nihaixia: a Ni Haixia TCM Agent Skill for Claude Code and Other Runtimes

What nihaixia packages, and who it is written for

The repository turns Ni Haixia's teaching corpus into an Agent Skill: a folder of Markdown and reference files that an AI agent loads when a trigger phrase appears. The README names the activation words directly, including 倪海厦, 海厦视角, 倪师, 经方思维 and 倪海厦会怎么看. Say one of those in a session and the skill is supposed to answer in Ni Haixia's frame rather than a generic one.

The intended user is not a patient. It is someone who already reads classical Chinese medicine and wants retrieval speed: which formula covers this presentation, what the composition and dose are, which channel pattern is at play, whether a comparable case exists. The README's own examples follow that shape. A question about chills, no sweat and neck pain returns 太阳伤寒, 麻黄汤证, with a one-line rationale about cold binding the exterior. A question about 小柴胡汤 returns the 少阳 indication and the phrase 但见一证便是.

The scale claimed is large: 129 Shanghan Lun clauses, 23 Jingui chapters, 72 Huangdi Neijing chapters, 374 materia medica entries split 137 upper, 110 middle and 127 lower, 1,257 structured cases plus 243 narrative cases, and 2,452 pages of lecture material. Those numbers are the project's own statement of coverage. Nothing in the repository layout lets an outside reader verify them without opening the files, which is the first thing a cautious adopter should do.

Six-channel reasoning, formula cards and the distillation pipeline

The mechanism is retrieval plus a fixed reasoning order, not a trained model. The repository root holds SKILL.md as the entry point, with modules/, references/, cases/ and distilled_cases.md alongside it. The README describes SKILL.md as carrying quick-reference blocks that were inserted at the entrance: the 开阖枢 diagram, the 原穴 full table and the 五输穴 formula, among six blocks added in v2.2.0.

The reasoning order is the part that distinguishes this from a plain document search. Version 2.1.0 is described as the upgrade from a knowledge-base query tool to a clinical reasoning assistant with a six-channel pattern mode. The README lists 8 diagnostic formulas, a rapid diagnostic flow chart, a pulse-and-tongue quick lookup and a seven-step thinking model. That sequence is what the agent is meant to walk through before naming a formula.

Formula cards are a separate layer. The v2.3.0 notes describe a 经方条文卡片体系 moved to a must-read position ahead of the answer, with version labels on cards and a mechanism to keep Chinese numbering from colliding. The v2.2.8d notes describe a reconciliation check that forces a card to appear the first time a formula is mentioned, plus session-level deduplication of assertion phrases. In practice this means the answer is assembled from cards rather than generated freely, which is a deliberate constraint on the model's improvisation.

The corpus itself was produced by a separate tool. The README badges a 中医思维蒸馏器 at v4.4.0, linking to jangviktor-web/tcm-distiller. That is worth noting because it tells you the knowledge files are generated artifacts with their own version history, not hand-written notes.

Installing nihaixia and running a first query

The README gives several install paths. The recommended one is ClawHub, and the README notes that ClawHub carries V2.2.0 while V2.3.1 was still in system review at the time of writing. Two commands are given:

openclaw skills install @jangviktor-web/nihaixia
npx skills add https://clawhub.ai/jangviktor-web/skills/nihaixia

A second hosted path is Tencent SkillHub, where the package is named @user_ff4d9420/nihaixia-pro and the README states it is already at V2.3.1. The install instruction is itself a prompt to paste into the agent, pointing at an installation document rather than a shell command.

For a local Claude Code setup, the manual path clones the repository and copies the folder into the skills directory:

git clone https://github.com/jangviktor-web/nihaixia.git
cp -r nihaixia/ ~/.claude/skills/nihaixia/

After that, a new session that mentions 倪海厦 or 倪师 should load the skill. A first real test is the README's own cold case: ask about chills, absence of sweat and pain at the back of the neck. The expected output is 太阳伤寒, 麻黄汤证, with the note that cold binds the exterior and that 麻黄 opens the pores while 桂枝 releases the muscle layer. If your agent answers with a general discussion of colds instead, the skill did not load, and the trigger phrase or the install path is the thing to check.

The mobile route is Tencent IMA. The README walks through downloading from the IMA Skill store, then updating by sending the SkillHub install text to IMA Copilot. That update step is a copy-paste prompt, not a package manager, so the installed version depends on what the store and the agent do with it.

Dose reconciliation is the project's real engineering problem

The release history is dominated by one issue: dosage. V2.2.2 describes extracting and double-checking doses for 258 formulas from Shanghan Lun and Jingui, correcting deviations between the commonly circulated text and Ni Haixia's taught version, plus module errors, disordered quick tables and wrong line-number annotations. V2.2.4 describes rebuilding three systems of weights-and-measures conversion and fixing a rule for direct clinical conversion. V2.3.0 describes completing the full dose correction for 258 formulas and freezing the three conversion systems.

