How to Use AI Projects for GMAT® Prep: The Complete Setup
A few months ago we published a guide to using AI for GMAT® prep, and its advice came down to two habits: give the AI real context about your situation, and manage the context window by starting a fresh chat before the current one degrades. Part of that system was a copy-paste ritual. You saved a context document on your computer, and every time you opened a new chat, you pasted your background back in.
That ritual works, and if you never set up a project, it's still the right move. It also creates a bottleneck. GMAT® prep runs for months, and pasting the same background into dozens of chats creates friction, the kind that makes the ritual easy to skip. After a few weeks it can drop out of the rotation, and once it does, a new chat may start without the background you meant to carry into it.
A project can retire most of that ritual. You may know the pattern by its brand names: ChatGPT Projects, Claude Projects, or Gemini Gems. The names differ, and the implementations differ in the details, but the shared idea is one pattern: a workspace where you save instructions and reference files once, and new chats inside it can draw on what you saved. We'll call it a project from here on.
This post is the project setup we'd use for GMAT® prep: three projects, what goes in each one, and the prompts that make them useful.
One provider note up front. Google has announced that Gems will transition to "skills" starting in November 2026 on personal accounts, and existing Gems will migrate automatically. The setup below should survive that transition, because it depends on what you save, not on what the button is called. Features move quickly in this space, so if a menu name has shifted since we wrote this, the provider's help center has the current version.
Why Chat Sessions Evaporate
An AI chat has a limited working context. Your scores, your gaps, and your target date can all be on the table at the start of a conversation, but after a few questions that background competes with the conversation itself for space, and the answers get less specific as the chat gets longer. Start a new chat and that background may be gone unless you reload it. Our April guide's answer was to refresh chats often and paste a context document back in each time, which keeps the answers sharp but hands the reloading work to you.
The problem isn't the strategy. It's the repetition tax. A GMAT® prep asks you to re-establish context over and over: the error log session on Tuesday, the study plan check-in on Saturday, and the question review after each practice test. Each one needs the same background, each re-pasting is a small enough chore to skip, and skipping it means generic answers that sound like the average of the internet instead of advice about your prep.
A project can take that reloading over. You describe your situation once, in the project's instructions. You upload the files the AI should know about, your score history, your error log, and your study plan, once. New chats inside that project can use all of it from their first message.
You'll probably still start fresh chats when the answers turn repetitive or fuzzy, because a chat's own context window still fills up no matter where the chat lives. The difference is that reloading your context stops being your job. The project can handle it every time, for as long as your prep runs.
The Three-Project Setup
One project for your error log, one for your study plan, and one for question review. Three works because each project accumulates a different kind of history, and mixing them means the AI's attention gets split between jobs that don't belong together. A chat that's simultaneously reviewing a missed question, tracking your habits, and deciding what to study next is doing three jobs shallowly instead of one job well.
Project 1: The Error Log
If you've read our error log guide, you know the two-list system: Category A for execution errors, questions you knew how to do but missed, and Category B for knowledge gaps, questions you couldn't do yet. The two categories need different fixes, which is where the project comes in. The AI can classify new misses into the right list as you report them, and it can compare this week's entries against the weeks before, which is the part a spreadsheet on its own won't do.
Instructions to paste in:
You maintain my two-list GMAT® error log. Category A means execution errors: misreads, computation slips, and questions I knew how to do. Category B means knowledge gaps: content I haven't learned well enough yet. When I paste my results, classify every miss into A or B, add it to the running lists, and tell me whether any Category A pattern is repeating across recent weeks. Do not generate new practice questions.
Files to upload: your log itself, as a simple document or spreadsheet. Keep it minimal: one phrase per mistake, the section it came from, and the rule you're working on. The project can read it each time you report new results, which keeps you from maintaining the same data in two places.
Chats that live here: end-of-session entries ("finished 20 questions, here's what I missed") and a weekly pattern review where you ask what the lists are telling you.
Project 2: The Study Plan
Your study plan is the document that tells you what to do next, and our study plan guide breaks a full prep into four phases: diagnostic, knowledge building, practice test cycling, and final preparation. The plan shifts as your scores move, which is why it deserves its own project.
Instructions to paste in:
You hold my GMAT® study plan and my progress against it. I'm targeting a [score] on [date], studying about [number] hours a week. When I check in, compare what I did to the plan, name which phase I'm in, and pick one to three focus areas for next week. Don't add more than three, and don't recommend a practice test just because a week has passed.
Files to upload: the plan itself, plus your two or three most recent practice test score reports. Add newer ones as you go and delete old ones. A small, current pile describes the test taker you are now; a complete but stale pile describes someone you were in September.
Chats that live here: a weekly check-in ("here's what I studied and my latest score, what should next week cover?") and the review you do after each practice test, where the project's memory of your previous scores can make the trend visible instead of something you're reconstructing from screenshots.
