# Master-skill: a source-grounded Buddhist AI persona framework for Claude Code and other agents

> Master-skill ships 15 pre-built Buddhist master personas with citation requirements, ethics gates and fidelity tests. It is aimed at developers and researchers who want lineage-specific answers tied to CBETA, BDRC or SuttaCentral IDs, not at casual chatbots.

**xr843/Master-skill** — FoJin-powered Buddhist AI persona framework — source-grounded, boundary-aware, fidelity-tested, runtime-ready.

- Repository: https://github.com/xr843/Master-skill
- Website: https://fojin.app/chat
- Stars: 439 · Forks: 85
- Language: Python
- License: NOASSERTION
- Published: 2026-09-10 · Updated: 2026-09-10 · Language: en
- Canonical page: https://hysenlabs.com/projects/xr843-master-skill

## The problem Master-skill targets: lineage answers without invented citations

Ask a general model about Huineng or Tsongkhapa and you get fluent prose with no way to check it. Master-skill exists to close that gap. Each persona declares a sources[] list, and the HARD-GATE rule in the README states that doctrinal assertions, practice instructions and text explanations must cite the declared sources (CBETA, BDRC, Toh, SuttaCentral, PTS or compliantly compiled teachings), with fabricated source IDs forbidden. The audience is narrow and specific: Buddhist studies researchers who need to quote a master's text, developers building a study tool on top of an agent runtime, and practitioners who want a named lineage voice rather than a generic assistant. The README is explicit that most users do not need to install anything, because fojin.app/chat exposes the same personas in a browser with a master mode dropdown.

## How the persona pipeline works: decision tree, live retrieval, two-stage review

The runtime shape follows the Anthropic Agent Skills convention: a SKILL.md file holds a decision tree and a quick reference, while references/ and sources/ load on demand so context stays small. When a persona answers, live retrieval against FoJin returns a text_id, and only then does the response carry a FoJin location link; the README states that every doctrinal assertion must be backed by a citation. If FoJin is unavailable, the offline excerpts under sources/ are the fallback. Generation runs through two independent review passes before anything is written, first doctrinal accuracy then style consistency, and a FAIL triggers automatic repair for at most two rounds. Fidelity is checked by tests/fidelity.jsonl, which holds 10 or more Q&A pairs per master (18 for the compare-masters meta-skill), and the CI dry-run validates structure on every push. Scoring a real run needs ANTHROPIC_API_KEY and is documented as a manual local or pre-release step. The README reports a first committed baseline of 59 of 84 tested passing (70 percent), with 40 percent coverage across 211 total items.

## Installing Master-skill and asking a master a first question

The README points developers at the NPX installer, which deploys a persona into the Claude Code skills directory. The npm package is master-skill and the CLI entry is bin/cli.mjs. Running list shows what is available before you install anything.

```bash
npx master-skill list
npx master-skill install master-zhiyi
```

After the install, the persona lands in ~/.claude/skills/master-<slug>/, so the example above creates ~/.claude/skills/master-zhiyi/. The README notes that the short form install zhiyi and the full form install master-zhiyi both work and resolve to the same target. Once installed, the slash command /master-zhiyi becomes available in Claude Code. The v0.6 changelog entry explains why every master command carries the master- prefix: with 50 or more skills installed, single-word slash commands blur into other skill lists, and the prefix clusters the masters under /m tab completion. The two meta-skills compare-masters and create-master deliberately keep their unprefixed names.

For a first real question, the README's own routing table is the fastest path. If you cannot sit still in meditation, it points to /虚云, /智顗 or /master-ajahn-chah for huatou, shamatha-vipashyana or mindfulness observation respectively. The documented Huineng example shows the expected output shape: a direct answer in the master's register, followed by bracketed citations such as 《六祖大师法宝坛经·坐禅品》 with a fojin.app link, and a closing note that the reader can consult the source texts on FoJin.

## Where Master-skill breaks down or is the wrong tool

The most obvious limitation is the one the project states itself: all generated dialogue is AI-synthesized content and does not represent the historical master's spoken teaching or written work. Anyone who needs a citable primary source, rather than a paraphrase that links to one, should read the linked CBETA or SuttaCentral text instead. A second constraint is the offline fallback. The sources/ directory holds key passages from core texts, but the README does not claim it covers every text a persona might need, so a FoJin outage can narrow what a persona can cite. Third, fidelity scoring is not part of the default CI path: the dry-run validates structure only, and the README describes the scored run as a manual step requiring ANTHROPIC_API_KEY, which means a fork can pass CI while its persona answers degrade. Finally, the ethics layer imposes hard refusals. The v0.5 notes record a NO_ATTAINMENT_JUDGMENT gate for Mahasi Sayadaw that forbids the AI from judging an individual's attainment, and v0.4 added no_esoteric_instruction and no_fabricated_quotes. If your use case depends on the model giving practice instructions or esoteric guidance, these gates are a deliberate obstacle, not a bug.