That is a meaningful signal. Classical formula doses are genuinely ambiguous: the Han-era measures do not map cleanly onto modern grams, and different teachers resolve the gap differently. A project that keeps re-correcting the same numbers across four releases is telling you the numbers are contested, not settled. The README even notes that V2.2.3 left 5 items where the source files record nothing, and that historical-source boundaries are annotated for later additions.

For an adopter, the practical consequence is that the dose in a card is a claim with a provenance, and the provenance matters more than the number. If you use this skill for anything beyond study, you need to know which conversion system a given card uses. The README says the boundaries between the systems were clarified in V2.2.4, but it does not reproduce the conversion tables in the README itself, so you have to read the module files.

Where nihaixia stops being the right tool

The clearest limitation is scope. This is a distillation of one teacher's reading of four classical texts plus his lectures, and the README frames the whole project that way. It is not a consensus reference. Where Ni Haixia's interpretation diverges from a mainstream textbook or from another school, the skill will follow him. If your question is what a hospital formulary or a modern pharmacopoeia says, this repository is the wrong source and will not tell you it is disagreeing.

The second limitation is the output-style layer. V2.3.0 explicitly fixed an academic-sounding output problem to restore 倪师味, and V2.3.1 restructured the format into spoken narration plus tables and bold marks. That is a design goal, but it means the skill optimizes for sounding like the teacher. Voice fidelity and calibrated uncertainty pull in opposite directions. A confident colloquial answer about a serious presentation is exactly the failure mode a reader should watch for.

The third is the licence. The README badge shows MulanPSL-2.0 and links to a LICENSE file, but the repository facts list the licence as unknown and the top-level entries do not include a LICENSE file. Those two things do not agree. Until that is resolved, anyone planning to redistribute the skill, bundle it into a product or mirror it on an internal server has an open question that the repository does not currently answer.

Finally, the medical boundary. The README's examples include cancer, cardiovascular and metabolic cases, and the case index is organized around those categories. A retrieval tool that returns a classical formula for a serious diagnosis is not a treatment plan, and nothing in the repository is described as validating outcomes.

How it differs from a plain TCM reference or a generic RAG setup

The obvious alternative is a general-purpose retrieval setup: put the same lecture transcripts and case files into a vector store and ask questions against them. The difference is the reasoning scaffold. A generic retrieval system returns passages ranked by similarity. nihaixia imposes an order: eight diagnostic formulas, a flow chart, a pulse-and-tongue lookup, then a seven-step model, and it forces formula cards to appear at fixed points with deduplication. The output is more predictable and easier to audit, at the cost of flexibility. If your question does not fit the six-channel frame, the scaffold is overhead.

A second alternative is a curated database of classical formulas with structured dose fields, the kind of thing a pharmacy or a research group maintains. That approach is stronger on provenance and weaker on reasoning: it will tell you the composition and the source line, but it will not walk a presentation toward a channel pattern. nihaixia is the inverse. It reasons first and cites through cards.

A third comparison is the upstream distiller. The README badges jangviktor-web/tcm-distiller at v4.4.0 as the 中医思维蒸馏器 that produced this content. If you want the pipeline rather than the output, that is the repository to read. nihaixia is the packaged artifact, and its releases are corrections to that artifact rather than changes to the extraction method.

Maintenance, upgrade cost and licence status

The last push to the default branch was on 2026-08-19, and the most recent release, v2.3.1, is dated the same day. Releases before it land in a tight cluster: V2.3.0 on 2026-08-15, V2.2.1 on 2026-08-11. That is a burst of correction work, not a steady cadence, and the repository is not archived.

Upgrade cost is real because the artifacts are versioned separately. ClawHub was at V2.2.0 while SkillHub was at V2.3.1, and the IMA store needed a manual prompt to move forward. If you install through one channel, you may not get the dose corrections from another. The README also publishes historical archives through a GitHub proxy, tagged from v1.0.0 onward, which is useful for reproducing an older answer but adds another version axis to track.

On licence, the README badge says MulanPSL-2.0 and links to LICENSE, while the repository's top-level entries do not include that file and the licence is recorded as unknown. MulanPSL-2.0 is a permissive Chinese open source licence, but a badge is not the licence text. Treat the terms as unconfirmed until a LICENSE file is present, and note that the corpus is derived from a named teacher's lectures, which raises a separate question about the underlying material that a software licence would not settle. This is not legal advice; it is a description of what the repository does and does not state.

Editorial conclusion

Adopt nihaixia if you already work inside an Agent Skill runtime and want Ni Haixia's six-channel reasoning available as a lookup layer for formulas, doses, points and cases. Do not adopt it as a diagnostic authority, and do not install it expecting a maintained library: the last push was on 2026-08-19, and the repository has no declared licence file, so redistribution terms are unresolved. Before relying on it, open SKILL.md and check which activation phrases and quick-reference blocks your runtime actually loads, then compare one formula card against the source text you trust.

Frequently asked questions

Official sources

1. Issues
2. jangviktor-web/nihaixia on GitHub
3. Project website
4. README
5. Releases