Prepare a source inventory
Give files meaningful names and record their titles, authors or organizations, dates and versions. Remove accidental duplicates from your working set, but keep distinct editions when comparing changes. Note whether a PDF contains selectable text or scanned images. Check extracted text for broken columns, missing tables and incorrect characters.
Ask a question that crosses documents
Frame a comparison around definitions, findings, assumptions or recommendations. Ask the agent to distinguish a statement appearing in one source from a conclusion supported across several sources. A synthesis should identify disagreements and gaps rather than combine every paragraph into a uniform narrative.
- List the files included in the task.
- Define the fields you want compared.
- Separate document evidence from web evidence.
- Require uncertainty when a source cannot be read fully.
Use files as sources in Humanizily
Attach the supported documents in the Research Agent and explain their role in the brief. Review the proposed outline before the run. If a file cannot be extracted reliably, obtain a readable version or use an appropriate text-extraction step, then inspect the result. Do not assume that every page or table was captured because upload succeeded.
Verify the cross-document synthesis
Open the cited files and locate the passages behind important statements. Where available, check page numbers and section headings; PDF viewer numbering and printed page labels can differ. Record which version supplied the evidence. Before sharing or exporting, inspect the source list and confirm that confidential or restricted files are appropriate for the intended workflow.
Worked example
Starting point
Summarize these PDFs.
Suggested wording or approach
Compare how the three attached reports define customer retention. Record each definition, measurement period and stated limitation. Do not treat different definitions as directly comparable, and flag unreadable tables.
Example file-research brief. The agent’s access to text depends on the actual files and extraction results.
Common questions
Does a successful upload mean every page was read?
No. Check extraction quality and any missing or unreadable content.
Should I include duplicate reports?
Use one copy of identical files; keep different editions when the version difference matters.
Can I mix uploaded files and web sources?
Yes, but identify their roles and check the provenance of each cited claim.
Start with a focused research brief
Define your question, review the outline and check the evidence before sharing the report.
Open the Research AgentHow the agent works