bohyy/academic-ai-prompt: a Chinese-language prompt library for the thesis pipeline

The problem this library targets: thesis work that stalls before writing starts

The README frames the project around six stated difficulties: choosing a topic, finding papers, lacking a way to organise them, AI-drafted text that reads as AI-written, repetitive low-value work, and deadline pressure. The intended users are listed explicitly: master's students, doctoral students, undergraduate final-year projects, academic researchers and content creators.

That is a narrower audience than a general prompt collection. The value proposition is not "prompts for research" in the abstract but a fixed sequence: generate candidates, score them, argue for a shortlist, then find and review literature. The repository layout reflects that sequence, with separate directories for topic selection, paper finding, literature review and writing. If your bottleneck is somewhere else, such as experiment design or statistical analysis, nothing in the listed directory structure addresses it.

How the repository is organised and what each directory contains

The top level holds a bilingual README, a LICENSE file, three PNG flowcharts (topic selection, paper finding, literature review), an index file named 索引, and four Chinese-named directories: 论文选题系列, 论文查找系列, 文献综述系列 and 论文撰写系列.

Inside those directories the README lists specific files: 论文选题AI提示词库.md with 20+ topic prompts, 论文选题论证方案模板.md with three argument templates, 论文选题5维度评估表.xlsx as a scoring tool, 论文查找和文献综述AI提示词库.md with 30+ finding prompts, AI直接找论文的Prompt.md with eight fast schemes, 文献综述AI提示词库.md with 20+ review prompts, and 顶刊论文写作AI提示词库-2.md with 34+ writing prompts including nine text-optimisation prompts added in v2.1.

The README claims a total of 89+ prompts, 165,000+ characters, three flowcharts, one Excel tool and three complete templates. Those totals are the project's own description of itself. The README also carries an explicit note that the size, timeline and effectiveness figures are retained from the original project overview and are not validated benchmark results.

Installing nothing: how to run your first topic-generation prompt

There is no package, no CLI and no server. The README's quick-start path is to open a prompt file, copy the prompt, fill in your own details, and paste it into ChatGPT or Claude. The tutorial below follows the README's own worked example for a master's student in machine learning and recommender systems.

First, open the topic-selection prompt file and copy the generation prompt. It asks the model for 100 candidate titles, each with a short description and a stated reason the topic has value.

我是一名硕士研究生，计划选择关于机器学习和推荐系统的论文选题。

我的专业背景和兴趣：
- 本专业：计算机科学与技术（机器学习方向）
- 研究兴趣：深度学习、推荐系统、自然语言处理
- 已有知识：深度学习基础、Python编程、论文阅读能力

After pasting the full prompt into the chat interface, the README says you should receive a list of 100 topics with descriptions and value statements. Treat that list as raw material, not as a decision.

The second step is scoring. The README directs you to 论文选题5维度评估表.xlsx, which it says computes a weighted score automatically and takes five to ten minutes for the full list. You then take the top ten and run the deeper analysis prompt on the top three.

The README's stated comparison is two to three weeks of indecision against three to five days of this process. That comparison comes from the project's own overview and the README flags it as unverified, so the useful part is the procedure, not the number. The concrete output you should expect is a ranked shortlist and a written argument you can take to a supervisor.

The paper-finding prompts trade recall for speed, and the README says so

The paper-finding directory offers eight schemes of varying length, from about three minutes to about fifteen, plus a prompt that asks the model to name 30 to 50 papers directly. The README's own example instructs the model to sort by a one-to-five recommendation rating and to append a summary of how the field developed, the open problems, and which papers to use as an entry point.

This is the design choice worth pausing on. Asking a language model to enumerate literature produces plausible titles, authors and venues, and the README repeatedly tells you to verify in Google Scholar and to check all literature information. It even labels the sample output as requiring verification. So the intended workflow is generate-then-verify, not trust-then-cite. If you want a tool that queries an index and returns retrievable records, this is a different kind of instrument, and the eight schemes exist precisely because the direct-ask approach has limits.

