# academic-ai-prompt disclaims its own headline numbers, and ships a detector-lowering prompt with a figure attached

> A prompt library for graduate research, mostly in Chinese, that claims to save fifty to seventy percent of a researcher's time. It also states in its first paragraph that those figures are not validated, and it contains one prompt whose stated purpose is lowering an AI-text detector score by a third to a half.

**bohyy/academic-ai-prompt** — 一套为研究生和学术研究者设计的完整AI Prompt库  📖 包含内容： ✨ 40+ 精心设计的AI Prompt ✨ 论文选题系统方法（生成、评估、论证） ✨ 论文查找快速方案（8个不同方案） ✨ 文献综述框架和工具 ✨ Excel自动评估表格 ✨ 3个完整的论证模板  🚀 核心优势： ⚡ 节省时间 50-70%（选题3-5天而不是2-3周） 🎯 科学方法（基于系统的5维度评估体系） 💡 即插即用（所有Prompt直接复制可用） 📚 全流程覆盖（从选题到出版的完整方案）  🎓 适用人群： 👨‍🎓 硕士研究生 | 博士研究生 | 本科毕业设计 | 学术研究者 | 内容创作者  

- Repository: https://github.com/bohyy/academic-ai-prompt
- Stars: 2,014 · Forks: 125
- Language: Unknown
- License: MIT
- Published: 2026-09-16 · Updated: 2026-09-16 · Language: en
- Canonical page: https://hysenlabs.com/projects/bohyy-academic-ai-prompt

## The README disclaims its own headline figures before describing them

The first two paragraphs of the document are caveats, in Chinese and English side by side. The first says the prompt files, templates and flowcharts in the subdirectories are currently primarily in Chinese, while the README itself is bilingual. The second is more unusual: it states that the scale, timing and effectiveness figures which follow are carried over from the original project overview, are not validated benchmark results, and that the reader should rely on the current repository contents and their own situation.

The same warning appears a second time before the worked examples, in slightly different words: the timings and dates in those examples are kept from the original description, are for reference only, and you should update your research time range and verify all literature information. The sample output block carries its own label saying the bibliographic information needs checking.

So the repository that promises to save fifty to seventy percent of your research time tells you three separate times not to believe its numbers. That is worth more than the numbers, because a prompt collection that shipped fifty to seventy percent without saying so would be asking for trust it had not earned.

## One of the nine new prompts lowers a detector score, and the repository takes no position on that

Version 2.1 added a section on text polishing and expression, nine prompts numbered 7.1 to 7.9. Seven of them are ordinary editorial work: polishing for academic quality, expanding a paragraph, condensing one, raising academic register, unifying writing style, a guide to improving an AI draft, and a combined pass.

The second one is different. It is described as reducing the AI rate, with a stated figure of thirty to fifty percent. The problem this library claims to solve includes that AI-generated drafts come back with too high an AI rate and inconsistent style, so the section's own framing is that the fix for machine-sounding prose is more machine assistance.

What the repository does not contain is any statement about academic integrity, submission rules, or what a programme's policy is. There is no note telling you to check your institution's rules, and there is no note telling you the opposite. The audience list puts content creators next to doctoral students and undergraduate capstone projects, and the writing section is titled for top-journal papers with a claimed outcome of a top-journal-quality manuscript.

Whether a detector-lowering prompt is appropriate is a question about the rules of the programme you are in, and this repository deliberately does not answer it. Anyone using these nine prompts has to answer it for themselves.

## The paper-finding prompt asks for fifty citations from memory and the fix comes afterwards

The flagship workflow is a prompt that asks a model to recommend thirty to fifty high quality papers on a chosen topic, specifying a research area, core techniques, application domains, four types of paper wanted, a year range, and preferred venues. It asks for each entry to carry a title, authors and year, the venue, a one-line core contribution, why it is recommended, and a star rating, sorted by that rating. It then asks for a summary of how the field developed over ten years, the main remaining problems, and which few papers to start with.

The sample output that follows shows two entries. U-Net from 2015 at MICCAI, and Attention U-Net from 2018 at MIDL, both correctly attributed and both real, followed by an ellipsis standing in for the remaining twenty-six to forty-seven, then a four-period summary of the field from 2010 to 2024.

Then the verification step: check the list in Google Scholar before using it. The README is clear that this is the workflow, and it labels the sample output as needing verification.

So the mitigation is a human check after generation rather than a constraint during it. The prompt contains no instruction to mark uncertainty, to refuse when it does not know, or to flag anything it is unsure of. For a technique whose known failure mode is confidently stated citations that do not exist, that ordering is the thing to keep in mind.

## Both worked examples hardcode a year range that has expired

The two examples in the document are the ones a reader is most likely to copy, and both bake in absolute dates.

The topic-generation prompt includes a requirement that the proposed topics should account for what is currently hot in the field, and specifies the years 2023 to 2024 as the reference for that. Its requirement list, copied out of the prompt, reads:

```text
请帮我生成100个可能的论文选题，要求：
1. 选题应该在机器学习和推荐系统范围内
2. 选题应该既有创新性（不是简单应用现有方法），又具有可行性（两年内能完成）
3. 选题应该包括不同的研究方向：理论突破、方法改进、应用创新、工程优化等
4. 选题应该考虑当前学界的热点（2023-2024年的研究方向）
5. 选题应该与推荐系统和深度学习有关联
```

The paper-finding prompt asks for the last eight years, defined as 2016 to 2024, with the emphasis on 2022 to 2024.

