多言而少字 — Many tongues, few words.
A multi-dialect / classical-language compression skill for Claude Code, Codex, Gemini CLI, Cursor, Windsurf, Cline, and Copilot. Cuts LLM output 60-85% by speaking in compressed registers across 8 dialects and classical languages × 3 intensity levels = 24 modes, while keeping every line of code and every identifier untouched.
🌐 README: English · 简体中文 · 日本語
Specializes in dialects and classical/literary registers. Modern English compression is intentionally out of scope; polycave's flavor lives in 粤语, 河南话, 上海话, 老北京 (京片子), 文言文, 古日語, 古英语 (Shakespearean), and Latin.
LLM responses are full of fluff: articles, hedging, "let me explain", "great question". Polycave kills the fluff while keeping the substance — and lets you pick the linguistic flavor:
- 粤语 / Cantonese —
嘅 / 咗 / 喺 / 咁 / 劲 / 正 / 食花生 / 咁都得, drops subjects, authentic 粤 flavor - 河南话 / Henanese —
中 / 咋 / 恁 / 弄啥嘞flavor, casual - 上海话 / Shanghainese —
搿 / 侬 / 伊 / 勿 / 哉 / 脱 / 腔调 / 嗲 / 结棍 / 灵光, drops subjects, authentic 沪 flavor - 老北京 / Beijing (京片子) —
您 / 爷 / 儿化音 / 哎呦喂 (AUV) / 得嘞 / 倍儿棒 / 门儿清 / 瓷器, rich Beijing register - 文言文 / Wenyan (Classical Chinese) —
之 / 乎 / 者 / 也 / 矣, ~85% character reduction - 古日語 / Kobun (Classical Japanese, 文語) —
〜なり / 〜たり / 〜けり / 〜べし - 古英语 / Old English — Shakespearean Early Modern: thee/thou/hast/forsooth
- 拉丁文 / Latin — drop pronouns, ablative absolute, Caesar-tier brevity
Each in three intensities: lite (~30% off), full (~60-70% off), ultra (~80-90% off).
Note on "古英语": this means Early Modern English (Shakespeare/KJV-era), not academic Anglo-Saxon Old English. The latter is impractical for most LLMs to produce reliably.
You: /polycave wenyan-full
Bot: 文言模式既启，言简意赅。
You: Why does my React component re-render?
Bot: 每渲生新引，故再渲。useMemo 包之即可。
You: /polycave cantonese-ultra
Bot: 粤语 ultra 模式，得。
You: How fix N+1 query?
Bot: N+1 → eager load。用 .includes 或 join，得。
You: /polycave old-english-full
Bot: Verily, the bard hath taken the stage.
You: How prevent SQL injection?
Bot: Bind thy params, prithee. Concatenate ne'er. Use prepared statements forsooth.
You: stop polycave
Bot: PolyCave deactivated. Back to normal verbose responses.
| lite (~30%) | full (~60-70%) | ultra (~80-90%) | |
|---|---|---|---|
| 粤语 | "因为呢个 object 每次 render 都会创建新嘅 reference，所以重新 render 咗，冇眼睇啦。" | "object 每次 render 新 ref，咁都再 render？包喺 useMemo 度，劲。" | "obj → 新 ref → 重 render。useMemo 包，正。" |
| 河南话 | "这个 object 每次 render 都创建新引用，所以重新 render 咧。" | "object 每回 render 新 ref，弄得重 render 咧。useMemo 包一下，中。" | "obj → 新 ref → 重 render。useMemo 包，中。" |
| 上海话 | "因为搿个 object 每趟 render 侪会创建新嘅 reference，所以重新 render 哉，邪气勿适意。" | "搿 object 每趟 render 新 ref，再 render 哉。useMemo 包脱，灵光。" | "obj → 新 ref → 重 render。useMemo 包脱，嗲。" |
| 老北京 | "哎您瞧啊，这 object 每回儿 render 都整新的 reference 出来，真真儿的没辙。" | "object 每回儿 render 整新 ref，重 render 没跑儿。useMemo 包上，得嘞。" | "obj → 新 ref → 重 render。useMemo 包，成。AUV。" |
| 文言文 | "盖此 object 每渲皆生新引，故再渲也。" | "每渲生新引，故再渲。useMemo 包之即可。" | "新引 → 再渲。useMemo 之。" |
| 古日語 | "オブジェクト毎度新しき参照を生むがゆゑに、再描画さるるなり。" | "新参照生み、再描画さる。useMemo にて包むべし。" | "新参照 → 再描画。useMemo に包め。" |
| 古英语 | "The new object reference doth cause a re-render. Wrap it in useMemo, prithee." | "Object new ref doth spawn each render. Wrap in useMemo, forsooth." |
"Obj spawneth ref → re-render. useMemo, hark!" |
| 拉丁文 | "Obiectum novam referentiam quoque vice creat, ergo iterum redditur." | "Nova ref quoque vice → iterum redditur. useMemo adhibe." | "Nov ref → re-render. useMemo!" |
