Check the PDF before you upload it
ChatGPT can use an attached PDF as source material for a summary. First confirm that you are allowed to submit this particular document. Follow your employer's policy for contracts, customer records, financial reports, health information, and restricted material. Remove credentials or unrelated personal details where you have authority to do so. If the document cannot be shared with ChatGPT, use an approved tool or work from an authorized, redacted excerpt.
- Open the PDF yourself. Record its title, version or date, and number of pages so you can recognize the right file in the answer.
- Try selecting a sentence on a middle page. If you cannot select text, the document may be a scan; the scan workflow below requires extra checks.
- Look at a table, chart, footnote, or appendix if one could affect your purpose. A summary of body paragraphs alone may miss it.
- Write down the question the summary should answer: an overview for yourself, a decision for a manager, research notes, or instructions to carry out.
Summarize your first PDF in ChatGPT
- Open a new ChatGPT chat and attach the PDF using the file control in the message composer. Wait until the attachment appears in the draft message. If it does not, resolve the upload before asking for a summary.
- Send a prompt that names the audience, purpose, length, and source boundary. Start with the short prompt below if you only need an overview.
- Check the reply for the document title, the requested topics, and page or section locations. Ask which pages or visuals it could not read if the answer seems to skip part of the file.
- Open the cited pages for every number, obligation, exception, or recommendation that matters. Check the adjacent text as well as the cited sentence.
- If the summary will be shared or used for a decision, run the omission check later in this guide and revise the summary before sending it.
Summarize the attached PDF for [reader] who needs to [purpose].
Give me [length or number of bullets]. Use only this PDF.
Put a PDF page number or section heading beside each main point.
Include the main conclusion and any exception that changes it.
If you cannot read part of the file, name the missing part.You do not need a long prompt for a quick orientation. For a summary that someone else will rely on, the output and checking rules below make the result easier to audit. OpenAI's prompting guidance recommends specifying the goal, relevant source, desired output, and boundaries when those details change the result.
Choose the summary that fits your task
| Your task | Ask ChatGPT to return | What to check in the PDF |
|---|---|---|
| Decide whether to read the whole file | A brief account of its purpose, sections, main claim, and conclusion | Whether the summary covers the introduction, final section, and any appendix that qualifies the conclusion |
| Brief someone on a decision | The recommendation, supporting numbers, alternatives, conditions, and unresolved questions | Whether a condition or excluded cost changes the recommendation |
| Understand a research paper | The research question, method, sample, results, and stated limitations | Whether the result is described with the same population and degree of certainty as the paper |
| Carry out a procedure | Required steps in order, dependencies, deadlines, and exceptions | Whether words such as “must,” “may,” and “unless” keep their original meaning |
These are different outputs, even when the same file is attached. A five-sentence executive summary is useful for a manager but too compressed for a procedure with several exceptions. If you need a particular section, name its heading or page range and ask for a focused summary of that section.
Use a source-checked prompt for consequential summaries
The following prompt asks for evidence beside each finding and makes missing information visible. Replace the bracketed text. Treat the page references as leads to check, not proof that ChatGPT read every page correctly.
Use the attached PDF as the only source for this task.
Purpose: Help [reader] [make a decision / understand the research / follow a procedure].
Scope: Summarize [whole PDF or named sections].
Length: [for example, 200 words plus a findings table].
Return:
1. the document title, date or version, and the pages you could read;
2. a summary that answers the stated purpose;
3. a table of important findings with PDF page or section references;
4. exceptions, limitations, and conditions that narrow those findings;
5. questions the PDF does not answer.
Use no outside facts. Separate what the author says from your interpretation.
Do not invent a quotation or a page number. If a passage, chart, or page
cannot be read confidently, say which one and leave that claim unverified.Use “PDF page” to mean the page shown by the viewer, unless you explicitly need the page number printed in the document. The two can differ when a report has a cover, unnumbered front matter, or Roman numerals. In your final summary, use one numbering system consistently, for example “PDF page 7 (printed page 5).” If ChatGPT cannot supply an exact page, ask for a section, figure, or table location and find the passage yourself.
Practice on a three-page decision brief
Download the synthetic practice PDF. It describes an eight-week support pilot with 120 agents in three teams. The requested decision is whether to expand to all 160 agents. Because its numbers and qualifications are known in advance, you can check a ChatGPT summary against the actual pages without uploading a real workplace document.
