Cover photo taken on 2026.1.22
While the reason for CS finals to exist can’t be entirely denied, that doesn’t change the fact that core-course finals in CS have always been a widely maligned, spirit-grinding affair. For one thing, the final score doesn’t reliably reflect how deep a student’s understanding of the course actually is (that is… deep learning); for another, the effort-to-reward ratio of a final can’t be guaranteed — plenty of the time you pour hours into review and still can’t pull a high score.
I consider myself one of the better exam-takers in my CS cohort, and now that I’ve finally landed what I consider one of the rarest A+s of all — Computer Networks (only 1 person >=98; in 24Fall DSAA, 3 people were >=98) — it feels like a fine time to share. The tips below are ordered by when you’d use them.
Step 1. AI + Slides + Handwritten Summaries
Use a Prompt that suits you, consume it together with the slides, plus a handwritten summary on the side.
I’ve touched up my earlier slide-review prompt; feel free to iterate on it a little based on your own understanding and taste. This is what I refined bit by bit through my own back-and-forth with AI (kept in the original Chinese — it directs the model to explain in Chinese):
现在我正在进行这门科目的学习，你将要作为学习助手帮助我总结课件内容。你的要求如下，必须严格遵守：
如果你可以设置这次对话的标题，将其设置为例如
[CS203(if exist), or Compiler(if not found)][chapter or lecture number] Overview。如果不可以，你则在第一行输出这个内容。对于日常语言，你需要积极的使用中文。但是如果遇到你认为必要或者关键的英语专业名词，可以适当使用英文表示或者双语对照（但是，如果专业名词或者关键词没有合适中文，可以不用双语），这不仅在这次解释中要遵循这个规则，在以后的对话中也要遵循。对于重要的概念或缩写，英文对应是必要的。
不需要前导语，直接进入正题。根据课件内的内容，首先大概的过一遍课件内所囊括的知识点，这一边是为了让我熟悉课件内容，并带有讲解。然后，仔细的讲每一页课件中所有知识点与细节（NOTICE: PAGE BY PAGE），如果必要，附上例子（可以为知识点应用例子，也可以是代码块）。
有些知识点稍微复杂和难理解，你可以在这之外适当加入一些容易理解的内容，但注意这种内容当且仅当你认为这个知识点很难理解。
课件内无关的课程信息不需要讲解，可以直接跳过。
在输出的最后部分，给出一些面向关键概念或者容易混淆、不易理解知识点的小测试（quiz）并附上解答。xml 语言在这里不起作用，不需要设计交互展开格式。
禁止在你的回答中加上 cite 或者
[cite_start]，也不需要前导语，你只用线性过一遍课件。同时记住这个提示词，我们的对话会包括多次课件内容解释以及答疑解惑，在我提出的问题中你要积极使用课件知识，并且在我传入新的课件内容的时候要继续严格遵循以上规则。
I genuinely recommend using Gemini inside AIStudio for this step. Judging by my experience with the various AIs, Gemini’s output is basically the most down-to-earth and easiest to understand. Opus might have the potential to take over that spot, but I don’t have a stable Pro subscription, and the Copilot student verification only comes with 100 Opus prompts, so I never tried it much. (Update: when I wrote this the student account could still use Opus — not anymore! You can register a fresh account + a HK Visa card and mooch a one-month Pro trial; start it mid-month, and once it refreshes on the 1st you get 200 Opus uses. With good prompt design, a single call can do a whole lot of work!)
Then — actually, this is the step I consider most important. Grab a few blank A4 sheets and a pen. Condense the knowledge through your own brain one more time, then handcraft a summary that is genuinely yours. The process is long and painful — one deck of slides can take over an hour — but only this forced pass through your head makes it actually stick. Another reason the trick works: reviewing slides, you’re bound to zone out, and maybe while we’re zoned out the slides just flow past like water, and then I feel like I’ve got it, when really I haven’t. A handwritten summary like this precisely logs your zone-out breakpoints, and pushing through linearly this way, the quality is genuinely guaranteed.
Step 2. Most Professors Are Lazy
That is to say: overfit to past exams, quizzes, and homework with complete peace of mind. Once you’ve fit them well, your score won’t be low, and you can straight-up carry over your high-school problem-drilling techniques. If you can’t find past papers, overfit the homework instead. And make good use of asking senior students for intel plus GitHub repo search — there are still plenty of treasure-trove repos holding fitting material. Also, from what I’ve observed, with AI spreading at flood level, professors have gradually stopped trusting students’ most basic abilities; questions are trending broader rather than deeper — perhaps a point worth noting.
Step 3. Have AI Generate Mock Exams
If the material is sufficient and the time is there, have the AI generate mock exams.
This is honestly a leftover habit from high school — a bit too drill-happy. But if you have the time, it really is worth doing. There’s also a lightweight alternative: the Gemini web app shipped an interactive Quiz feature. Throw the output from Step 1 in there and just tell it to make a quiz, and you quickly get one of those clickable multiple-choice sets. Early in review, when you want to find your footing fast, doing a few of those isn’t bad.
But this definitely works less well than just building a full mock yourself. More often than not, the final is really just a crossover mutation of the past few years’ papers plus some small tweaks, and — granted the overfitting was effective — this is the most familiar training regimen (for me). So: dump {past papers, quizzes, homework, (optional) the slide summaries generated in Step 1} wholesale into a model with a bigger context window, have it generate an exam, print it out, and work through it yourself. Tiring, but effective.