computes the combined INR total for forex transactions 5. Builds a table with all invoices, dollar values, rupee totals, and a formula-style breakdown column 6. Saves the output file

Real example: I dropped twenty-two office reimbursement invoices into a folder — a mix of direct INR payments and forex USD transactions with markup fees and IGST. Claude read every PDF, identified which were forex and which were INR-only, extracted all the component values, and created a table. For forex invoices, the breakdown column showed the exact computation (e.g., ₹15,000 + ₹450 + ₹81 = ₹15,531), while INR-only payments had a single amount. This would have taken over an hour manually — Claude finished in under five minutes with 100% accuracy on every extracted number.

Why the breakdown column matters: For reimbursement submissions and audit trails, showing the individual components — base amount, markup, and tax — is often required by finance teams. Asking Claude to produce this in a second pass (after the initial table) is a good example of iterative refinement: start with the core data, then add the presentation layer that makes it submission-ready.

Workflow 2: Long Report Synthesis

As a Claude prompt:

Read this research report:
/Users/krishna/Desktop/Reports/sector-outlook-2025.pdf

It's approximately 60 pages. Produce a synthesis that:
- Captures the 8-10 most important findings
- Notes specific data points and statistics
- Identifies the report's key thesis and supporting arguments
- Highlights any predictions or forecasts
- Flags anything that contradicts conventional wisdom
- Formats the output in my one-pager style (scannable, bold
  topic headers, concise bullets)

Save as RTF.

What Claude does: 1. Reads the entire PDF (Claude can handle documents of 100+ pages) 2. Identifies the most important findings 3. Extracts specific data points 4. Structures everything in the requested format 5. Saves as RTF

Real example: I synthesized a sixty-plus page published investment research report into a structured one-pager. Claude identified the non-obvious insights, extracted key statistics, and formatted everything in my specific one-pager style. I then used this synthesis as the basis for a LinkedIn post and supporting infographics.

Workflow 3: Merging and Converting Documents

As a Claude prompt:

I have a folder of presentation slides exported as individual images:
/Users/krishna/Desktop/Presentation/

Merge all images (1.png through 11.png) into a single PDF.
The slides should be in filename order, each image as a full page.

What Claude does: 1. Reads the folder and sorts by filename 2. Converts each image to a PDF page 3. Merges into a single document

Other common conversions: - CSV to formatted Excel with headers - Markdown to PDF - Multiple text files to a single document - Image files to PDF - Parquet to CSV (for data files)

Workflow 4: Comparative Document Analysis

As a Claude prompt:

I have two year-end sector newsletters:
/Users/krishna/Desktop/Content/consumer-newsletter-2025.pdf
/Users/krishna/Desktop/Content/fintech-newsletter-2025.pdf

For each newsletter:
- Summarize the key themes and predictions
- List all companies mentioned and the context
- Identify any data points or statistics cited

Then compare:
- What themes appear in both sectors?
- Where do the outlooks diverge?
- Which newsletter makes bolder predictions?

Save each summary separately, plus a combined comparison document.

What Claude does: 1. Reads both documents 2. Produces individual summaries 3. Cross-references for common themes 4. Writes a comparative analysis 5. Saves three separate files

Workflow 7: Generating Styled Reports from Free APIs

Beyond analyzing existing documents, Claude Code can pull live data from free APIs and produce polished PDF reports — without any paid services.

The pattern is: free API → data collection → processing → styled PDF output. Once you see this pattern, you will find applications everywhere.

As a Claude prompt:

Generate a professional PDF report with the following:

1. Pull the latest data from [free public API — e.g., Open-Meteo
   for weather, a government statistics API, a public financial
   data API]
2. Process the data: compute averages, identify trends, flag
   anomalies
3. Create a styled PDF report with:
   - A cover page with title and date range
   - A summary section with key metrics
   - A detailed section with charts and tables
   - A sources/methodology section

Use reportlab for PDF generation. The style should be clean and
professional — dark headers, clear typography, data tables with
alternating row colors.

Save to /Users/krishna/Desktop/Reports/weekly-report.pdf

What Claude does: 1. Writes a Python script to fetch data from the API (using requests) 2. Processes the data — calculations, trend analysis, formatting 3. Uses reportlab (a free, open-source Python library) to generate a styled PDF 4. Applies professional formatting — cover page, headers, tables, page numbers 5. Saves the final PDF

Why this matters: This is the foundation for recurring automated reports. Any data source with a free API becomes a report you can generate on demand: - Government economic data (via MoSPI, FRED, World Bank APIs) - Weather and climate data (via Open-Meteo — no API key required) - Public company financial data (via free financial APIs) - Public health data (via WHO, CDC, or NCBI APIs) - Social media metrics (from your own exported data)

The first time you run it, Claude writes the script. After that, the same script can be re-run to generate updated reports — turning a one-time effort into a recurring capability.

Pro Tips

Tip 1: For large documents, be specific about what you need. “Summarize this 100-page report” will give you a generic summary. “Extract all data points about market size projections in this report” gives you exactly what you need.