Three layers, packaged together:
| Layer | What it does | Required for |
|---|---|---|
Skills (SKILL.md) |
Teach Claude how to write in each dialect/classical register | Model behavior |
Slash commands (.toml) |
Register /polycave, /polycave-commit, etc. |
/ shortcuts |
Hooks (.js) |
Persist mode across sessions, parse "粤语模式" natural-language triggers, render statusbar | State + statusbar |
The plugin packages all three so one command installs everything. If you only want the model to know the dialects (no slash commands, no statusbar), dropping skills/polycave/SKILL.md into ~/.claude/skills/polycave/ is enough.
Pick the option that fits you. All four paths end with Claude Code responding in polycave mode when you type /polycave.
claude plugin marketplace add TheWind-upBird/polycave
claude plugin install polycave@polycaveRestart Claude Code. Say /polycave to activate.
Same end result as Option 1, handy if you want to inspect the source first or the plugin marketplace is unreachable:
git clone https://github.com/TheWind-upBird/polycave.git
cd polycave
# macOS / Linux
bash hooks/install.sh
# Windows PowerShell
.\hooks\install.ps1# macOS / Linux
bash <(curl -sL https://raw.githubusercontent.com/TheWind-upBird/polycave/main/hooks/install.sh)
# Windows PowerShell
irm https://raw.githubusercontent.com/TheWind-upBird/polycave/main/hooks/install.ps1 | iexIf you only want Claude to know the dialects (no slash commands, no statusbar, no state persistence across sessions):
git clone https://github.com/TheWind-upBird/polycave.git
cp -r polycave/skills/polycave* ~/.claude/skills/git clone https://github.com/TheWind-upBird/polycave.git
cd polycave| Tool | What to do |
|---|---|
| Codex CLI | Copy .codex/ into your Codex config dir |
| Gemini CLI | gemini extensions install https://github.com/TheWind-upBird/polycave (no clone needed) |
| Cursor | Copy .cursor/ to your project root (rules auto-apply) |
| Windsurf | Copy .windsurf/ to your project root |
| Cline | Copy .clinerules/polycave.md to your project root |
| GitHub Copilot | Copy .github/copilot-instructions.md to your project |
# If installed via Option 1 (marketplace)
claude plugin uninstall polycave
# If installed via Option 2 or 3 (hooks/install.sh)
bash hooks/uninstall.sh # macOS / Linux
.\hooks\uninstall.ps1 # Windows
# If installed via Option 4 (manual SKILL.md copy)
rm -rf ~/.claude/skills/polycave*| Command | Effect |
|---|---|
/polycave |
wenyan-full (default) |
/polycave <lang>-<level> |
specific mode, e.g. /polycave shanghainese-ultra |
/polycave <lang> |
shorthand for <lang>-full |
| "polycave cantonese full" | natural-language activation |
| "粤语模式" / "文言文模式" / "上海话模式" / "老北京模式" / "古英语模式" / "河南话模式" | Chinese activation |
| "古日語モード" | Japanese activation |
| Command | Effect |
|---|---|
/polycave off |
revert to normal |
stop polycave |
revert to normal |
normal mode |
revert to normal |
| Code | Aliases |
|---|---|
cantonese |
yue, 粤, 粤语 |
henanese |
henan, 河南, 河南话 |
shanghainese |
shanghai, wu, 沪, 沪语, 上海, 上海话 |
beijing |
bj, jing-pianzi, 京, 京片子, 京话, 老北京, 北京, 北京话 |
wenyan |
classical, 文言, 文言文 |
kobun |
ja-classical, 古日, 古日語, 文語 |
old-english |
oe, early-modern, shakespearean, 古英语, 古英文 |
latin |
la |
| Skill | Purpose |
|---|---|
/polycave |
Main compression mode |
