How to Analyse Your Own Documents Without Leaving the Page: Gemini Notebook's Code Execution Feature

If you regularly need to pull insights from reports, spreadsheets, or project summaries, you know the drill: export here, paste there, reformat, lose something, start again. This is for managers and ops professionals who are tired of that cycle. By the end of this article you will have a specific prompt and a working method to analyse your own uploaded documents inside Gemini Notebook, with no tab-switching and no reformatting.

Why This Matters

The standard workaround for document analysis has always involved moving your data somewhere else. You copy figures into ChatGPT and the formatting breaks. You export to a separate analytics tool and spend twenty minutes getting the columns right. Or you do it manually, which works but costs time you do not have. The deeper problem is that most AI tools answer questions from general knowledge, not from the specific documents in front of you. That gap is where hallucinated figures and invented patterns come from. When the AI is not actually reading your file, it is guessing from context. That is a reliability problem, not just an accuracy problem.

The Workflow

Google renamed NotebookLM to Gemini Notebook on 16 July and added the ability to write and run code directly inside a notebook, against your own uploaded files. That changes the practical equation.

Here is the workflow to use today. Upload your source documents directly into Gemini Notebook. These can be PDFs, reports, or project summaries, the kind of material that comes out of quarterly reviews. Then run this prompt against them:

PROMPT:

Based only on the documents I have uploaded, identify which activity areas appear in more than one report and show me a count of how often each one is mentioned.

Gemini Notebook will write and execute the code to answer the question, return a structured breakdown, and show you the code it used. That last part matters more than it sounds. You are not just getting an answer. You are getting an auditable record of exactly how the answer was produced. The output is grounded in your documents, not generated from the model's general knowledge.

In testing this against three quarterly PDF summaries, the cross-referencing task that would have taken thirty to forty minutes manually came back in under five. No invented figures. No patterns that were not actually there. It counted what was in the documents and showed its working.

One important limitation: when the prompt shifts from specific and document-referenced to interpretive, for example asking which area looks most at risk, the code execution drops out and the response reverts to a standard language model answer. Still useful, but no longer grounded in the same rigorous way. Keep your prompts specific and tied to what is in the documents for the most reliable results.

This is the kind of practical AI workflow covered every week in the AI at Work newsletter on LinkedIn. Subscribe to get one tested system per week, directly in your feed.

What to Expect

Used consistently, this workflow removes the biggest friction point in document-based analysis: the gap between where the information lives and where you can ask questions about it. You will stop losing formatting in transit, stop second-guessing whether the AI is drawing on your actual data, and start getting structured outputs with a transparent audit trail. For anyone who reviews reports regularly, the time saving compounds fast.

Start Here

Open Gemini Notebook, upload one document you already have open today, and run this prompt against it:

PROMPT: 

List every specific action, decision, or number mentioned in this document, in the order they appear.

It is a small test. But it will tell you immediately whether the tool's document-grounding is useful for the kind of material you actually work with. That is all you need to know to decide whether this belongs in your regular workflow.

The tool I use for this: Gemini Notebook.