Using ChatGPT to analyse documents without losing control

ChatGPT can read a PDF, a Word document or a set of notes and help you turn them into a summary, a task list or a plan of work.

But uploading a file and asking for “a summary” does not always produce something useful. ChatGPT can organise the text, but it does not automatically know which part is a decision, which part is an idea, and which part still needs developing.

I found this out working on a real digital project.

The problem I hit in my own project

The project had a master plan with priorities, ideas, decisions and documents to guide the building of a website.

The intention was to use ChatGPT to organise that material and then hand the work to an automation. The problem appeared when I reviewed the first published piece.

Some of the planning information had been turned straight into editorial content. The result looked like an article, but it was not ready for a real reader:

• the topic was barely developed;
• the content was far too generic;
• the image did not match the article;
• placeholder text had made it through;
• concrete examples were missing;
• no full editorial review had been done.

The text had headings and paragraphs, but that did not mean it solved anything for anyone.

The mistake was not only in the writing. It had started earlier: information that was not yet resolved had been sent to the automation.

Since then I have used one rule:

First you understand the document and resolve the content. Then you automate its preparation.

A priority is not an article

This was the most important distinction in the project.

Item	What it means
Priority	Something that matters to the project
Idea	A possibility that still needs developing
Task	A specific action someone has to carry out
Article	Finished content written for a reader

For example:

Create content about useful accessories

is a priority or an idea. It is not yet an article.

To turn it into something useful you would have to decide:

• what problem it will solve;
• who has that problem;
• what options exist;
• what mistakes the reader should avoid;
• what information needs verifying;
• which image best represents the topic.

Skip that step and ChatGPT will fill the gaps with sentences that are correct but far too general.

The method I use now

After that first attempt, I split the process into five steps.

1. Sort

First you have to identify what the document contains:

• decisions;
• ideas;
• tasks;
• articles;
• requirements;
• visual assets;
• information still outstanding.

You can ask:

Read this document without drafting any article yet. Classify the information as strategy, decision, idea, task, editorial content, technical requirement or outstanding information. Do not automatically turn ideas into articles.

2. Define

Then you pick one single piece.

Choose one idea that could become an article. Explain what problem it will solve, who will read it, what question it answers, and what information is missing before it can be written.

This is what stops a whole document turning into a collection of generic texts.

3. Draft

Once the topic is defined, you prepare the draft:

Write a complete article for someone who wants to solve this problem. Include concrete examples, use clear language, and do not invent experiences, results or sources. If important information is missing, say so.

4. Review

The text is reviewed as if you were the reader:

• Does it answer a specific question?
• Does it have enough depth?
• Do the examples help?
• Can the claims be verified?
• Does the image match the topic?
• Is there any placeholder text?
• Does the reader know what to do next?

5. Approve

Only when the article is resolved does it go to the automation.

The automation can prepare files, previews and technical checks. It must not decide that an incomplete note is a finished article.

Before and after

An instruction that is too open would be:

Read these priorities and create the content.

That sentence does not explain what should stay as planning and what should become a publication.

A safer instruction would be:

Read these documents and separate the strategy, the decisions, the ideas, the tasks and the possible articles. Do not draft anything yet. Say what information is missing to turn each idea into useful content.

Then:

Work only with the chosen idea. Define the reader, the problem, the main question, the structure and the examples needed. If information is missing, do not invent it.

And finally:

Write the article following that definition. Do not include placeholder text, provisional sentences, or claims that cannot be verified.

The difference is that you understand the material first and decide what to produce second.

How to ask for a useful summary

“Summarise this document” can fall short because it does not say what has to be kept.

A more precise request would be:

Summarise this document in plain language. Explain its purpose, the three main ideas, the decisions taken and the outstanding tasks. Keep the doubts and the unconfirmed data separate. Do not invent information.

If the summary is meant to support a decision:

Summarise this report for someone who has to decide whether to continue with the project. Explain what is being proposed, what advantages appear, what risks are mentioned, and what information is missing.

You can also choose the format:

• five points;
• a table;
• a 300-word summary;
• an explanation for a non-expert;
• conclusions kept separate from the facts.

How to pull out tasks and dates

To turn a document into a work list:

Extract only the outstanding tasks. Give the description, the owner, the deadline, the dependencies and the status. If something is not stated, write “not specified”.

Then:

Order the tasks by priority and dependency. Say which have to be done first and which are blocked by a pending decision.

This helps you tell an outstanding task apart from one that cannot start yet.

For example, reviewing an image may depend on the article being approved first. If that relationship is not identified, the work can be done in the wrong order.

How to separate facts, opinions and suggestions

Documents usually mix confirmed information with provisional comments.

You can ask:

Separate the confirmed facts, the opinions, the suggestions, the hypotheses and the information that needs confirming. Do not turn a suggestion into a decision.

Also:

Flag any sentences that are ambiguous, incomplete or contradictory. Say what question would have to be answered before we can continue.

That separation mattered in the project because it showed which information could be used and which needed a decision first.