The time claim, five minutes instead of ten to twenty hours, is again from the project overview and is not a measured result.

The nine text-optimisation prompts and the AI-detection question

Version 2.1 adds a block of nine prompts under 文本润色与表达 in the writing library: polishing, reducing AI-detection rate, expanding a paragraph, condensing one, raising academic register, an improvement guide, unifying style, an accuracy check, and a combined optimisation prompt.

The README states the AI-rate prompt can reduce detection by 30 to 50 percent. That figure is presented without a stated method, sample or detector, and the README's general disclaimer covers it. Whether any particular detector responds the way the prompt intends is not something the repository documents.

There is also a boundary issue the README does not resolve. Most venues and institutions have their own rules about AI assistance in manuscripts, and a prompt that specifically aims to lower a detection score sits close to that line. The library gives you the tool; it does not tell you whether your programme permits it. That is a judgement you make against your own institution's policy, not something the repository can settle.

Where this library is the wrong tool

The README states plainly that the prompt files, templates and flowcharts are currently primarily in Chinese, even though the README itself is bilingual. If you write in English, you can still read the instructions, but the artefacts you would copy are Chinese, and machine-translating a prompt changes the register cues that make it work. For English-language thesis writing, the practical benefit drops sharply.

A second limit is scope. The four directories cover topic selection, paper finding, literature review and writing. Nothing in the listed structure covers data analysis, experiment design, statistics or submission logistics. A researcher whose bottleneck is in one of those areas will find the library largely irrelevant.

A third is that the method is prompt-driven, not data-driven. The five-dimension Excel sheet scores candidates you generated, and the paper lists come from a model rather than a bibliographic database. Neither produces a citation you can trust without checking it yourself.

Compared with a general prompt collection

The obvious alternative is a general-purpose prompt repository or a curated list of academic writing prompts, of which the search data suggests many exist. The difference is structure. A general collection gives you prompts to pick from; this project gives you a pipeline with a fixed order and a scoring artefact in the middle.

The Excel sheet is the clearest divergence. A generic prompt list has no equivalent of a five-dimension weighted evaluation that you fill in between generation and shortlisting. That artefact is what turns a brainstorm into a ranked decision, and it is also the part most tied to a particular evaluation scheme, so it will fit some disciplines better than others. The three flowcharts play a similar role: they describe a sequence rather than a set of options.

The cost of that structure is rigidity. If your process does not match the four stages, you are using a subset of the library rather than the library.

Licence, maintenance and what upgrading costs you

The repository is MIT-licensed, and the LICENSE file sits at the top level. MIT permits reuse and modification with attribution and without warranty. That is a permissive arrangement, and it means you can adapt the prompts into your own materials. It is not legal advice, and it says nothing about your institution's rules on AI-assisted writing, which are separate.

The last push to the default branch was on 2026-09-02, two weeks before this writing, and the repository is not archived. There are no retrieved releases, so version 2.1 exists as a README label rather than a tagged artefact. Upgrading therefore means pulling the default branch and diffing Markdown files yourself. Because the prompts are plain text, a merge conflict is readable and resolvable by hand, but any local edits you made to a prompt file will collide with upstream changes. If you customise the library heavily, keep your edits in separate files rather than editing the originals.

Editorial conclusion

Adopt it if you write a thesis in Chinese and want a structured, copy-paste starting point for topic generation and literature work rather than another list of prompt examples. Skip it if your writing is in English, since the README states the prompt files, templates and flowcharts are currently primarily in Chinese. Before relying on it, open the topic-selection prompt file and the five-dimension Excel sheet and check that the fields match your discipline, and treat every paper the prompts suggest as unverified until you confirm it in Google Scholar.

Frequently asked questions

Official sources

1. bohyy/academic-ai-prompt on GitHub
2. Issues
3. License: MIT
4. README