Chinese-Grant-Writer-Skills: AI writing skills for NSFC and NSSFC grant applications

The gap between a research idea and a fundable Chinese grant application

NSFC and NSSFC applications are judged on structure as much as on the idea. A 立项依据 section has to trace every claim to a policy or a data source. A 研究内容 section has to break into 专题 and 孙研究内容 in a fixed semantic order. A 技术路线 section has to align with the research content and produce something a reviewer can follow visually. Most researchers write these sections from scratch each time, or copy a successful application from a colleague and adapt it badly.

This repository takes a different route. According to the README, it distills writing patterns from nearly 20 successful management-field applications into four self-contained skills, each covering a specific set of chapters. The target user is a researcher or a research group preparing an NSFC, NSSFC, or provincial-level natural science or social science application, who wants the drafting process broken into stages rather than one long prompt. The skills are not a template library. Each one carries its own knowledge base, style guide, anti-pattern list, and validation rules.

Four skills, one pipeline, and why the order is enforced

The README recommends a fixed sequence: fund-research-content-writer, then fund-technical-route-writer, then fund-background-writer, and finally fund-literature-review-writer. The reason given is dependency. research-content-writer defines the research skeleton (目标, 内容, 关键科学问题), and every later skill anchors to that skeleton. technical-route-writer comes next because the methods and technical route are the execution plan, and background-writer uses the technical route as input for mapping methods to existing limitations. literature-review-writer runs last because it needs the full context (content, problems, methods, gaps) before it can organize references into an argument.

Each skill has a defined input and output. research-content-writer takes a research topic and optionally a project type (面上, 重点, 青年, or provincial), and outputs Markdown for research goals, research content with a lead paragraph, and key scientific questions. technical-route-writer takes the three outputs above plus optional method drafts or flowcharts, and returns research methods, technical route text, a Mermaid diagram, and innovation points drawn from six types (method, perspective, theory, application, framework, integration). background-writer takes six inputs and returns structured 立项依据 text plus a policy list and a data source list. literature-review-writer takes a topic, content, and questions, and returns a review, a critical commentary, and a BibTeX file.

There is a closed-loop check inside each skill. research-content-writer verifies that 研究目标, 研究内容, and 关键科学问题 stay consistent with each other. technical-route-writer verifies the chain from research content to methods to technical route to innovation. background-writer checks four-dimension coverage and scans for inflated phrases such as 首次, 填补空白, and 国际领先. literature-review-writer enforces anti-patterns: no abstract stacking, no timeline review, no references disconnected from the argument.

Running the first skill and configuring the literature search

The repository has three top-level entries: LICENSE, README.md, and skills/. There is no package to install for the writing skills themselves. The README states that Python 3.8+ is required only for the literature search scripts, and that the search dependencies are listed in fund-literature-review-writer/requirements.txt. The skills themselves are directories containing SKILL.md, config.yaml, knowledge_base/, and references/.

Each skill has an information form at references/info_form.md listing required and optional inputs. The README says the more you provide, the less the agent has to guess. The skill generates an outline first, and the README advises reviewing and adjusting that outline before generating the full text. When the draft is done, each skill has a built-in validation menu at references/validation_menu.md and an anti-pattern check.

For the literature review skill, the search pipeline needs configuration. The README instructs you to copy search/.env.example to search/.env and fill in the values. The configuration keys it lists are EXA_API_KEY, OPENALEX_EMAIL, USE_LLM_SUPPORT, LLM_API_KEY, LLM_API_ENDPOINT, and LLM_MODEL.

EXA_API_KEY=your-exa-api-key-here
[email protected]
USE_LLM_SUPPORT=true
LLM_API_KEY=your-llm-api-key-here
LLM_API_ENDPOINT=https://api.example.com/openai
LLM_MODEL=your-model-name-here

EXA_API_KEY and OPENALEX_EMAIL are marked as required in the README. The LLM variables are only needed if USE_LLM_SUPPORT is set to true. After configuration, the search pipeline runs across OpenAlex, Crossref, Exa, and Google Scholar in parallel, filters by journal tier using CSVs in search/data/, and outputs a BibTeX library you can import into Zotero. The README does not document a dry-run mode or a way to inspect intermediate search results before the BibTeX file is written.