Working with PDF, Word, Excel and images

PDF

Useful for reports, manuals and contracts. If it contains selectable text, it will usually be easier to analyse. If it is a scan, names, figures and tables have to be checked by hand.

Word

Lets you work with headings, lists and paragraphs. You can ask it to keep the structure or to reorganise the document.

Excel

Works best when each column has a clear name and each row is one record. You can ask it to analyse dates, amounts, statuses or categories.

OpenAI recommends using descriptive headers and well-organised data to improve the analysis.

Images

ChatGPT can analyse screenshots and visual documents, but it is worth checking small text, figures and complex tables by hand.

How to spot contradictions

You can ask:

Look for internal contradictions in the document. Say which parts conflict and explain why. Do not decide which one is right.

If one section says a delivery is on 10 September and another says the 15th, ChatGPT should flag the conflict. It should not pick a date on its own.

How to prepare an email

Once the information has been reviewed:

Prepare a short email confirming that we have received the proposal. Say that we will review the budget and that we need to clarify the delivery date. Do not accept any terms, do not promise dates, and do not invent information.

Before sending it, check whether it:

• accepts something you did not mean to accept;
• promises a date;
• mentions a wrong figure;
• assigns a task to the wrong person;
• uses a different tone from the one you intended.

How to work with private information

Before uploading a document, check whether it contains data you do not need to share.

It might include:

• full names;
• addresses;
• phone numbers;
• passwords;
• banking details;
• contracts;
• medical information;
• client data.

If that data is not needed, you can replace it with labels such as [CLIENT], [COMPANY] or [DATE].

The rule is simple: use only the information needed to solve the task.

Common mistakes

That first attempt left five lessons:

1. Priorities are not articles.
2. A template is not finished content.
3. A text with structure can still be shallow.
4. An attractive image may have nothing to do with the topic.
5. An automation must not fill in information that is missing.

The fix was not adding more paragraphs at random. It was separating the stages and putting an approval step before publishing.

What the automation should do

The automation can:

• prepare files;
• generate previews;
• run checks;
• create a working branch;
• prepare a pull request;
• report errors;
• carry on with independent tasks.

But it must not:

• decide which priority deserves an article;
• invent an experience;
• pick an unrelated image;
• publish a template;
• fill gaps with generic text;
• treat unreviewed content as finished.

If information is missing, the task should stay open and the system should explain what is missing.

How to know whether an article is finished

An article is ready when it:

• answers a specific question;
• solves a real need;
• has enough depth;
• includes useful examples;
• contains no invented claims;
• uses a related image;
• tells the reader what they can do;
• has been reviewed by a person;
• contains no placeholder text;
• has a title and description that match.

The final review

You can ask ChatGPT:

Audit this article as an editor. Check whether it answers a specific need, whether the examples are useful, whether there are unsupported claims, whether the structure helps the reader, and whether the image matches the topic. Say what should be corrected before publishing.

Then:

Rewrite only the parts that need improving. Keep the correct facts and do not invent information to fill the gaps.

Reviewing is not only about checking the grammar. It also means checking the usefulness, the accuracy, the image and the intent of the content.

Frequently asked questions

Can ChatGPT turn a plan into articles?

It can help develop one specific idea, but it should not automatically turn a whole plan into articles.

Why can a well-written text still be bad?

Because it can have good grammar and structure and still not answer a problem or provide enough information.

Who decides whether an idea becomes an article?

That decision belongs to the editorial process. ChatGPT can propose options, but it should not decide automatically.

When does the automation come in?

Once the article has been defined, drafted, reviewed and approved.

What happens if information is missing?

The piece should stay open and what is missing should be clearly explained. It must not be filled in with invented information.

Can ChatGPT get a document wrong?

Yes. It can miss data, misread a sentence or confuse information. The important points have to be checked against the original document.

Checklist before publishing

Before sending an article to publication, check:

• Is the topic defined?
• Does the reader have a specific problem?
• Does the article answer that problem?
• Have facts and opinions been separated?
• Has unconfirmed data been flagged?
• Does the image match the content?
• Has all placeholder text been removed?
• Are the dates and figures correct?
• Are the examples useful?
• Has a person reviewed the text?
• Does the title really describe the article?
• Does the description summarise its content?
• Is the article approved?

Conclusion

This project showed that using ChatGPT with documents is not just about summarising information or generating text.

First you have to understand what the document contains. Then you have to separate priorities, ideas, tasks and editorial content. Only then can you write an article and send it to an automation.

The correct process is:

Document → sorting → specific topic → article → review → approval → automation → publication

Artificial intelligence can speed up the work, but it does not replace the editorial decision.

The most important lesson was this:

An automation can produce content that has been resolved. It cannot decide on its own what a document was supposed to mean.

Sources

• File Uploads FAQ — OpenAI
• Data analysis with ChatGPT — OpenAI
• Extracting Insights with ChatGPT Data Analysis — OpenAI

Keep reading

• Getting started with ChatGPT without learning prompts
• Automating what is worth automating

Review

Responsible for publication: Ibrahin. Date: 2026-09-02. Last checked: 2026-09-02. Status: published and approved.