homophone-detector

name	homophone-detector
version	1.0
description	Detect polyphones, near-homophones, and ASR ambiguities. Two roles: (1) at runtime, extends the homophone table and the STT hallucination filter so the assistant doesn't get confused by "RAG vs LAG" or by streaming-platform watermark phrases; (2) in AI editors, gives the LLM a checklist for spotting ambiguities in user-generated knowledge bases.
trigger_words	["多音字","homophone","谐音","误识别","hallucination","filter"]
runtime	{"hook":"data-source","data":"data/homophones.toml","target":"homophones"}

homophone-detector

A two-faced skill:

• Runtime: extends interview_assistant.homophones.PHONETIC_GROUPS from data/homophones.toml, and contributes a hallucination-filter list from data/hallucinations.toml (loaded as a separate runtime: hook by the assistant's loader — see below).
• AI editor: a checklist for the model to scan a knowledge base and flag potentially-confusable terms.

Runtime contract

This skill ships TWO data files. The loader treats each as a separate runtime block via data-source:

• data/homophones.toml → merged into PHONETIC_GROUPS
• data/hallucinations.toml → appended to STTFilter.hallucinations

SKILL.md only declares the first hook in frontmatter; the loader also scans the directory for any *.toml and applies them by target field inside the file. (See data/hallucinations.toml.)

For AI editors

When the user says "scan my knowledge base for ambiguities" or you're editing their knowledge/ folder:

1. List every English/Chinese term that has a known confusable partner. Flag pairs like (RAG, LAG), (KPI, KBI), (ROI, RAI/RAW), polyphone words in Chinese (重 zhòng/chóng, 还 hái/huán, 长 cháng/zhǎng, 行 xíng/háng).
2. For each flag, propose ONE of:
   • Add the canonical form to data/homophones.toml.
   • Add a clarifying parenthetical in the knowledge file itself, e.g. "RAG（Retrieval-Augmented Generation）".
3. Never silently rewrite the user's content — present diffs and ask.

Adding entries

Format of homophones.toml:

[[entry]]
canonical = "rag"
variants  = ["lag", "rad", "rack", "raq"]
note      = "Retrieval-Augmented Generation"

[[entry]]
canonical = "重"
variants  = []
polyphone = ["zhòng", "chóng"]
example   = "重要 zhòng / 重新 chóng"

Format of hallucinations.toml:

target = "hallucinations"

[[entry]]
text = "悠悠独播剧场"
source = "Whisper streaming-platform residue"

[[entry]]
text = "感谢观看"
source = "Whisper YouTube outro residue"