The latest articles on DEV Community by Practical AI Workflows (@practicalaiwo). https://dev.to/practicalaiwo https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4022681%2F47c34928-dea0-4c57-81a7-4f55d33b2212.jpg https://dev.to/practicalaiwo en Practical AI Workflows Thu, 09 Jul 2026 14:00:57 +0000 https://dev.to/practicalaiwo/a-source-first-workflow-for-studying-pdfs-with-chatgpt-and-notebooklm-nbl https://dev.to/practicalaiwo/a-source-first-workflow-for-studying-pdfs-with-chatgpt-and-notebooklm-nbl <p>AI summaries are useful, but they can hide the source trail.</p> <p>When I use ChatGPT or NotebookLM with a long PDF, the fastest thing to ask is usually:</p> <blockquote> <p>summarize this</p> </blockquote> <p>The problem is that this compresses the document too early. A clean summary can look useful while making it harder to tell which claims came directly from the source and which parts were inferred or smoothed over.</p> <p>For studying, that matters. I do not just want a nicer explanation. I want notes I can check.</p> <h2> The source-first order </h2> <p>The workflow I have been testing is:</p> <ol> <li>map the source</li> <li>extract source-backed claims</li> <li>ask retrieval questions</li> <li>generate the study guide last</li> </ol> <h2> 1. Map the source </h2> <p>Before asking for a summary, ask the model to map the document:</p> <ul> <li>what sections exist?</li> <li>what does each section cover?</li> <li>what claims appear in each section?</li> <li>where should I check those claims?</li> </ul> <p>A source map is not the final answer. It is a way to keep the document visible before the model starts synthesizing.</p> <h2> 2. Extract source-backed claims </h2> <p>Next, separate the claims into buckets:</p> <ul> <li>directly supported by the PDF</li> <li>reasonable inference</li> <li>needs verification</li> <li>not supported</li> </ul> <p>This makes weak points easier to catch before they get turned into polished study notes.</p> <h2> 3. Ask retrieval questions </h2> <p>Instead of jumping straight to flashcards or a study guide, ask retrieval questions first:</p> <ul> <li>what does this section define?</li> <li>which paragraph supports this claim?</li> <li>what evidence is given?</li> <li>what is missing?</li> </ul> <p>If the questions are vague, the source map probably is not good enough yet.</p> <h2> 4. Generate the study guide last </h2> <p>Only after the source is mapped and checked should you generate:</p> <ul> <li>a summary</li> <li>flashcards</li> <li>practice questions</li> <li>exam-style explanations</li> </ul> <p>The summary is still useful. I just do not want it first.</p> <p>I wrote up the full workflow here:</p> <p><a href="https://practical-ai-workflows.pages.dev/" rel="noopener noreferrer">https://practical-ai-workflows.pages.dev/</a></p> ai chatgpt learning productivity