What does this skill do, and when should you use it?
nihaisha is an open-source Agent Skill for Claude Code / Codex that turns over a dozen Ni Haisha courses — Shang Han Lun, Jingui Yaolue, acupuncture, Shennong Bencao, Huangdi Neijing, Tianji and more — into lightweight Markdown modules. It supports retrieval by symptom, formula, acupoint, lesson, or board-screenshot evidence, producing pattern-differentiation summaries, formula comparisons, and lesson-by-lesson review plans backed by 2,986 indexed screenshots and page-level PDF citations. The skill is explicitly educational: it refuses personal diagnosis, prescriptions, dosing, or self-acupuncture guidance. The author built it to help his father study, and ships a 240-question evaluation to back its accuracy claims.
Reads distilled course modules under references/ (learning entry, symptom index, formula-patterns, six-channel, and per-course chapters); runs scripts/search_screenshots.py for ranked screenshot-evidence search and scripts/search_pdf_evidence.py for page-level traceability, citing as pdf-evidence:<doc_id>#p<page> and text-evidence:<doc_id>#s<section>; produces pattern analysis, formula comparisons, study outlines, or Markdown notes that can be appended back into references; automatically runs a second-pass search of Ni-recommended supplemental books when primary course material matches, kept in a separate labeled section; forces medical-care warnings for toxic herbs, emergencies, and dangerous needling-depth topics.
- A TCM learner who wants the course-based distinction between Guizhi Tang and Mahuang Tang formula-patterns, not generic answers
- A beginner asking in plain language ("cold with chills", "diarrhea") who wants questions translated into course differentiation terms first
- A student reviewing a course module who needs a lesson map, key concepts, review questions, and screenshot keywords
- Someone verifying which lesson/page a statement comes from, or locating board screenshots for Xiao Chaihu Tang or acupuncture demos
- A learner organizing notes on meridians, dangerous needling-depth safety references, or Shennong Bencao herb properties
- A technical family member or self-learner who wants dense lecture videos distilled into a searchable, correctable knowledge base
How do you install this skill?
- This is a static source review; no scripts or evals were executed, and scores are low-confidence.
- License metadata is unknown — only copyright/use statements exist; verify redistribution terms before any commercial use.
- The skill is explicitly educational; do not use it for diagnosis, prescriptions, or self-treatment.
- Self-reported weak metrics (8.3% boundary compliance, 62.5% consistency) suggest possible false triggers or inconsistent answers in practice.
- The RAG mode is decommissioned; the installer excludes data/ assets, so full-featured semantic retrieval is currently unavailable.
- Shell / CLI
- Local filesystem
Python (for search scripts)
The default lightweight mode has no extra dependencies — it only reads references/, assets/, and the bundled search scripts. To install, send your AI assistant the prompt: "Help me install the nihaisha skill: https://github.com/JuneYaooo/nihaisha-nishi-tcm". The agent clones the repo and places the folder into its skills directory; restart the agent so metadata reloads. Python is needed to run the screenshot/PDF evidence search scripts (no specific version documented). The RAG + knowledge-graph mode (previously ~3.68 GB) is temporarily offline with no data downloads offered.
git clone --depth 1 https://github.com/JuneYaooo/nihaisha-nishi-tcm.git ~/.claude/skills/nihaisha-nishi-tcmGenerated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Use nihaisha to organize the difference between Taiyang zhongfeng and Taiyang shanghan
- Use nihaisha to find Xiao Chaihu Tang board screenshot evidence
- Use nihaisha to trace chest bi, water qi, and phlegm-fluid in the Jingui course
After installation, trigger it with natural language, e.g. "Use nihaisha to organize the difference between Taiyang zhongfeng and Taiyang shanghan" or "Use nihaisha to find Xiao Chaihu Tang board screenshot evidence" or "Use nihaisha to trace chest bi, water qi, and phlegm-fluid in the Jingui course". The skill detects the entry point (symptom / formula / lesson / evidence) and answers from the matching module; it replies in the language of your request while preserving original Chinese terminology.
What are this skill's strengths and limitations?
- Broad coverage: 13 course modules in text, with 2,986 screenshot evidence entries plus PDF page-level and non-PDF text evidence layers for traceability
- Exceptionally detailed safety boundaries: explicit refusal of prescription/dosing/self-acupuncture, forced warnings for toxic herbs, emergencies, pregnancy, children, and dangerous needling depths
- Published 240-question evaluation (91.4% overall answer score, 98.5% on clinical safety) with public rubrics and judgment files
- Zero-dependency lightweight install; terminology professionally proofread with systematic errata history
- License is unknown; commercial boundaries rest only on the author's non-commercial notice
- RAG + knowledge-graph mode is offline, so cross-corpus semantic retrieval is currently unavailable
- Specialized evaluation shows an 8.3% (5/60) capability-boundary pass rate, 62.5% consistency on similar questions, and 7 source mis-attribution observations
- Liang Dong dialogue and Stanford lecture modules are text-only with no screenshot evidence; adaptation beyond Claude Code/Codex is unverified
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Ni Haisha TCM Course Agent Skill this page | 62 · Recommended | ★ 2.2k | 1mo ago | — |
| Empty Slack Chat Skill | 28 · Not recommended | ★ 15k | 1mo ago | MIT |
| Obsidian Wiki Query | 50 · Use with care | ★ 16k | 11d ago | MIT |
| Book-to-Skill Converter | 57 · Use with care | ★ 37k | 3d ago | MIT |
| Vault Hybrid Retrieval | 54 · Use with care | ★ 16k | 11d ago | MIT |
The source names one alternative route in the same space: Google NotebookLM as recommended by Dr. Andy Lee (Ni's senior disciple) — users upload course PDFs themselves for low-barrier Q&A, but without traceability or sharing. The author also notes the distillation methodology comes from his own lineage-skill project.
How did FollowSkills review this skill?
SKILL.md explicitly reads only bundled references/scripts/assets in a lightweight default mode with no downloads or external side effects; multi-layer medical safety boundaries (no diagnosis/prescription, mandatory referral for emergencies/toxic herbs, externally sourced acupuncture-depth data kept separately attributed). Deducted: unknown license metadata (copyright statement only), unverified publisher, script permissions not fully documented in static view.
Workflow is self-consistent; script invocations, flags and citation formats match across sections; RAG decommission instructions are explicit and non-contradictory. Deducted: scripts and evidence files not inspected/executed; failure-feedback quality on abnormal input unverifiable statically.
Clear audiences, plain-language entry, bilingual policy, explicit non-fit (real medical decisions), all data local to the repo with no overseas service dependency — good mainland-China fit. Deducted: self-reported boundary-compliance rate of only 8.3% and no independent trigger-precision evidence.
Well-layered docs with progressive disclosure, update log, maintenance statement, and stable document IDs. Deducted: no formal LICENSE file/metadata, no semantic versioning or changelog discipline, and a mixed 0.1.0 RAG-prototype pyproject that may confuse.
Rich declared capabilities (2986 screenshot index, PDF page-level traceability, 240-question eval at 91.4%) with clear output formats. Deducted: static review cannot reproduce key paths; evals are self-run with weak boundary/consistency sub-scores, so direct usability lacks independent verification.
README links auditable eval files, rubrics and judgments; citation formats are traceable; course claims are separated from external sources. Deducted: all evidence is self-reported with no third-party execution, and eval file contents were not verified in this static read.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →