Feed the Syllabus to Your AI: A G-Test Knowledge Base in Gemini Notebook and Claude Projects

The official G-test textbook is one thick book covering the whole scope, and the syllabus splits into 55 mid-level items. Read everything before touching a question and half of your 40 days is gone to reading. So you ask the AI. But hand it only the name of a major item and ask for the important terms, and it will enthusiastically list terms that are not in the syllabus at all. When this site tried it with Claude, a glossary produced without the syllabus covered 169 of the 495 keywords and returned 438 terms that were not on the list.

This article contains affiliate links. As an Amazon Associate, we earn from qualifying purchases.

This article covers the countermeasure, which we call ingest: feed the syllabus and the official sample questions to the AI and build a knowledge base that answers only within the material it has been given, with citations. The tools are Gemini Notebook, formerly NotebookLM, and Claude Projects, compared against ChatGPT Projects and Gemini Gems. Every limit below was taken from the vendor help pages on 21 September 2026. The second half reports what we measured when we built glossaries with and without the syllabus.

What ingest means: fix the scope before you ask

A general AI chat answers from its training data. With no scope specified, asking about the G-test returns general knowledge about AI. That is not wrong, but it does not coincide with the exam scope.

Ingest changes the premise. You hand over the syllabus and the samples as files and establish the relationship: answer on the basis of this material. How the material is treated differs by tool. Gemini Notebook quotes the relevant passage as grounding, attaching a citation number you can click to return to the original. Claude Projects places the material directly in the conversation context, and on paid plans switches to retrieval once the material approaches the context limit.

The benefit of this arrangement is that when the answer is not in the material, you can tell. Gemini Notebook states in its official FAQ that if the answer is not in the source material it will not provide a response, so an out-of-scope question is reported as out of scope by the silence. With the Projects-style tools you get the same effect by asking for the syllabus item number and discarding whatever comes back without one.

The name change, and the plan limits

Start with the name. Google renamed NotebookLM to Gemini Notebook on 16 July 2026. The announcement describes it as the same standalone product that can now do more across the Google ecosystem. Alongside the rename it gained a secure cloud computer that can write and run code natively, rolling out first to Ultra and some Workspace customers and then to Pro users on the web.

This site covered the feature set under the old name in an earlier article. Here we keep both names, because most people searching still search for NotebookLM, and we cover only the part that matters for exam study.

The plan limits, from the official Upgrade Gemini Notebook help page, retrieved 21 September 2026:

The ranges for Ultra exist because the official table separates the 20 TB and 30 TB plans; the lower figure is the former and the higher the latter.

For the G-test you need the syllabus PDF, the official samples, the JDLA introductory deck, the table of contents of the official text and your own notes. That will not exceed ten sources, so the free tier of 50 is plenty. What changed is how usage is counted. From 2 September 2026, following an announcement on 28 August, chat and Studio usage moved to a compute-based measure that factors in prompt complexity, chat length, number of sources and the features used, and the allowance refreshes every five hours rather than daily, up to a weekly cap, with the remainder visible under Settings and Usage. Note that the Upgrade page still shows per-day figures; Google has not reconciled the two pages. Either way, a long chat full of complex questions burns the allowance faster than short questions asked in separate chats.

Four tools on the same axes

Gemini Notebook is not the only tool that ingests. Claude, ChatGPT and Gemini each have a workspace you can attach files to. Here is what the official help pages said on 21 September 2026.

The citation row is decisive for exam study. Gemini Notebook and Gems attach, by default, a reference showing which part of which source an answer came from. Claude Projects and ChatGPT Projects return answers that draw on the material, but their official help does not document an explicit citation feature. Writing an instruction that asks which syllabus item a claim belongs to gets you close, but the tool is not guaranteeing it.

Claude Projects has a different advantage. You can create up to five even on the free plan, and on paid plans the material switches automatically to retrieval as it approaches the context limit, expanding capacity by up to ten times. If your use case is dumping in the table of contents of the official text plus a large pile of your own notes, that suits better. ChatGPT Projects allows sharing even on Free, which matters if you are studying with someone.

Which to use comes down to a single question. If you want citations attached automatically, Gemini Notebook. If you want to stay in the AI you already use, the Projects of whichever vendor that is. The procedure below assumes Gemini Notebook, but the same order works in the others.

What to ingest: public material, and the book you bought

Split the material into what is public and what you paid for.

There are three public sources. The PDF of the G-test scope, the syllabus, from the JDLA site, currently revision 1.4 dated 11 May 2026 and applied from the November 2024 sitting. The page of twenty official sample questions. And the PDF of the JDLA introductory deck, the August 2026 edition. The syllabus lists an objective and a set of keywords for every major and mid-level item, and that listing is the backbone of the knowledge base. The samples teach the shape of the questions, and the deck carries the exam system and the study-time figures.

Whether to ingest the official textbook you bought is a copyright question. Digitising a book you own for your own study is generally understood to fall within private use. Sharing that notebook with other people, or publishing the summaries it generates, goes beyond private use. Keep the notebook to yourself.

The procedure, in five steps

1. Create a notebook and add the three public sources. Name it after the exam and the sitting so that later notebooks do not blur together.
2. Ask for a summary per major item, with the instruction that every claim must carry the mid-level item number it comes from. Anything that arrives without a number is out of scope and gets discarded.
3. Ask for a glossary per major item, restricted to the keywords that appear in the syllabus, with the instruction not to add terms that are not there.
4. Ask for confirmation questions on the summary and the glossary. These are not for memorising answers; they are for checking whether the summary and the glossary are actually understood. Building questions properly is covered in a separate article.
5. Click through the citation to the original. This round trip is what connects reading an AI summary to reading the material.

