Choosing the Right AI Tool · Article Three
There are forty PDFs on the desktop, several articles you wrote previously, two recorded interviews and a spreadsheet. You want AI to help you understand the material, locate disagreements within it and eventually support a writing project that can continue to develop.
Uploading the files to ChatGPT Projects, Claude Projects or Gemini Notebook appears to begin the same process. All three products use similar words such as sources, instructions and chats. Their implied answer to “what is this collection?” is not quite the same. A Project resembles a room in which continuing work accumulates: sources, conversations and tasks develop together. A Notebook was originally designed more like a bounded reading desk. Sources chosen by the user sit at the centre and answers repeatedly lead back to them.
Before choosing, ask whether you want to perform many kinds of work around the material, or understand the material as far as possible on its own terms. That difference matters more than the maximum number of files in a plan.
The difficult work begins after a successful upload
Moving a file into an AI system does not establish a dependable knowledge base. A scanned PDF may be recognised poorly. Table structure can change during extraction. Importing a webpage may omit attachments. Audio transcription can introduce mistakes. Two files with the same name may represent different revisions. Less visibly, an answer may not use all the material that the user assumes has been considered.
An environment for personal sources therefore needs to handle at least four matters. The source boundary should be intelligible. The system needs a way to find relevant passages in a larger collection. Answers should lead back to particular sources. Instructions, discussions and provisional conclusions must continue over the life of the project.
The three products distribute those responsibilities differently.
| Working environment | Primary unit of organisation | Relationship between answers and sources | Natural direction of extension |
|---|---|---|---|
| ChatGPT Projects | Files, chats, project instructions and memory within a continuing project | Files become part of the project context; tools such as search and Canvas can also be used | Research, writing, planning, images and other mixed tasks |
| Claude Projects | A self-contained history, knowledge base and project instructions | Project knowledge can serve separate chats; retrieval expands the usable collection as it approaches context limits | Analysis, writing and repeated discussion against a common knowledge background |
| Gemini Notebook | A collection of sources selected and organised by the user | In the standalone Notebook, answers are grounded in selected sources and provide inline citations | Reading, learning, synthesis and multiple study or presentation forms derived from sources |
The table does not name a winner. It identifies a difference in emphasis: Projects preserve continuing work; a Notebook preserves a source world that can be questioned.
ChatGPT Projects: when sources are one part of longer work
ChatGPT Projects puts reference files, related conversations and project instructions into one workspace. Projects have built-in memory and can use tools available in ChatGPT, including search, Canvas and image generation; paid plans may provide deep research and agent capabilities. Existing chats can be moved into a project, after which they inherit the project’s instructions and file context.
This structure is useful when a source collection will pass through several stages. A project may begin with reading, continue into an outline and bilingual drafts, then require diagrams or a check against recent web information. The sources are not the only object. They become durable background for a broader production process. Different chats can handle evidence, structure and editing so that a single thread does not grow without limit.
Built-in memory needs careful interpretation. Remembering past discussions saves repetition, but it can also turn an early mistake into later background. Important interim conclusions should be saved in clear project files and labelled as verified findings or working hypotheses. “The AI should remember” is not an adequate record for a consequential decision.
Source discipline presents another issue. A Project can use web search and other tools. That makes it possible to update the collection, but an answer may also move beyond the uploaded files. When a strict boundary is required, project instructions and the immediate task should say that answers must rely only on project sources, identify gaps rather than fill them from general knowledge, and name the files and locations used. Sampling remains necessary because an instruction guides behaviour; it is not a technical guarantee of isolation.
ChatGPT Projects is consequently a strong candidate when sources will enter a varied production process. Its central value is continuity plus a broad toolset, not a promise that every answer remains permanently sealed inside the supplied material.
Claude Projects: when one knowledge background must serve many discussions
Claude Projects also provides self-contained workspaces with chat histories, project knowledge and project instructions. Documents, text and code placed in project knowledge can supply background across the project’s conversations. Anthropic says that when the project knowledge approaches the context limit, retrieval-augmented generation is enabled to expand the usable capacity.
That structure fits a relatively stable body of knowledge supporting repeated analysis. A course collection, interview archive and group of existing articles can be placed in one project. Separate conversations then address definitions, conflicts in the evidence and revisions to particular chapters. Project instructions preserve terminology, tone and evidentiary expectations without re-entering them in every chat.
“Available as project knowledge” does not mean that every answer reads every file to the same degree. A large collection generally requires retrieval, and retrieval determines which passages enter the current context. Material may be missed when the user’s vocabulary differs greatly from the source. Before beginning extensive drafting, ask for a short inventory of each file and try questions whose answers can only be found in one particular source.
Claude Projects should also be distinguished from Claude’s Research capability and Claude Code. A Project provides persistent knowledge background. Research performs multi-step investigation across the web and connected sources. Claude Code enters files and a terminal environment. A common brand does not give the three entry points identical source boundaries or permissions.