## Master-skill versus a plain RAG chatbot over Buddhist canon

A generic retrieval-augmented chatbot over a canon dump retrieves passages and summarizes them in one voice. Master-skill adds a persona layer on top of retrieval: the answer must match the master's register, and the two-stage review checks style consistency separately from doctrinal accuracy. That separation is the real architectural difference. A plain RAG system has no equivalent of the Layer 0 HARD-GATE, no declared sources[] per persona, and no per-persona fidelity fixture. The trade-off runs the other way too. A plain RAG pipeline is easier to point at an arbitrary corpus, while Master-skill's 15 personas are fixed and adding a new one means going through /create-master with its citation and ethics constraints. If your corpus is outside the Indian, Chinese, Tibetan and Theravada traditions the project covers, the persona scaffolding buys you nothing and the validation scripts become overhead.

## Maintenance cadence, licensing and what a fork inherits

The repository is not archived and the last push was on 2026-09-08, so the project is still moving. The release history shows v0.11.0 on 2026-08-31, v0.10.1 on 2026-07-17 and v0.10.0 on 2026-07-10, with package.json at version 0.12.2. The v0.10.1 title, the gates that never ran, is a useful signal about how this project treats its own CI: release names describe verification failures, and the npm test script chains a long list of validators including check-gate-liveness.py, validate-citation-references.py, validate-fidelity.py and validate-persona-fidelity.py. Upgrading means re-running that chain, and any persona you authored yourself has to keep satisfying the citation-reference and fidelity validators as they evolve. The code is MIT licensed, so the framework itself is permissive. That licence does not settle the status of the underlying scriptures or of any compiled teaching excerpts: the README points to ETHICS.md for the Tier A to D copyright classification, the dual-track content licensing and the takedown channel. Read ETHICS.md before publishing generated output commercially; this article is not legal advice.

## Conclusion

Adopt Master-skill if you need lineage-specific Buddhist answers that carry CBETA, BDRC, Toh, SuttaCentral or PTS identifiers, and you are willing to run the validation scripts before shipping a persona. Do not adopt it if you want a general-purpose religious chatbot, if you cannot accept AI-synthesized speech attributed to historical figures, or if you need guaranteed offline citation coverage for texts outside the sources/ directory. Before relying on it, run python3 scripts/validate.py --strict and python3 scripts/test-fidelity.py --all --dry-run against your checkout, and read ETHICS.md to confirm the copyright tiers and the takedown channel match how you intend to publish outputs.

## FAQ

### What is Master-skill?

It is a FoJin-powered Buddhist AI persona framework that ships 15 pre-built masters across Indian, Chinese, Tibetan and Theravada traditions, plus the compare-masters, master-debate and master-curriculum meta-skills. It is delivered as an AgentSkills package for Claude Code, Cursor, Codex CLI, OpenCode and Gemini CLI, and also runs in the browser at fojin.app/chat.

### How do I install Master-skill?

The README gives the NPX installer: npx master-skill install master-zhiyi deploys the persona to ~/.claude/skills/master-zhiyi/. The short form install zhiyi works too, and npx master-skill list shows the available masters first.

### Does Master-skill need an API key to run its fidelity tests?

Yes for scored runs. The README states that the CI dry-run only validates structure, while real scoring requires ANTHROPIC_API_KEY and is performed as a manual local or pre-release step.

### Is the dialogue from Master-skill the master's actual words?

No. The README states that all dialogue generated through Master-skill is AI-synthesized content and does not represent the historical master's spoken teaching or written work, and it directs readers to the cited source texts for the original wording.

## Sources

- [License: MIT](https://github.com/xr843/Master-skill/blob/main/LICENSE)
- [Project website](https://fojin.app/chat)
- [README](https://github.com/xr843/Master-skill/blob/main/README.md)
- [Releases](https://github.com/xr843/Master-skill/releases)
- [xr843/Master-skill on GitHub](https://github.com/xr843/Master-skill)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/xr843-master-skill