Documents & Data

Drop the document into the chat — and ask

Drag a PDF, a spreadsheet, a screenshot, or a photo of a document into the chat — Claude will read it and answer from the contents. Remember the lesson on hallucinations: working from an attached document is the model's most reliable mode. The text is in front of it, there's nothing to reconstruct from memory.

Standard scenarios:

• A contract: "Summarize the key terms. Separately: payments, deadlines, penalties, termination conditions. Flag the clauses that look non-standard or risky for me as the contractor."
• A 60-page report: "A one-page summary: main conclusions, numbers, risks. Then answer questions from the text."
• A device manual: "How do I set up X? Answer only from this PDF, with a section reference."
• An invoice or receipt as a photo: "Pull the line items and amounts into a table."

Demand grounding in the source: "which section of the document are you saying this from?" — it disciplines the answers and simplifies checking.

Screenshots and images

Claude reads images on par with text: a screenshot of an error, a photo of a whiteboard after a meeting, a diagram from a presentation, a program's interface.

• A screenshot of an error + "what's wrong and how do I fix it?" — faster than retyping the error text.
• A photo of handwritten notes → "transcribe into a structured task list."
• A diagram or chart → "explain what's shown here and what conclusions follow."

Tables and data

Here's the main rule of the lesson. For calculations over data, ask for analysis with code execution, not an answer "by eye":

• "Analyze the table: revenue by month, top 5 clients, anomalies" — Claude will write and run code, returning exact numbers and charts.
• Asking it to "count in its head" over a thousand rows means getting plausible but approximate numbers (the lesson on hallucinations and arithmetic).

What works well with tables:

• data cleaning: "find duplicates, empty fields, typos in names";
• turning data into a conclusion: "state the three main insights from this table for a presentation to management";
• format conversion: "assemble these three CSVs into one summary table by month."

Limitations

• Very long documents run into the context window. The fix: work chapter by chapter or ask for a summary of each part first.
• Poor-quality scans are read with errors in the numbers — cross-check critical figures against the original.
• Complex PDF layouts (multi-column, tables within tables) sometimes break the reading order. If the answers are strange, ask how the model understood the document's structure.

Key takeaways

• An attached document is the most reliable mode of work: the model answers from the text, not from memory.
• Images are equal to text: screenshots, photos of documents, diagrams.
• Calculations over data — only via analysis with code execution, not "in the head."
• Demand grounding of answers in the document's sections; cross-check critical numbers from scans.