Documents demo
The PDF Evidence Table Reveal
Upload a PDF and ask Claude to separate claims, supporting pages, uncertainty, and follow-up checks into a reusable evidence table.
"Extract text, analyze charts"
Setup: what the audience sees
The spectacle is not summarization. It is the moment when a 40-page document becomes a table of claims, page anchors, counterevidence, and next questions.
The setup matters because Claude demos fail when the prompt hides the goal, mixes private data into a public reveal, or asks for a finished answer without giving the model a way to expose assumptions.
- Pick a public PDF, internal policy copy, or synthetic report you are allowed to upload.
- If the source is high stakes, use Claude to organize evidence, then verify in the original document.
- Ask for page references and uncertainty labels instead of a fluent summary alone.
What this does not prove
It does not prove legal, medical, or financial truth. It proves that Claude can help structure document review and surface where humans need to check the source.
<task>Analyze the uploaded PDF as an evidence clerk.</task>
<output>
Return a table with these columns:
1. Claim in the document
2. Page or section reference
3. Evidence type: direct text, chart/table, image, or inference
4. Confidence: high, medium, low
5. What I should verify manually
</output>
<rules>
- Do not invent page references.
- If a claim relies on a chart or table, say so.
- After the table, list the three most important unresolved questions.
- Keep direct quotes short and only when they are essential.
</rules> The reveal: what should happen
- Claude should preserve the difference between text, charts, images, and inference.
- The answer should be useful even when the PDF is messy because uncertainty is part of the output format.
- The unresolved questions often become the best follow-up prompt.
Verification steps
A demo becomes useful when the audience can inspect it. Use the steps below to turn the reveal from theater into a repeatable workflow.
- Open the PDF and spot-check the highest-impact rows.
- Ask Claude to produce a second pass limited to low-confidence rows.
- For formal use, compare the table against a human-created sample.
Guardrails
- Do not upload documents you lack permission to process.
- Do not treat generated page references as final without spot checks.
- Use structured outputs or a schema when downstream software depends on the table shape.
Variations
- Turn a vendor security PDF into a procurement risk register.
- Turn a research paper into methods, limitations, and replication notes.
- Turn a policy document into obligations, owners, and evidence needed.
FAQ
Why ask for uncertainty labels?
They make the demo more honest and more useful. A citation-backed low-confidence row is better than a polished unsupported claim.
Can Claude cite images inside PDFs?
Anthropic citation docs describe text citations; image citations are not currently supported, so chart/image findings should be manually checked.
Related demos
Research
Research Brief
Ask Claude to research a current question, use multiple sources, and separate established facts from open questions.
Long Context
Long Context
Give Claude a large transcript or document bundle and ask for a structured synthesis with decision logs, contradictions, and missing evidence.
Primary Sources
These demos cite official Anthropic docs or help-center pages for capability claims. They do not claim identical output across accounts, plans, models, or dates.
Claude Platform Docs
PDF support
Claude can process PDFs and answer questions about text, pictures, charts, and tables in those documents.
Claude Platform Docs
Citations
Claude citations can reference text from provided documents, including text extracted from PDFs.
Claude Platform Docs
Increase output consistency
When exact JSON conformance matters, Anthropic recommends structured outputs rather than prompt-only formatting.
Claude Platform Docs
Prompting best practices
Anthropic guidance emphasizes explicit instructions, output formats, tool-use direction, and model-specific prompt updates.