Tip 2: Use Claude’s accuracy, then verify. For financial documents (invoices, budgets), Claude’s extraction accuracy is remarkably high — in my experience, effectively 100% for clearly formatted invoices. Still, spot-check critical numbers.

Tip 3: Process documents in batches. If you have twenty invoices, do not process them one at a time. Point Claude at the folder and ask it to process all of them in a single sweep.

Tip 4: Save document analysis for future reference. Ask Claude to save its analysis as a separate markdown file. This creates a searchable, text-based version of insights from documents that might otherwise be locked in PDF format.

Tip 5: Use the /document-skills:pdf skill for polished output. When you need to create a new PDF rather than just analyze an existing one, the document-skills plugin includes a /document-skills:pdf skill that generates professional-quality PDFs with custom layouts, fonts, and formatting. Perfect for creating reports, proposals, and deliverables.

Tip 6: Chain document processing with other workflows. The real power of document intelligence appears when you chain it with other capabilities. For example:

As a Claude prompt:

1. Read the quarterly financial report at /Reports/Q4.pdf
2. Extract all revenue figures and growth percentages
3. Create a line chart showing revenue trajectory
4. Write a one-paragraph executive summary
5. Combine the chart and summary into a new one-page PDF
6. Save both the PDF and the raw data as CSV

This chain — read document, extract data, visualize, synthesize, produce new document — is one of the most common professional workflows. Claude handles it in a single instruction.

Workflow 5: Bulk Document Processing Pipeline

When you need to process many documents of the same type, Claude excels at batch operations.

As a Claude prompt:

I have a folder of client proposals at:
/Users/krishna/Desktop/Proposals/

Each PDF follows a similar structure: executive summary, scope
of work, timeline, and pricing.

For each proposal:
1. Extract the client name, project scope (2 sentences), and
   total budget
2. Identify the proposed timeline

Then compile everything into a summary spreadsheet (XLSX) with
columns: Client, Scope, Budget, Timeline, Filename.
Sort by budget descending.

What Claude does: 1. Scans the folder and identifies all PDF files 2. Reads each document, extracting structured data 3. Handles variations in formatting across documents 4. Creates a clean spreadsheet with all extracted data 5. Sorts and formats for easy review

This pattern works for any batch of similarly-structured documents — contracts, invoices, resumes, research papers, compliance documents.

Workflow 6: Creating Polished Documents from Scratch

Claude Code is not just about reading documents — it can create them from scratch with professional formatting.

As a Claude prompt:

Create a project proposal PDF for [project description].
Use the /document-skills:pdf skill for professional formatting.

Include these sections:
1. Executive Summary (1 paragraph)
2. Problem Statement
3. Proposed Solution
4. Timeline (table format: phase, duration, deliverables)
5. Budget Overview (table format)
6. Team and Contact

Use a clean, professional design. Blue accent color.
Include a header with the project title on each page.

What Claude does: 1. Invokes the PDF skill for professional layout 2. Generates content for each section 3. Formats tables, headers, and page structure 4. Produces a polished PDF ready for distribution

The document-skills plugin supports PDF, DOCX, XLSX, and PPTX formats — so you can produce whatever format your audience expects.

Workflow 8: The Deck Rebrand — Redesign Every Slide, Change Nothing

This workflow closed the biggest format gap in my document work: PowerPoint. In June, a colleague’s data-heavy deck landed with me — every slide had to match the brand, and not one number was allowed to move.

Picture the equivalent in your world. Your colleague’s quarterly business review arrives in the template from three rebrands ago: thirty slides of revenue tables and funnel charts in tired grey, while the brand guide says deep forest green, cream backgrounds, and the logo bottom-right on every slide. The content is finished; only the design is wrong.

The ask that makes it work puts the design system first:

Say to Claude:

First build a slide master: the right slide types, heading styles,
fonts, logo placement, and background treatments from my brand guide.
Then redesign every slide of this deck to match it. Do not change a
single piece of content — same words, same charts, same data, same
order.

How Claude actually does this is less exotic than it sounds. It writes small Python scripts (using a library called python-pptx) that restyle the slides programmatically, renders a preview image of each slide, looks at those previews, and fixes what is off. It is the screenshot verification loop from Chapter 5, applied to slides instead of web pages.

Three practical notes from doing this for real. First, slide-master-first is the discipline: get the design system approved on one slide before it touches thirty.

Second, give Claude two inputs — a reference deck whose design you admire and your brand guide. Third, a big deck can take a couple of sessions; when context runs low, ask for a session handoff note (Chapter 12 covers these) and resume fresh.

Setup is minimal, and Claude installs everything when asked. The two requirements are the python-pptx library and LibreOffice (for rendering the slide previews):

In your terminal:

pip install python-pptx

Say “install whatever you need for PowerPoint work” and Claude handles both. The payoff is the point: one instruction redesigns an entire deck without risking a single number.