Full Breakdown
A Structured Prompt That Turns ChatGPT Into a Document-Comparison Assistant
By Drooid · · How we work
A Structured Prompt for Document Comparison
The Tom’s Guide article describes a prompt that improves ChatGPT’s usefulness when evaluating lengthy contracts, insurance policies, or other multi-page PDFs. Instead of asking the model to “compare” two documents in a single step, the author recommends a four-stage workflow: extract key data from each file independently, cite the source page for every item, compare the extracted tables, and flag any missing information. The prompt explicitly tells the model to write “Not found” for any detail that is not stated in the source, preventing the model from fabricating answers.
Step-by-Step Workflow
1. Upload both PDFs.
3. After extraction, request a side-by-side table that lists the differences.
4. Finally, ask the model to identify five differences or unclear terms that merit further investigation.
Benefits Over Traditional Summaries
By separating extraction from comparison, the approach forces the model to treat each document uniformly, reducing the risk that it will prioritize information subjectively. Page citations let the user verify figures—e.g., a $500 deductible versus a $1,000 deductible—without rereading the entire text. The author notes that this method eliminates the need to scan 60 pages line-by-line and makes discrepancies easier to spot.
Practical Applications
The technique is useful for any decision that hinges on detailed contractual terms. The article cites examples such as switching wireless providers, evaluating insurance policies, comparing apartment leases, analyzing streaming-service plans, reviewing AI-subscription offers, and scrutinizing travel-package inclusions. Users can adapt the category list to fit the specific documents they are reviewing.
Cautions and Best Practices
The author warns that ChatGPT can still misinterpret clauses, overlook table entries, or present confident-sounding but inaccurate statements. Including page numbers provides a quick verification path, and preferring “Not found” over guessed information helps maintain trustworthiness. The method is intended to surface the handful of pages where human judgment is truly required, not to replace that judgment entirely.