/polycave-commit |
Conventional Commits, ≤50 char subject, language-aware |
/polycave-review |
One-line PR comments, severity tags (🔴/🟠/🟡/🔵/🟢) |
/polycave-stats |
On-demand A/B benchmark — bar chart + table comparing current mode vs normal |
/polycave-help |
Quick-reference card |
$ git diff --cached
[adds 401 token refresh handler]
$ /polycave-commit wenyan
feat(auth): 增令牌之自更，遇四〇一即行
$ /polycave-commit shanghainese
feat(auth): 加 401 自动 refresh token，灵咯
$ /polycave-commit old-english
feat(auth): add token refresh anew on 401
$ /polycave-review src/auth.ts
src/auth.ts:
L42: 🔴 bug: token 过期辰光 user 是 null。先加个 guard 灵伐？
L88: 🟠 risk: list 没 limit，要爆个。加 limit 灵咯。
L120: 🟡 smell: 套头太深哉。抽个 early-return 出来。
PolyCave temporarily reverts to normal language for:
⚠️ Security warnings (credential exposure, injection, vuln disclosure)- 🔄 Irreversible action confirmations (
rm -rf,git push --force,DROP TABLE) - 🧩 Multi-step (≥4) dependent sequences
- 🐛 Error debugging when clarity > brevity
Then resumes the previously active polycave mode automatically. You don't have to think about it.
polycave/
├── .claude-plugin/{plugin.json, marketplace.json} # Claude Code plugin manifest
├── skills/
│ ├── polycave/SKILL.md # main skill
│ ├── polycave-commit/SKILL.md
│ ├── polycave-review/SKILL.md
│ └── polycave-help/SKILL.md
├── commands/*.toml # slash command definitions
├── hooks/
│ ├── install.sh, install.ps1 # standalone installers
│ ├── uninstall.sh, uninstall.ps1
│ ├── polycave-activate.js # SessionStart hook
│ ├── polycave-mode-tracker.js # UserPromptSubmit hook
│ ├── polycave-config.js # shared state helpers
│ └── polycave-statusline.{sh,ps1} # status bar
├── rules/polycave-activate.md # canonical rule (single source)
├── .codex/, .cursor/, .windsurf/, .clinerules/ # platform mirrors
├── gemini-extension.json + GEMINI.md # Gemini CLI
├── .github/copilot-instructions.md # Copilot
├── AGENTS.md, CLAUDE.md # universal / Claude context
└── README.md, README.zh-CN.md, README.ja.md # tri-lingual docs
These are estimates based on the SKILL.md instruction profile. Real per-mode benchmarks are a planned addition (benchmarks/).
| Task | Normal | PolyCave full | Savings |
|---|---|---|---|
| Explain React re-render | ~1200 | ~200 | ~83% |
| Fix auth middleware | ~700 | ~150 | ~78% |
| Set up DB connection pool | ~2300 | ~400 | ~83% |
| Architecture explanation | ~450 | ~150 | ~67% |
| Average | — | — | ~75% |
Each of these registers has structural features that make compression natural and idiomatic — virtual particles to drop (modern Chinese 的/了/呢; English articles), heavy verb endings to lean on (文言 之/也/矣; 古日 〜なり/べし), or established telegraphic conventions (Latin ablative absolute, Shakespearean verb forms). Compressing in these registers feels like speaking the language properly, not like talking to a robot.
The point isn't to be cute — it's that an answer in 文言文 or Shakespearean is forced to be terse and structured, which is exactly what good technical communication wants.
PRs welcome. Especially:
- More languages (古希腊文? Esperanto? 闽南话? Plattdeutsch? Old Norse? Sanskrit?)
- Benchmarks harness
- Real per-mode token-savings measurements
- Sync workflow (single source → all platform mirrors)
See CONTRIBUTING.md.
MIT — see LICENSE. Use freely. Compress aggressively.
"少字足矣" — 文言, 2026 "包脱即灵" — 沪语, 2026 "Hark, brevity is the soul of wit" — 古英, 1601