Use this workflow when the answer should come from a file rather than from general model knowledge. File availability and limits can vary by plan and workspace.
1. Attach only the source you actually need
Open a chat, attach the relevant PDF, DOCX, CSV, XLSX or other supported file, and avoid mixing unrelated files before the task is clear.
2. Define the job before asking for a summary
Say whether you want extraction, comparison, transformation, calculation or a decision-support brief. Name the audience and output format.
3. Set an evidence rule
Tell ChatGPT to rely on the uploaded file, cite page/section/row references when possible, and explicitly say when the file does not contain an answer.
4. Ask for a first-pass inventory
For long files, first request the structure, key fields, missing information and obvious inconsistencies. Then ask the narrower question.
5. Verify the important claims
Open the original file and spot-check the cited passages, rows and calculations before using the result in a high-stakes decision.
Use only the attached file as the source of truth. First list the sections relevant to my question. Then answer in a table with: claim, evidence location, confidence, and what is missing. If the file does not support a claim, say “not found in the file.”
Asking “summarize this” without a purpose; uploading many versions without naming the source of truth; accepting a polished answer without checking the underlying pages or rows.
Finish check
You should be able to point from every important output claim back to a specific part of the uploaded source.