nihaisha: A Claude Agent Skill for Studying Ni Haisha's Traditional Chinese Medicine Courses

What nihaisha Is and Who It Is For

nihaisha is an agent skill designed for students studying the lecture series of Ni Haisha (also known as Ni Haixia), a practitioner of classical Chinese medicine. The skill organizes the course content into a searchable reference accessible through natural language queries in Claude Code, Claude Desktop and similar agents that support the skills format.

The core audience is someone actively working through Ni Haisha's courses on Shanghanlun (Treatise on Cold Damage), Jingui Yaolue (Essential Prescriptions of the Golden Cabinet), acupuncture, Huangdi Neijing and Shennong Bencao Jing, who wants to look up a formula's symptom pattern, compare two prescriptions, find the lecture and timestamp where a specific acupoint was discussed, or trace a classical citation back to its source document.

The skill operates with a defined safety boundary: it will not provide individual diagnosis, prescriptions, dosage guidance, or self-medication advice. Dangerous acupuncture depths for thoracic and back points are flagged with non-individual threshold warnings drawn from a 2006 Japanese acupuncture safety paper, which is indexed separately and labeled as external acupuncture safety reference, distinct from Ni Haisha's own course material.

Course Coverage and Screenshot Evidence

The skill indexes thirteen course modules. Each module has both a text references file and a screenshot evidence index. The Shanghanlun module has 649 screenshot evidence entries in `references/screenshot-evidence.md`. Jingui Yaolue has 656. Zhongjing Xinfa has 68. Clinical cases have 88. Bagang (Eight Principles differentiation) has 33. Fuyang Forum has 37. Yijinjing has 28. Tianji has 527. Huangdi Neijing has 272. Shennong Bencao has 127.

In total the repository documents 2,986 screenshot evidence entries, with the images stored as compressed WebP files in the repository. Each entry can be searched by formula name, acupoint, lesson number, pathomechanism, numerological keywords or timestamp. Beyond screenshots, the skill supports PDF page-level tracing through `references/pdf-evidence/`, allowing retrieval by course module, keyword and page number.

The text processing layer underwent a systematic correction pass in 2026-06 based on proofreading material from the Qihuang Shengxian wisdom collection, correcting transcription errors in terminology, formula names, acupoint names and classical citations that arose from ASR (automatic speech recognition) in earlier versions.

Installing the Skill

The repository provides an installation script:

bash install_as_skill.sh

The `pyproject.toml` defines the package as `nihaisha-rag-prototype` at version 0.1.0, requiring Python 3.11 or newer. Optional dependencies include `faiss-cpu>=1.8.0` for vector search and `FlagEmbedding>=1.3.0` for local BGE-M3 embeddings. The `.env.example` file shows the configuration keys: `SILICONFLOW_API_KEY` for the recommended SiliconFlow BAAI/bge-m3 embedding backend, and `SILICONFLOW_RERANK_MODEL` defaulting to `BAAI/bge-reranker-v2-m3`. Setting `LOCAL_BGE_M3_USE_FP16=false` uses the local BGE-M3 backend after installing the `.[local]` extras.

Text and knowledge search require no API key. Vector and hybrid search modes need the SiliconFlow key. The CLI entry point is `nihaisha-rag`.

RAG Mode Removal and Current Functional State

On 2026-08-08, the RAG plus knowledge graph mode was taken down based on user feedback that it had unresolved issues. The README states that while the mode is offline, no data package download is available for it. The standard course Q and A, formula and pattern comparison, per-lesson review, screenshot search and PDF page-level tracing continue to function using lightweight in-repository data.

The full documentation and architecture diagram for the RAG mode remain accessible at `docs/RAG_GRAPH_MODE.md`. Users who previously downloaded the local data to `data/pdf_rag_bge_m3/` can keep those files for when the mode is re-enabled. The README states that an announcement will be made in the update log when the mode returns.

This removal is a real limitation. The RAG mode with a 3.68 GB data package enabled vector-based semantic search, which allows finding related content by meaning rather than by exact keyword match. Without it, users rely on the text-based index, which covers the structured course notes and screenshot evidence indexes but will miss content that is not indexed by keyword.

Evaluation Results for the Standard Skill Mode

The repository includes a 240-question evaluation set covering five capability areas, nine common user task types and five core course modules. Each question was evaluated independently across three rounds with blind scoring.

The standard skill mode achieved an overall answer score of 91.4% (95% CI 90.1 to 92.7%) across 240 questions times three rounds. Citation precision was 93.2% and citation accessibility was 97.3%, meaning that when the skill cited a source, that source could be located 97.3% of the time. The clinical safety category scored 98.5%, indicating that the skill correctly refused to provide diagnosis, prescription or dosage guidance across 60 safety-boundary questions.

Notably, the capability boundary pass rate was only 8.3% (5 out of 60 attempts), meaning the skill successfully enforced the boundary only 8.3% of the time in that test set. The README does not resolve this apparent inconsistency between the 98.5% clinical safety score and the 8.3% capability boundary pass rate; the two test sets appear to measure different things. Answer consistency across similar questions was 62.5%, which the README documents but does not explain.

What the Skill Does Not Do

The skill is explicitly not for clinical decision-making. It will not give individual diagnoses, prescriptions, dosage figures, or instructions for acupuncture procedures such as needle depth, especially for dangerous locations in the thoracic and back regions. The external acupuncture safety reference, labeled separately from Ni Haisha's course material, provides non-individual threshold warnings from a peer-reviewed Japanese source, but these are warnings, not instructions.

For users who need a general TCM reference database rather than content specific to Ni Haisha's courses, a broader repository of classical texts would be more appropriate. nihaisha is specifically organized around Ni Haisha's interpretation and teaching style. Passages from other practitioners or commentaries are indexed in a separate layer, clearly labeled, so they do not mix with Ni Haisha's own words.

The project has no GitHub releases. It is described as still being iterated on, with the README explicitly advising users to return periodically to check for updates and install the latest version.

Maintenance Status and Licence

The repository was last pushed on 2026-09-16. It is not archived. The README does not specify a licence; the SKILL.md file is referenced but its licence terms are not quoted in the README.

The project is a Python package named `nihaisha-rag-prototype`. The ongoing nature is reflected in the update log, which shows changes from 2026-06-25 through 2026-08-08 covering safety improvements, evaluation additions, terminology corrections and the temporary removal of the RAG mode. The update log is the primary mechanism for tracking functional changes.

For a comparable approach to organizing dense course material into a searchable agent skill, the author's own `lineage-skill` repository at github.com/JuneYaooo/lineage-skill documents the broader distillation method used here.

Editorial conclusion

nihaisha is well suited to students and practitioners working through Ni Haisha's lecture series who want to search across course modules, trace claims to specific screenshots or PDF pages, and compare related formulas and acupoints without switching between video and text sources. The safety boundary is explicit and enforced: the skill will not give dosage guidance, acupuncture depth recommendations for dangerous needle locations, or individual diagnoses. Anyone who needs clinical decision support rather than course study retrieval is explicitly outside the intended scope. The RAG plus knowledge graph mode was taken down on 2026-08-08 and is not available as of the last repository update on 2026-09-16; users who installed that data package can keep the local files for when it returns, but the mode itself is non-functional until further notice.

Frequently asked questions

Official sources

1. Issues
2. JuneYaooo/nihaisha-nishi-tcm on GitHub
3. README