A useful summary of this file should say that average handle time moved from 12.4 to 10.8 minutes, while customer satisfaction moved from 4.28 to 4.31 and was treated as essentially unchanged (PDF page 1). West improved by only 0.7 minute and its quality score fell from 92 to 90 (page 2). The report recommends a controlled 30-day extension with five required gates, not an immediate full rollout; procurement for the remaining 40 agents stays on hold (page 3). Those are source facts, not a generated answer you should copy without checking.
To repeat the retained run, use the prompt below with the practice PDF. The more flexible detailed prompt earlier in this guide was added during this revision and was not part of the August test.
Read the attached PDF as the only source for this task.
Purpose: Help me [make a decision / prepare a briefing / understand the research].
Audience: [Describe the reader and their existing knowledge.]
Return:
1. a five-sentence executive summary;
2. the document's purpose and main conclusion;
3. the most important findings in a table with page references;
4. limitations, exceptions, or conditions that could change the interpretation;
5. unanswered questions that the PDF does not resolve.
Accuracy rules:
- Do not add outside facts.
- Distinguish the author's statements from your interpretation.
- If a page cannot be read, say so.
- Do not create quotations. Use quotation marks only for text you can locate exactly.Before that run, the editorial team wrote checks for the decision, the two headline measures, West's result, the five extension gates, the causal limitation, and a page 3 measurement exception. The reported handle time excludes chats under 60 seconds and over 90 minutes, and the PDF gives no sensitivity analysis for that exclusion.
| Prewritten check | Source fact | Observed first pass |
|---|---|---|
| Decision | Controlled 30-day extension; no full rollout | Kept procurement for 40 agents on hold |
| Efficiency | 12.4 to 10.8 minutes, down 12.9% | Matched with a page 1 reference |
| Customer outcome | 4.28 to 4.31, treated as unchanged | Did not describe this as a meaningful gain |
| Weakest team | West: 0.7-minute handle-time improvement, two-point QA decline, 63% training completion | Matched with a page 2 reference |
| Measurement exception | Excluded chats under 60 seconds and over 90 minutes | Found on page 3 and flagged the missing sensitivity analysis |
The first response also kept the report's uncertainty: the pilot was not randomized, had no concurrent control group, and overlapped with a redesigned knowledge base and a new coach for Central. Supervisor review time was not measured separately, and the cost model excluded several expenses. The lower handle time therefore supported more testing, not a claim that the assistant caused the improvement.
Ask what the first summary left out
After the first response, run a separate omission check in the same chat. Name the draft you want reviewed and keep the PDF available. This second prompt can reveal an underemphasized fact, but its answer still needs a source check.
Compare your draft summary with the attached PDF. List material facts
that the draft omits or underemphasizes, especially negative results,
exceptions, eligibility rules, dates, dependencies, and uncertainty.
For each item, give its PDF page or section and explain how it could
change the stated purpose. Do not repeat points already covered well.In the retained run, ChatGPT listed 15 omitted or underemphasized details. Among them were North's one-point quality decline despite its strong handle-time result, differently sized team samples, policy wording supervisors usually caught, the case types where suggested responses were used most, and the fact that all five extension gates were mandatory. Open pages 2 and 3 to decide which of these belong in the final brief. The first draft's recommendation did not change, but omitting those facts would make a rollout discussion less well informed.
For example, a corrected decision sentence could read: “The brief recommends a controlled 30-day extension before any full rollout. West's QA score fell two points and North's fell one; the extension requires five checks, including a comparison group and separate measurement of supervisor review time (PDF pages 2 and 3).” This is an editorial example assembled from the practice PDF, not a quote from the ChatGPT response.
Verify claims in the original pages
Start with claims that could change what a reader does. Copy the summary sentence into a ledger, note the cited page, find the supporting passage, and record the result. Check nearby paragraphs, table headers, units, footnotes, and defined terms. Recalculate a percentage if the summary uses it: the practice PDF's 1.6-minute decrease divided by the 12.4-minute baseline is about 12.9%.
| Result | When to use it | Next action |
|---|---|---|
| Supported | The page supports the same number and scope | Keep the claim and page reference |
| Partly supported | A date, group, qualifier, or unit differs | Narrow the wording to match the source |
| Contradicted | The source says something materially different | Correct the claim before sharing |
| Not found | No adequate passage can be located | Remove the claim or mark it unverified |
| Material omission | An absent fact changes how a claim should be read | Add the fact with its location |
Download the verification ledger. Its first five rows show completed checks against the practice PDF, including a material omission; five empty rows follow for your own file. The example rows are editorial checks of the PDF. They do not claim that each sample sentence appeared verbatim in the retained model response.