The README also notes that Markdown or plain text documents such as .md and .txt are preferred as input, and that if you need to extract content from a PDF, it recommends converting it with MinerU to get cleaner structured text.

Where the literature pipeline stops: English sources and rate limits

The README puts a warning box directly under the literature review skill description. Two constraints matter. First, the current search pipeline only supports English-language literature search and does not support automatic search of Chinese literature such as CNKI. For an NSFC application, where Chinese-language sources often carry weight in the 国内外研究现状 section, that is a real gap. The README recommends that users provide their own literature and abstract lists whenever possible, because automated search coverage is limited and user-supplied references improve the review's quality and focus.

Second, the README notes that each search source may rate-limit, and advises configuring API keys in advance and preparing to supplement manually. There is no documented retry policy, backoff behavior, or result cache in the README. If you run the pipeline on a large topic, you should expect to fill gaps by hand.

The background writer has its own constraint. It performs automated policy retrieval from sources including 国务院, 科技部, NSFC, and various 主管部门, and automated data retrieval from 国家统计局, industry blue books, and authoritative white papers. The README does not state what happens when a retrieval returns nothing, or how the skill handles a policy that has been superseded. Treat the policy and data lists as a starting point to verify, not as citations you can paste in unchecked.

How this differs from writing the application in a general chat assistant

A general assistant such as ChatGPT will draft a 立项依据 section if you paste in your research topic. The difference here is the structure imposed before generation. Each skill carries a knowledge base derived from successful management-field applications, a style guide, an anti-pattern list, and a validation menu. The research content skill enforces a three-level structure and a six-step semantic order for 孙研究内容: complex problem, real-world difficulty, research limitation, essential contradiction, breakthrough need, and problem landing point. A general assistant has no reason to follow that sequence.

The literature review skill is the clearest contrast. A chat assistant typically produces a summary of the papers you paste in. This skill organizes references by research content rather than by timeline or author, and produces four types of critical judgment: methodological, data applicability, model explanatory power, and research boundary. It also runs a four-source parallel search rather than relying on whatever you paste. The trade-off is that the search covers English sources only, so a chat assistant with browsing enabled may actually reach Chinese sources this pipeline cannot.

Within the repository, the skills are also designed not to be copy-pasted into each other. The README says the skills reference each other's output by design but each has its own writing conventions, so you should let each skill perform its own transformation rather than pasting text between them.

Licence, maintenance, and what an upgrade would cost you

The repository is MIT licensed, which permits commercial and non-commercial use, modification, and redistribution provided the copyright notice and permission notice are included. The LICENSE file sits at the top level alongside README.md and skills/. This article does not give legal advice; if you plan to redistribute the skills inside an institution or a company, read the LICENSE text yourself.

On maintenance: the last push to the default branch was on 2026-06-25. The repository is not archived. There are no releases retrieved, so there is no versioned upgrade path documented in the README. Because the skills are plain directories with SKILL.md, config.yaml, knowledge_base/, and references/ files, an upgrade would mean pulling new versions of those files and re-reading the style guides, since the writing conventions themselves are the substance of the project. If you have customized config.yaml or added your own knowledge base entries, those changes would need to be reconciled by hand. The README does not describe a migration process or a changelog for skill content.

The Python dependency set is confined to the literature search scripts, so a change to the search pipeline would not affect the three writing skills. That separation is worth noting if you only care about the drafting workflow.

Editorial conclusion

Adopt this if you already have a research design and want a structured drafting pipeline with validation menus and anti-pattern checks for NSFC or NSSFC chapters. Do not adopt it if you need Chinese-language literature search, since the README states the built-in pipeline only covers English sources, or if you expect the skills to invent your research content. Before committing, check whether the search scripts run on your Python 3.8+ environment, and confirm that your institution's requirements for 立项依据 and 研究现状 match the four-skill split.

Frequently asked questions

Official sources

1. HuiyuLi-2000/Chinese-Grant-Writer-Skills on GitHub
2. Issues
3. License: MIT
4. README