A Claude Code plugin that strips AI-flavor phrasing from the model's writing across the whole chain — dialogue, commits, PRs, docs, README, code comments — while leaving code logic untouched. Auto-injected on session start, always on, no command.
LLMs statistically favor the safest, most generic phrasing. The result is screen after screen of 「赋能 / 助力 / 携手 / 标志着……的重要里程碑」 in Chinese and serves as a testament to / nestled in the landscape in English. Low information density, tiring to read.
baihua lists these patterns as a blacklist and injects it into session context on start, so the model sees the list before every reply and every file it writes. The list has to be explicit — otherwise the model won't recognize 「赋能」 as a tic, because statistically it likes that word.
Inspired by the humanizer skill (33 patterns for de-AI-ing text), but baihua targets Chinese (with English patterns too) and auto-injects as a plugin instead of requiring a manual /humanizer call.
Applies everywhere the model writes for humans:
- Dialogue replies
- Commit messages
- PR descriptions
- Docs, README
- Code comments
The only exemption: code itself (identifiers, statements, structure) — that's not writing.
baihua stays independent of other rule systems in the user's setup; each does its own job and the model reconciles them.
Add two blocks to ~/.claude/settings.json:
"extraKnownMarketplaces": {
"baihua-marketplace": {
"source": { "source": "git", "url": "https://github.com/F2077/baihua.git" }
}
},
"enabledPlugins": {
"baihua@baihua-marketplace": true
}Takes effect on the next session. To disable, flip true to false.
- Full-inline injection. The pattern list is in context every turn. De-flavoring isn't like adding a style (where a principle suffices) — the blacklist must be visible each turn, or the model's statistical bias wins.
- Workflow untouched. Only a SessionStart hook; pure prompt injection, no tool-call interception.
- Always on, no command.
For the expanded pattern table with before/after rewrites, see references/平实要略.md.
MIT