When the PDF is long, check coverage before synthesis
A single smooth answer can hide skipped chapters. Ask for a section map before summarizing a long report. Compare the returned headings and page ranges with the PDF's contents page or your own quick scan. If an appendix, methods section, or table is missing from the map, repair the map first.
- Record the file's page count and the sections relevant to your purpose. Include appendices or exhibits that set definitions or exceptions.
- Ask ChatGPT for a map of those sections and their PDF page ranges. Treat an absent or uncertain range as a coverage gap.
- Summarize one relevant section at a time. For each, request the key claim, numbers, limitations, and exact location. Save the section notes.
- Ask for a final synthesis from the section notes and the attached PDF. Tell ChatGPT to flag disagreements, repeated claims, and dependencies between sections.
- Compare the final answer with the section map. Every conclusion that matters should have a traceable source location; a section you never checked should remain outside the claimed coverage.
Before summarizing, map the PDF's sections, page ranges, tables,
and appendices relevant to [purpose]. Mark any section or page you
cannot locate or read. Do not infer its contents. Then summarize only
the sections I select, one section at a time.Section notes are a working aid, not a replacement for the PDF. A final synthesis can compound an error in those notes. Return to the original pages for any recommendation, total, quotation, or exception you intend to pass along. If you only review selected chapters, label the result “summary of selected sections” and name the selection.
When pages are scanned or visually complex
| What you see | Practical next step | Do not claim completion until |
|---|---|---|
| Text cannot be selected or copied from a page | Obtain an accessible source or run approved OCR. Compare a paragraph near the start, middle, and end of the affected material with the original image. | Names, numbers, and headings in those samples are readable and match the source. |
| Two columns, sidebars, or footnotes read out of order | Ask for the specific passage's heading and surrounding text; inspect the page image yourself. Split the affected pages or provide clean extracted text if needed. | The summary keeps each qualification with the claim it modifies. |
| A table carries the main evidence | Ask for the exact row and column labels, units, and footnotes used for each reported value. Check totals and any excluded rows in the original table. | The numbers and their denominators agree with the table. |
| A chart carries the main evidence | Check the title, axes, units, period, legend, and source note. If an exact value is not printed or otherwise available, describe only the visible trend. | The wording does not turn an approximate visual reading into an exact figure. |
These checks matter when a PDF mixes selectable text with scanned exhibits: a readable text layer does not prove that a chart or image-based appendix was understood. If extraction remains unreliable, give the reader a short verified summary and a named list of pages needing manual review.
Fix common summary failures
| Failure you can observe | What to do next |
|---|---|
| The attachment is absent from the sent message or the answer names the wrong document | Attach the right file again, confirm its title or date, and start a fresh request tied to that file. |
| The answer is vague or gives equal space to irrelevant background | State the reader's exact purpose and request a bounded output, such as three decision points and the conditions for each. |
| A page reference points to the wrong passage | Check whether it uses viewer or printed numbering. If neither matches, ask for a section or nearby wording and locate the claim yourself. |
| An appendix, exception, or negative result is missing | Run the omission check, read that part of the PDF, and revise the affected sentence rather than merely adding a generic caveat. |
| The file is password protected, incomplete, or unreadable | Obtain an authorized accessible copy or an approved excerpt. State which pages the summary could not cover. |
Hand over a summary someone else can check
Put the document title, version or date, and summary purpose at the top. Lead with the answer to that purpose. Keep page or section locations beside the relevant claims, followed by material limitations and unresolved questions. For the practice brief, a handoff that says only “handle time improved 12.9%” would leave out the unchanged customer-satisfaction result, team variation, and required extension gates.
Revise the checked summary for [reader]. Put the answer to [purpose]
first. Keep each verified number and qualification with its PDF page
or section reference. Add a short "Not established by this PDF"
section for unresolved questions. Remove claims marked contradicted or
not found in my verification ledger. Do not imply that unreadable or
unchecked pages were reviewed.Before sharing, open the final summary and verify every consequential claim once more against the PDF. For the practice brief, that means the 30-day extension and its five mandatory gates, the two headline measures, West and North's quality results, the lack of a causal comparison, and the excluded handle-time records. If the source cannot support one of those details, leave that detail visibly unresolved.