Project 3: Question Review
The review conversation after a missed question is different from the log entry that records it. The log captures the miss. The review works through the reasoning, and that's where guardrails matter most, because the AI is doing live reasoning work rather than sorting entries.
Instructions to paste in:
You help me review GMAT® questions I've missed. I'll paste a question and the official answer, and I'll tell you where I got stuck or what I picked. Walk me through the logic of the correct answer, then trace where my reasoning diverged from it. When we finish, ask me to restate the logic in my own words, and suggest one line for my error log. Never invent your own practice questions.
Files to upload: none required. The question, the official answer, and your attempt belong in the chat itself, and a fresh chat per question keeps each review clean anyway.
Chats that live here: one chat per question, started fresh each time. Two guardrails come with this project, and both come from our April guide. For math, an AI explanation is a supplement rather than a substitute, so verify the computation and lean on explanations from real humans when the question is GMAT®-level. And give the AI the official answer before you paste the question, especially on verbal, so it's confirming logic instead of re-deriving it.
A Week With the Projects
Say you've just finished a 20-question quant set and missed five. You open the error log project and type:
Finished 20 quant questions, missed 5. Here they are with what I did: [paste]. Classify each as Category A or B, add them to my log, and flag anything repeating from the last two weeks.
The project can compare those five misses against every entry it has seen before, so "anything repeating" gets an answer drawn from a record rather than a guess. If computation slips under time pressure show up for the third week running, that's the kind of pattern worth hearing about before your next practice test.
Saturday, you take a practice test. The review happens in the question review project, one fresh chat per missed question, official answer pasted first. Then the scores go into the study plan project:
Week 6 check-in. This week I covered [topics] and took Tuesday's practice test: [scores]. Where am I against the plan, which phase am I in, and what should next week cover?
The reply should come back with your history attached: the focus areas from your last two tests, the phase you're in, and one to three things to work on next week. The record stays put. Next week's chats start from it instead of from zero.
Where Projects Don't Help
One-off questions don't belong in a project. A project is a memory for ongoing work, and a question you'll ask once, like how to set up a rate chart or why an answer choice weakens an argument, doesn't accumulate history. Dropping it into a project buries your ongoing work under noise, so keep a plain chat for one-offs and save the projects for threads that run for weeks.
Plan tiers vary, and sometimes they decide for you. ChatGPT includes projects on free accounts, though free projects hold fewer files (five per project at last check, versus 25 on Plus). Claude's project features are documented under its paid plans. Gemini offers Gems on personal accounts, with the skills transition coming in November as described above. If your tool restricts something, the three-project structure shrinks gracefully: the error log project is the one we'd keep, because that's where accumulated history helps most.
A project also doesn't fix accuracy, so the main cautions from our April guide still apply inside one. AI-generated GMAT® questions are a trap, math answers deserve a calculator check, and official materials stay your practice source. A better-informed AI is not automatically a more reliable one.
And it doesn't diagnose. A project may surface the fact that your Category A list is dominated by misreads, but it can't watch you solve and tell you whether a knowledge gap, an execution habit, or time pressure is behind a missed question. That analysis is still yours (or a coach's), and the projects make it easier by keeping the evidence in one place instead of scattered across months of chats.
Frequently Asked Questions
What is an AI project?
An AI project is a persistent workspace inside an AI tool where you save instructions and files once, and new chats within it can use them from the start. ChatGPT calls the feature Projects, Claude calls it Projects, and Gemini calls it a Gem (transitioning to "skills" starting in November 2026). The shared idea is context that persists between chats.
Do AI projects cost money?
It depends on the provider. ChatGPT includes projects on free accounts with tighter file limits (five files per project at last check), while higher tiers raise the limit. Claude documents its project features under paid plans. Gemini offers Gems on eligible personal Google accounts. These boundaries move, so check the provider's current plan page.
What should I put in my project instructions?
Three things: who you are as a test taker (target score, test date, and weekly study hours), how you want the AI to behave (the rules and definitions for that project, like the two error categories), and what you never want it to do (generate practice questions, grade you, or pad its answers). Write it like a briefing for a helpful teammate who starts Monday and has never met you.
Do I still need to start new chats if everything is in a project?
Probably, yes. A project reloads your instructions and files into each chat, but the chat's own conversation still fills its context window as you talk. When answers start getting repetitive or fuzzy, start a fresh chat inside the same project. Your setup reloads without any pasting, which is the part that used to be manual labor.
Can an AI project replace my error log spreadsheet?
No. The log document is the source of truth, and it needs to be something you own and can read. The project can read that document, help you fold in new entries, and spot patterns across weeks that are tedious to find by eye. Keep the log simple, keep it yours, and let the project do the remembering around it.