PaperForge: four prompt files that reconstruct a paper's reasoning

Four text files, no licence file, nothing to execute

The entire repository is `README.md`, `SKILL_CHN.md`, `SKILL_EN.md` and `System_Prompt.txt`. There is no source directory, no test suite, no CI configuration, no package manifest and no release. Repository metadata records the licence as unknown and the primary language as unknown, which tracks with a README written mostly in Chinese.

That makes the adoption question different from most projects here. There is nothing to install and nothing to break; what you take is the wording. The last push was on 2026-09-25 and the repository is not archived, so the files are being edited, but with no releases and no tags there is no version boundary to fall back to. If you need to show a colleague exactly which wording you used, copy the file and keep it.

Section 3 forbids using the paper's own contributions as premises

The flow asks the model to reconstruct the authors' thinking path before it explains the method, and it does that under an unusual constraint. That section must not take the paper's own contributions as given. Only background that already existed, known failure modes, empirical observations and related work are allowed as premises.

This is the part that separates the prompt from a summariser. A summary starts from what the paper claims and works backwards to a plausible story. This reconstruction has to argue that the idea was reachable from what came before, using only prior knowledge. If the authors' contribution cannot be derived that way, the exercise exposes a gap in the motivation rather than smoothing it over.

It is also the reason the copyable excerpt in the README puts this step third, before the intuition and the method. The opening of that copyable prompt reads:

你的任务是：清晰、易懂、深入、详细的总结这篇论文（读取PDF、搜索arxiv等各种信息源获取论文）。
你的总结需要条理清晰的包含下面环节：
1. 论文提出并解决的研究问题是什么（适当搜索调研和补充背景）？为什么这个问题是重要的？解决这个问题能带来哪些价值？
2. 这个问题之前被解决了吗？之前的研究为什么存在不足？
3. 在正式讲方法之前，先重建作者可能的思考路径。这个部分不要使用论文自己的贡献作为前提，只使用论文之前已有的背景、失败模式、经验观察和相关工作。
4. 这篇论文提出方法的Intuition是什么？易懂清晰concise的告诉我这篇论文核心idea的本质。

Item 4 asks for the core idea in plain terms, item 5 walks a real example through input, processing and output, and item 6 asks for the derivation from zero with the theory background filled in, with a requirement that every formula use correct markdown so it renders.

Twelve sections, and the author names only four as load-bearing

The installed skill is described as triggering a complete 12-section analysis flow when you send a paper link or a title in conversation. The README then singles out four sections as the ones that matter: the reconstruction of the author's thinking path, the most fragile assumption, the minimal reproduction experiment and the strongest counterexample design.

That list is a good map of the intent. Three of the four are adversarial rather than descriptive. Naming the weakest assumption and designing the counterexample that would break the claim are the moves that separate a reader who understood the paper from one who absorbed its framing.

The copyable version embedded in the README shows only the opening of the flow, running from the task statement through items on the research problem, the prior work and its shortcomings, the thinking-path reconstruction, the intuition, the worked method and the derivation, and stopping partway through the item on experiment design. The remaining sections live in the three files themselves.

Option B is labelled the personal recommendation and then reports the other method winning

Option B is the system prompt route, and its heading marks it as the author's personal recommendation. Inside it sit two methods. Method 1 creates a Project and puts the prompt in Custom Instructions, after which you send a link, a title or a PDF in the conversation; the stated advantage is reuse, since the prompt does not need pasting each time. Method 2 pastes the prompt straight into the chat and then uploads the PDF, the link or the title.

The anecdote that follows cuts against the heading. The author says Method 1 had been the habit for a long time and felt convenient, then Method 2 was tried recently and produced summaries of somewhat higher quality, with no explanation offered beyond a guess that the position of a system prompt versus a user prompt might matter. That is the state of the evidence: an unresolved personal observation, not a finding.

The practical reading is that the two setups are cheap to compare, since the difference is only where the text sits.

The Chinese file targets Claude and Codex, the English one only Claude

The three deliverables are not translations of one another with different filenames. `SKILL_CHN.md` is the Chinese version, adapted for the Claude and Codex skill systems, installable directly. `SKILL_EN.md` is the English skill, formatted for the Claude skill system alone. `System_Prompt.txt` is meant to be pasted into ChatGPT or Claude custom instructions.

The asymmetry is worth noting before you pick a file. Someone on Codex has the Chinese path and no English one; someone on ChatGPT has the plain prompt and no skill file at all. Whether the English skill is simply a translation or has been reworked for the Claude format is not stated.

The prompt also assumes it can go and get the paper. Its opening task line instructs the model to read the PDF and search arxiv and other sources to obtain the paper, which means the workflow expects network access and a model willing to fetch rather than only answer from what you pasted.

Adapting it to the humanities is a copy-paste instruction, not a branch

Option D addresses the fact that the current version is aimed at science and engineering papers heavy on experiments and methodology. The adaptation route is not a maintained variant. You paste the repository URL into GPT with web search and reasoning enabled, along with a request to research how reading humanities and social science papers in a chosen field differs from reading science papers, and how the skills in the repository should be changed to suit it. The instruction asks for four things to be preserved: reverse-engineering the author's line of thought, decomposing the argument structure, identifying key assumptions, and finding questions that extend the work.

One placeholder does the customisation, the specific field, with ten offered values: history, sociology, political science, philosophy, anthropology, education, communication, economics, law and literary studies.

Option C covers the other self-improvement path, for readers who already have a prompt of their own. The model is asked to compare your version against PaperForge, keep the parts of your workflow that fit you, and merge a version matched to your own research direction and reading habits.

The provenance is one social media post

The Reference section points at a single Xiaohongshu post, titled in the author's framing as the best prompt for reading papers with AI after two and a half years of use. The link carries share parameters and a token in its query string, which is the shape that platform hands out for a share URL.

That is the entire citation trail. There is no paper, no evaluation and no comparison in the repository, so nothing here tells you whether the 12 sections produce better retention than a shorter prompt, or whether the longer flow costs you attention on short papers.

Which is not necessarily a defect for what this is. The artefact is a prompt, the audience is a reader who intends to edit it, and Option C exists precisely because the author expects people to fork it into their own workflow. The honest framing is that you are adopting a starting point, and `System_Prompt.txt` is the one file worth editing first.

Editorial conclusion

Use PaperForge if you already read a lot of papers and want a prompt you can edit rather than a tool you have to install. Do not expect reproducibility, tests or a licence: the repository root has four files, the metadata names no licence, and there is no LICENSE file. Before adopting it, read SKILL_EN.md or SKILL_CHN.md end to end, copy System_Prompt.txt into a scratch chat to compare it against a Project-level setup, and ask the author about licensing if you plan to redistribute any of the three files.

Frequently asked questions

Official sources

1. FeijiangHan/PaperForge on GitHub
2. Issues
3. README