Step five earns its place. Take model selection and evaluation: the objective in the syllabus is not merely knowing the evaluation metrics but being able to choose an appropriate metric for the purpose. A summary tends to round that down to understand the evaluation metrics. Going back to the source shows that what is asked is the choosing, not the defining. The habit of returning to the citation is how you pick up that information about how things are asked.

If your study session is 45 minutes, an early-stage split of 20 minutes on ingest and checking the summary, 15 on the glossary and 10 on confirmation questions works. Ingest finishes in the first few days; after that you repeat the glossary and the confirmation questions major item by major item.

The five steps above are prompt design with a source attached. If you want to go further into structuring context and instructions so the model stays inside your material, this book is a solid reference.

Prompt Engineering for LLMs

USD 54.51 on Amazon.com (as of 2026/09/22)

Amazon

＼楽天ポイント4倍セール！／

楽天市場

ポチップ

Measured: with the syllabus and without it

To check that the procedure earns its keep, this site had Claude (claude-sonnet-5, no tools and no project settings, a bare model) build a glossary for all ten major items under two conditions. Condition A supplied the syllabus text for that major item, objectives and keywords, with the instruction to include every keyword listed and add nothing that was not. Condition B supplied only the name of the major item, with the instruction to list whatever came to mind. Scoring was against the 495 keywords of revision 1.4, counted per mid-level item: how many generated terms matched (coverage), how many did not (out of scope), and whether the mid-level item number and name attached to a matching term were correct (attribution).

Twenty calls, a few minutes, a few tens of yen. The cost of ingest is not the tokens; it is deciding to hand over the file.

Coverage is the headline, but attribution matters more for study. Under condition B, half of the terms that did land inside the scope were filed under the wrong mid-level item. A glossary sorted into the wrong drawers cannot be used to count your weak areas, because the counting axis does not match the exam. Ingest fixes the scope, and it also aligns the coordinate system of your records with the exam.

Measuring answers with and without the syllabus is a small evaluation. For a fuller treatment of how to evaluate model outputs systematically, see this book on building applications with foundation models.

AI Engineering: Building Applications with Foundation Models

USD 52.40 on Amazon.com (as of 2026/09/22)

Amazon

＼楽天ポイント4倍セール！／

楽天市場

ポチップ

Traps: answers without citations, and the wording of objectives

Some traps showed up in the measurement.

The first is on the evaluating side. We initially scored by mechanically splitting the extracted syllabus text into objective lines and keyword lines. But in the PDF extraction, ten mid-level items had their keyword column run on to the end of the objective line, and 36 terms were counted as part of an objective. Scored that way, condition A appeared to have 33 out-of-scope terms, when in fact those were syllabus keywords. An independent check caught it, and the table above is the recomputed figure after separating the columns properly. Material you hand to an AI should be looked at by a human first to check that the structure survived extraction. Tables and formulas break easily.

The second is that tools differ in how they handle out-of-scope questions. Gemini Notebook states in its FAQ that it returns no answer when the answer is not in the sources. The Projects tools will sometimes fill in from general knowledge, and that part has no grounding in your material. Rather than treating either as a weakness, use both as out-of-scope detectors. Decide in advance: what Notebook declined to answer, and what came back from Projects without an item number, does not get memorised.

The third is the change in how limits bite from September 2026. Opening a new chat per major item and asking short questions uses the five-hour window better than a single long chat.

The fourth is that ingesting the twenty samples with their answers can leak the samples into generated confirmation questions. The samples are material for learning the format, so ingesting them is fine, but check that generated questions are not reheated samples.

Doing the same thing in Claude Projects

If Claude is what you already use, the same procedure fits into Projects. Create one project and put the syllabus PDF and the introductory deck into the project knowledge. In the instructions, write that every answer must carry the syllabus mid-level item number, and that anything not in the syllabus must be labelled out of scope.

Free accounts can hold five projects, which is enough for one exam. On paid plans, as the material approaches the context limit, retrieval kicks in and capacity expands by up to ten times, so adding the table of contents of the official text and a growing pile of your own notes does not break it. What you do not get is a documented citation feature, so the item number you asked for in the instructions is doing that job. Treat an answer without a number the way you would treat silence from Notebook.

Summary

Ingest is the first step because it is the only step that fixes the boundary. Feed the syllabus, the samples and the introductory deck to the tool, require a mid-level item number on every claim, and discard what arrives without one. With the syllabus, a generated glossary covered 99.0 percent of the keywords and attributed 97.4 percent of them correctly; without it, 34.1 percent and 50.0 percent. The tool does not know the exam scope until you tell it.

Tomorrow, the material this builds gets used: explaining terms back to the AI so it can point out the holes.

Sources

• JDLA, About the G-test (syllabus, sample questions and the introductory deck)
• Google, NotebookLM is now Gemini Notebook
• Google help, Upgrade Gemini Notebook (plan limits)
• Google, flexible usage limits for Gemini Notebook
• Google help, Manage your Gemini Notebook usage limits
• Google help, Gemini Notebook FAQ
• Anthropic help, What are Projects?
• Anthropic help, RAG for Projects
• OpenAI help, Using Projects in ChatGPT
• Google help, Use Gems in Gemini Apps

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.