opencode-speak-human-tw: A Plugin That Strips the AI Smell Out of Writing
opencode-speak-human-tw is an OpenCode plugin port of Raymond Hou's speak-human-tw — a rewriting skill that removes AI-flavored phrasing, mainland-China vocabulary and half-width punctuation from Traditional Chinese copy, so the result reads like a person wrote it.
What it does
For anyone publishing Chinese content out of an AI-assisted workflow, the tell is in the rhythm: canned transitions, over-polite hedges, borrowed mainland idioms. This plugin targets exactly that. speak-human-tw reviews and rewrites text to strip the AI smell — de-genericizing phrasing, correcting Simplified-Chinese borrowings, normalizing punctuation — while leaving code, logs and config files alone. Trigger it with /speak-human-tw, pass a file with @filename.md, or just say "這段去 AI 味" and the skill auto-loads.
Mechanics
- Registers a
/speak-human-twcommand and skills on OpenCode startup. - Hourly auto-update check against npm, downloadable and restart-prompted (disable with
SPEAK_HUMAN_TW_AUTOUPDATE=0). - MIT-licensed, same as the original.
Dogfood: where it's already in use
This is not a shelf tool. The ax-mcp project runs all its Taiwanese-Chinese documentation through opencode-speak-human-tw for proofing and tone tuning — the README's "去 AI 味聲明" is exactly that workflow. On this hub, it is the designated tool for the planned Traditional-Chinese content track: the editorial standard is geek-style English for now, and when the hub starts publishing zh-TW articles, the de-AI pass on drafts is this plugin's job.
opencode plugin @jimmyyen/opencode-speak-human-tw --globalRepo: https://github.com/cawa0505/opencode-speak-human-tw — MIT.