Both ranges are now behind the field. A reader who copies either prompt today sends a model an instruction to weight two or three specific years as the frontier, which is precisely the staleness a prompt library exists to avoid. Requirement four is the whole problem in one line: it names two years as the definition of the state of the art, and it is the line a user would have to notice and edit.

Nothing in the document flags the dates, and there is no placeholder mechanism: the years sit in the middle of prose the way they would in any example. This is the failure mode of copy-and-paste prompt collections in general, and it is worse in a library whose entire value proposition is that the prompts work as written. The fix is a find and replace on four numbers, which nobody in the repository has done.

## The five evaluation dimensions are the one thing in the repository you cannot read

The methodology rests on a scoring spreadsheet. The workflow is: generate a hundred candidate topics with a prompt, score them in a five-dimension evaluation workbook that automatically computes a weighted score, take the top ten, then run another prompt to analyse the top three in depth. The stated effect is that the selection takes three to five days instead of two to three weeks, that the resulting topic is chosen with confidence, and that it saves two to three months of later revision.

The workbook is in the repository as a spreadsheet file, the only binary artefact in the tree. Its contents are described in one sentence: five dimensions and automatic weighted scoring.

So the criteria that decide which of a hundred generated topics you work on for two years are not in the README, not in any prompt, and not readable as text. You have to open the file. Everything else in this repository is Markdown you can read in a browser, and the single component that embodies its central claim about being systematic is the one you cannot inspect without a spreadsheet application.

The claim attached to the outcome, that the traditional approach reaches a topic quality of fifty percent and this one ninety percent or better, has no definition of quality anywhere in the repository either.

## Three different prompt counts appear in three different places

The repository summary says the collection contains more than forty carefully designed prompts. The README says more than eighty-nine. The README's own file listing breaks the number down by directory, with more than twenty for topic selection, more than thirty for finding papers, more than twenty for literature review, and more than thirty-four for writing. Those four figures add to more than one hundred, and the listing's own total line says eighty-nine.

There is also a character count of more than one hundred and sixty-five thousand, three flowcharts, one spreadsheet and three complete templates. The flowcharts check out against the tree, which holds three images, one each for topic selection, paper finding and literature review. The spreadsheet checks out too.

None of this matters much for a document collection, where counts are promotional rather than contractual. It matters because the numbers are the first thing the summary leads with, and a reader who checks any two of the three figures against each other will find that they disagree.

## The README's own file listing stops mid-line and omits a flowchart that exists

The project contents section prints a text tree of the repository, and it ends in the middle of an entry: the paper-finding flowchart line opens a parenthesis describing the paper-finding process and never closes it, with nothing after.

The entry after it, the literature review flowchart, is not in the listing at all, even though the tree contains the image. So the listing is both truncated at its end and missing an item that exists.

Everything else in the repository uses Chinese names for directories and files: four series directories for topic selection, paper finding, literature review and writing, an index file with no file extension, and the three flowcharts in English. There is no English version of any prompt file. A reader who works in English gets a bilingual README that points at filenames they will have to paste into a file manager, and every file they open is in Chinese.

There is also no code, no build and no release. The version, 2.1, appears in the README heading and nowhere else, and the repository publishes no releases.

## Conclusion

academic-ai-prompt is a set of copy-and-paste texts, not a system, and the repository is honest about the difference in its opening paragraphs: the sizes and the time savings are carried over from an earlier write-up and are not measurements. Use the prompts as drafting scaffolding and ignore the percentage claims entirely, including the ninety percent quality figure for topic selection, which has no definition behind it. Two things to check before relying on it. The paper-finding prompt asks a model for thirty to fifty citations from memory, so verify every reference before it reaches your bibliography. And decide for yourself where a detector-lowering prompt sits relative to the rules of the programme you are in, because the repository states no position on that and it is the one item here that could cost you more than the hours it saves.

## FAQ

### What is bohyy/academic-ai-prompt?

A collection of prompt texts for graduate and academic research, covering topic selection, paper finding, literature review and paper writing, plus three flowcharts and one spreadsheet. The prompt files themselves are primarily in Chinese, while the README is bilingual. The licence is MIT and the repository publishes no releases.

### Are the time-saving figures in academic-ai-prompt verified?

The README says they are not. Its second paragraph states that the scale, timing and effectiveness figures are carried over from the original project overview, are not validated benchmark results, and that the reader should rely on the repository contents and their own situation. A second warning repeats this before the worked examples.

### How does academic-ai-prompt find papers, and how do you check them?

A prompt asks a model to recommend thirty to fifty papers with titles, authors, years, venues, contributions and star ratings, plus a summary of how the field developed. The README then tells you to verify the list in Google Scholar, and labels the sample output as needing verification. The check comes after generation, not as part of it.

### What does the academic-ai-prompt topic selection workflow involve?

Generating a hundred candidate topics with a prompt, scoring them in a five-dimension spreadsheet that computes a weighted score, taking the top ten, and running another prompt to analyse the top three. The five dimensions themselves are described only as being in the spreadsheet, which is the one binary file in the repository.

### Does academic-ai-prompt say anything about academic integrity rules?

No. One of the nine text-polishing prompts added in version 2.1 is described as reducing the AI-text rate by a stated thirty to fifty percent, and the repository contains no statement about submission rules or programme policy. Whether that is appropriate depends on the institution you are working in.

## Sources

- [bohyy/academic-ai-prompt on GitHub](https://github.com/bohyy/academic-ai-prompt)
- [Issues](https://github.com/bohyy/academic-ai-prompt/issues)
- [License: MIT](https://github.com/bohyy/academic-ai-prompt/blob/main/LICENSE)
- [README](https://github.com/bohyy/academic-ai-prompt/blob/main/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/bohyy-academic-ai-prompt