Working with files and data in Claude: strong at reading, careful with numbers
You can hand Claude a pile of files — a CSV export, a quarterly PDF, a spreadsheet, a screenshot of a table — and ask questions in plain English. That is…
Claude guide · as of July 10, 2026 · 4 minutes read · details change — confirm current specs on claude.ai
You can hand Claude a pile of files — a CSV export, a quarterly PDF, a spreadsheet, a screenshot of a table — and ask questions in plain English. That is genuinely useful. But handling files and analyzing data are two different jobs, and Claude is much better at one than the other. This guide draws the line clearly so you don't ship a wrong number.
What you can upload
- Spreadsheets and data files:
.csv,.xlsx,.tsv - Documents: PDF, Word, plain text, Markdown
- Images: PNG, JPG, screenshots — Claude reads charts and tables inside them (vision)
Some models take very large context (up to ~1M tokens on some models), so long documents often fit in one pass. Exact upload sizes, row limits, and file counts vary by surface and plan — check the current limits at claude.ai and anthropic.com (as of July 2026 — confirm there).
Two different jobs
Reading and understanding — a real strength
Claude is genuinely good at making sense of a document: summarizing a 60-page filing, pulling the one clause you need, explaining what a confusing spreadsheet's columns mean, turning messy text into a clean structure, or comparing two versions of a contract. If the task is "read this and tell me what it says," you are on solid ground.
Computing exact numbers — handle with care
A language model predicts text. If you paste a column and ask "what's the total?", it may answer by pattern-guessing rather than adding — and it can be confidently wrong. The fix is simple: make it write and run code instead of eyeballing.
- On claude.ai, the analysis tool can write and run code to work through a CSV.
- Claude Code runs Python (pandas, and so on) directly on your machine.
- The Anthropic API offers code execution for developers.
Feature names and availability change — check the current set. When real code runs, the number is computed, not predicted.
| What you want | Reading and reasoning | Writing and running code |
|---|---|---|
| Summarize a report | Strong | Overkill |
| Explain what the columns mean | Strong | Not needed |
| Exact sum, average, or count | Risky — can be wrong | Reliable — actually computed |
| Filter or pivot 50,000 rows | Truncates and guesses | Reliable |
| Spot trends worth investigating | Good starting point | Confirm with code |
Failure modes to watch
- Invented totals. If it answers a math question instantly without running code, treat the number as a draft, not a result.
- Silent truncation. Very large files can be cut off; rows you care about may never be read.
- OCR mush. Scanned PDFs, merged cells, and multi-row headers get misread.
- Hallucinated cells. It can fill in a plausible-looking value that isn't actually in your file.
- Wrong assumptions. It may guess units, date formats, or which column is the key.
One habit fixes most of this: ask it to show the code, then spot-check one row by hand.
A 30-second pre-flight
- Ask for the calculation to run as code, not be estimated
- Confirm how many rows it actually read
- Check that column names and units match what you meant
- Re-compute one value yourself as a sanity check
- Keep the original file — Claude's output is a copy, not your source of truth
Charts and visuals
Claude has no native image generator. It doesn't paint a chart the way an image model would; it writes charting code that renders live in an Artifact. That is great for a quick, editable visual — but the picture is only as right as the numbers behind it, so verify those first.
When this is the wrong tool
- Live or recurring reporting. For a dashboard that refreshes daily, use a spreadsheet with formulas or a BI tool. Claude is a one-off analyst, not a pipeline.
- Audited or regulated numbers. When you need a reproducible, inspectable trail, keep the math in visible spreadsheet formulas.
- Millions of rows. Beyond what fits in context, use a real database and hand Claude a representative sample or the query results.
- Data you cannot send to the cloud. Claude is cloud-only. Uploading sends the file to Anthropic's servers. Turning off training-data use does not make the request local — the data still leaves your machine. For strictly private data, a local model may be the safer call.
- A single lookup. If you just need one cell or a quick sort, a spreadsheet is faster than a conversation.
The short version
Use Claude to read, explain, extract, and explore. When exact numbers matter, make it run code and then check its work. And when the job is a recurring pipeline, audited math, or data that can't leave your machine, reach for a different tool.
Related
- Artifacts — how live charts and rendered docs work
- Local vs cloud — when to keep data on your own machine
- Privacy — what actually happens to files you upload
- Limits — why Claude can be confidently wrong