For work centred on analysis, discussion and long-form prose in which a common knowledge base must support many conversations, Claude Projects is a natural candidate. The final decision should still be tested with the actual collection, particularly retrieval, citation behaviour and management of long discussions. Claims about a model’s context length are not a substitute for that test.
Gemini Notebook: when the sources are the centre of work
In July 2026 Google renamed NotebookLM as Gemini Notebook. It remains a standalone research and learning product. Users can add PDFs, Google Docs, Slides and Sheets, Word files, webpages, YouTube videos, audio, images and other source types, then select which sources should be active for a particular question.
The most important difference concerns grounding. Google’s explanation of Notebook integration with the Gemini app makes a specific distinction. In standalone Gemini Notebook, responses are grounded exclusively in notebook sources. When the same notebook is used through the Gemini app, the response may also draw on web search and other tools. The new unified naming can conceal this difference. A user who needs source-constrained reading should confirm which entry point is active.
Inline citations help the reader move from an answer back to the underlying passage. That feedback matters when comparing reports, learning a course or organising historical material. Notebook can also derive briefings, timelines, mind maps, audio overviews, quizzes and other forms from the sources. These can provide different routes into the same collection, but an engaging derivative remains an interpretation of the evidence, not new evidence.
Import limitations matter. Google explains that a web URL generally contributes the text of that page rather than nested pages or embedded material. YouTube sources depend on captions. Footnotes and comments in Google files may not be imported. A synced Drive source also behaves differently from a local uploaded copy. Before relying on a collection, open several imported sources and check that the decisive material is present.
If the primary requirement is “answer from this collection and make it easy for me to return to the text”, standalone Gemini Notebook expresses the task most directly. If the work soon needs open-web research, complex file editing or external action, the result may need to move into the Gemini application or another production environment.
Is the collection an archive, a workbench or an evidence boundary?
The same files can play three roles. They may form an archive that supplies background when needed. They can be raw material on a workbench where research, writing, planning and production occur. In a stricter arrangement, they constitute an evidence boundary: the system is not allowed to import knowledge quietly from elsewhere.
ChatGPT Projects and Claude Projects naturally cover the first two roles. Standalone Gemini Notebook expresses the third more explicitly. This is not an absolute classification. Projects can be instructed to apply tighter source discipline, while Notebook is becoming more closely connected to Gemini’s other tools. Actual behaviour at the chosen entry point matters more than the product name.
Sensitivity of the material may change the decision entirely. Consumer, team and enterprise plans can differ in data use, sharing, retention and administrative control. A contract, unpublished study, client record or health document does not become authorised for upload merely because an interface accepts the file type. Organisational policy, product terms and account settings should be checked first, and approved workspaces used where required.
A small acceptance test using your own material
Before migrating the full archive, prepare a ten-file set containing three deliberate difficulties. Let two sources report different numbers for the same issue. Place an important conclusion only in an appendix. Include an old, superseded version of another document.
Begin with factual-location questions and see whether the system names the file and passage. Then ask about the conflicting numbers and observe whether it smooths them into a single answer. Require a distinction between what a source explicitly says and what can be inferred across several sources. Finally, start a new conversation and see whether project instructions and material persist as expected.
Accuracy is not the only result to record. Count the steps needed to return to the original text. See whether the old version is easy to recognise and whether an updated source is reflected correctly. Over a long project, the cost of deleting, replacing and archiving sources can exceed the convenience of the initial upload.
This test often reveals that the best environment for reading is not the best environment for making the final work. Sources can be checked in a Notebook, then clearly identified extracts and citations transferred to a Project for writing. Alternatively, the whole task can remain in a Project while a separate evidence register protects the important claims. Two tools are not necessarily wasteful when they carry different responsibilities and the handoff is explicit.
Conclusion: choose the distance permitted between source and answer
If the material will support research, writing, planning and a variety of tool use, ChatGPT Projects offers a broad continuing workspace. If one knowledge background must serve repeated analysis and writing conversations in Claude, Claude Projects deserves an early trial. When the priority is reading within a selected collection and returning quickly from an answer to its text, standalone Gemini Notebook is most directly aligned with the task.
The decision should not stop at file limits. What matters is how the system locates material, whether it may cross the source boundary, whether the user can inspect its basis, and whether a mistaken conclusion is carried forward by project memory.
A collection becomes a dependable knowledge environment only when its sources remain identifiable, versions can be distinguished, citations return to the text and provisional judgements remain open to correction. AI can help organise that environment. The environment does not come into existence merely because the upload has finished.
Primary sources
- OpenAI: Projects in ChatGPT
- Anthropic: What are projects?
- Anthropic: How can I create and manage projects?
- Google: Learn about Gemini Notebook
- Google: Add or discover new sources for your notebook
- Google: Notebooks in Gemini Apps
Continue reading: Choosing the Right AI Tool
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.