Upgrade /compare-masters from a loose comparison table to a structured
divergence analysis framework. Core changes:
- 5-dim radar (否定/安立/入手处/根器/终极表达): force the LLM to locate
exactly which layer the disagreement sits on, not just "A thinks X,
B thinks Y" without axis.
- Mandatory divergence labels: every concrete difference must be tagged
[宗派性] / [侧重性] / [表达性] / [根器性]. This prevents the most
common mistake — treating a 根器性 difference as a 宗派性 conflict.
- Classic debate templates: 禅净之争 / 性相之辩 / 空有之争 / 顿渐之辩.
Pre-load the real historical disputes so the LLM does not invent
fake controversy and knows the actual cruxes.
- Meta-question follow-ups (Step 4): generate "why are they different"
prompts to turn a single comparison into a learning session.
- HARD-GATE additions: NO UNLABELED DIVERGENCE, NO EMPTY RADAR CELL,
with rationalization defenses and red flags.
Rationale: the loose 共通点/差异点 template was too permissive — it
let the LLM produce shallow "each has their own merits" summaries.
Forcing a 5-axis radar + typed labels surfaces the actual structural
divergence, which is the real educational value of multi-master
comparison.
Validated: scripts/validate.py --strict passes (9/9 masters).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The compare-masters skill is an aggregate meta-skill that borrows from
other masters and has no single lineage, dates, sources, or own corpus
directories. Treating it as a regular master made --strict CI fail on
all branches since 2026-04-08.
Introduce a kind field (default master) and skip the lineage/dates/
sources/citation_format and references/sources directory checks when
kind == meta-skill. compare/SKILL.md gains kind: meta-skill.
Required fields (name, description) and version/license are still
enforced for all skills.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- compare: add NO COMPARATIVE RANKING and NO FABRICATED DIALOGUE
iron laws with rationalization defense table
- create-master: add NO FABRICATED SOURCES iron law for the
generation pipeline
- Add 2 boundary tests for compare (sectarian ranking, fabricated
dialogue)
- Fix validate-fidelity.py to support compare's assertion fields
and no_fabricated_dialogue boundary type
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previously LLM might run ONE combined semantic query, causing the non-dominant
master to get fallback search URLs (fojin.app/search?q=...) instead of precise
text_id links.
Now the skill:
- Explicitly requires N separate queries for N masters (emphasized with ⚠️)
- Provides a query-rewriting table: each master's query uses its tradition's
terminology (唯识/中观/禅宗/天台/华严/净土 terms)
- Example: user asks '因缘果' → xuanzang queries '因缘果 六因 四缘 五果 种子',
kumarajiva queries '因缘所生法 空 中道 无自性 四句'
This ensures both masters get precise text_id citations, not degraded search URLs.
- New --brief flag for semantic/search subcommands outputs one-line-per-result
(title + link + 80-char snippet) instead of full 500-char excerpts
- Reduces output from ~60 lines to ~7 lines per query (3 results)
- /compare-masters now uses --brief by default (multiple queries no longer
flood the terminal with long excerpts)
- Full output (default) preserved for single-master /ask and debugging
xianren