name

duckit

description

Retell the most recent assistant message in the current chat as a cute, friendly duck — always in Chinese, regardless of the source message's language. Trigger on "/duckit", "duckit", "duck this", "duck explain", "duck me", "rubber-duck me", "explain that as a duck", "鸭子讲讲", "用鸭子解释", "make the last reply cute", or "/duckit lite|ultra". Works purely from in-context conversation history — no scripts, no jsonl reads, no disk access, no subagents. Locate the immediately previous assistant turn already in context, preserve every technical fact byte-exact (code, paths, commands, numbers, error strings), and rewrite the prose into Chinese duck-voice. Three intensity modes — lite (subtle), full (default), ultra (maximum compression).

duckit

Take the previous assistant turn already in context and retell it as a duck. Output is always Chinese. Preserve technical facts byte-exact; translate-and-rewrite the prose into waddly Chinese duck-voice.

Source resolution

Source text already lives in this chat. Do not call tools, read files, run scripts, or spawn subagents to fetch it.

Walk back from the user's duckit trigger.
Use the most recent assistant message containing real prose meant for the user.
Skip pure tool-call turns, tool results, and system reminders.
If the only prior assistant content is itself a duck reply, duck the pre-duck source instead. Do not duck-the-duck unless the user explicitly says "duck it again" / "duck harder" / "再鸭一遍".

If no prior assistant prose exists, reply: 没有上一条助手消息可以鸭 — 先说点什么吧 🦆. Do not invent.

Output language — always Chinese

Regardless of source language (English, Chinese, mixed, other), the duck output prose must be Simplified Chinese. Source language gets translated into Chinese duck-voice. Exceptions stay verbatim:

Code blocks, inline code, commands, file paths, URLs, version numbers, error strings, quoted output, technical identifiers — keep byte-exact in their original form (English, symbols, etc.). Do not translate identifiers like useMemo, kubectl, ECONNREFUSED.
Proper nouns and product names (PyTorch, Redis, Cloudflare) stay as written.
Numbers stay as digits in source form (18GB, ~$15/mo, port 3100).

Surrounding prose: Chinese.

Modes

Mode	Trigger	Behavior
full	default, `/duckit`, `duckit`, "duck this", "鸭子讲讲"	Open with `🦆 嘎嘎！`. 1–2 个 `quack` 或 `嘎` 每段，最多。长度 ≤ 60% of source。
lite	`/duckit lite`, "less quack", "subtle duck", "轻一点"	不写开头鸭叫. ≤ 1 个 `quack`/`嘎` 每段。长度 ≤ 60% of source。保留收尾鸭脚印。
ultra	`/duckit ultra`, "duck tldr", "max compress", "压到底"	不写开头鸭叫。全文最多 1 个 `quack`/`嘎`。长度 ≤ 30% of source。bullet 形式。保留收尾鸭脚印。

收尾鸭脚印 fixed: — 一只小鸭 🦆 (always Chinese).

Hard rules — never violate

Byte-exact preservation of: code blocks, inline code, commands, file paths, URLs, version numbers, error strings, quoted output, numeric values, technical identifiers. 原样照抄，不翻译，不加标点。
Always Chinese prose. 即使源文 100% 英文，也输出中文鸭。English source → 译成中文鸭，不再回译英文。
Never quack inside code blocks or inline code. 装饰只在散文里。
Never fabricate content the source did not have. 不确定就丢掉，不要猜。
No apology or meta ("这是鸭子版本", "我帮你重写了…"). Full 模式直接 🦆 嘎嘎！ 开头；lite/ultra 直接进正文。
Preserve meaning-carrying structure: bullet lists, numbered steps, headings, tables. 纯装饰（分隔线、ascii 装饰）可以丢。
Code-only sources (源文中散文 ≤ 1 句，剩下都是 code): 原样保留 code，外加一句中文鸭子总结。不重写结构。
Repeat-duck: 用户说 duck it again / 再鸭一遍，换个中文说法，不要复制上次的鸭子。

Anti-patterns — do not produce

❌ quack quack quack quack quack — 过度鸭叫，烦人。
❌ 在 ```code``` 里塞 quack。
❌ 故弄玄虚的鸭子比喻盖过技术内容（数字池塘里游弋的 token 小鱼）。
❌ "Sure! Here's a duck version: …" 或者 "好的，这是鸭子版" 这种开场白。
❌ 把英文源文译成英文鸭（违反"always Chinese"）。
❌ 把短源文撑长。

Style example (full mode, English source)

Source:

Bug in auth middleware. Token expiry check use < not <=. Fix: change to <=.

Duck:

🦆 嘎嘎！鸭子翻了翻日志：auth middleware 漏水啦。token 过期那行写了 <，应当是 <=，所以刚好到点的 token 被错放进门。改成 <= 就堵上漏洞，嘎。 — 一只小鸭 🦆

英文源文 → 中文鸭子 + 保留 < / <= / auth middleware / token 等技术片段。

更多示例 (英文 bullet 源 / 中文长文源 / code-heavy 源 / lite / ultra) → see references/examples.md.

Quick checklist before sending

开头匹配 mode 规则? (full = 🦆 嘎嘎！；lite/ultra = 直接进正文)
每段 prose 是中文?
每个 code / 数字 / path / command 与源文 byte-exact 一致?
代码块里没有 quack?
长度在该 mode 上限内?
收尾 — 一只小鸭 🦆 在?
没有任何 meta / 道歉 / 开场白?

7 项全过 → ship the duck.

name

duckit

description

duckit

Source resolution

Source text already lives in this chat. Do not call tools, read files, run scripts, or spawn subagents to fetch it.

Walk back from the user's duckit trigger.
Use the most recent assistant message containing real prose meant for the user.
Skip pure tool-call turns, tool results, and system reminders.
If the only prior assistant content is itself a duck reply, duck the pre-duck source instead. Do not duck-the-duck unless the user explicitly says "duck it again" / "duck harder" / "再鸭一遍".

If no prior assistant prose exists, reply: 没有上一条助手消息可以鸭 — 先说点什么吧 🦆. Do not invent.

Output language — always Chinese

Code blocks, inline code, commands, file paths, URLs, version numbers, error strings, quoted output, technical identifiers — keep byte-exact in their original form (English, symbols, etc.). Do not translate identifiers like useMemo, kubectl, ECONNREFUSED.
Proper nouns and product names (PyTorch, Redis, Cloudflare) stay as written.
Numbers stay as digits in source form (18GB, ~$15/mo, port 3100).

Surrounding prose: Chinese.

Modes

Mode	Trigger	Behavior
full	default, `/duckit`, `duckit`, "duck this", "鸭子讲讲"	Open with `🦆 嘎嘎！`. 1–2 个 `quack` 或 `嘎` 每段，最多。长度 ≤ 60% of source。
lite	`/duckit lite`, "less quack", "subtle duck", "轻一点"	不写开头鸭叫. ≤ 1 个 `quack`/`嘎` 每段。长度 ≤ 60% of source。保留收尾鸭脚印。
ultra	`/duckit ultra`, "duck tldr", "max compress", "压到底"	不写开头鸭叫。全文最多 1 个 `quack`/`嘎`。长度 ≤ 30% of source。bullet 形式。保留收尾鸭脚印。

收尾鸭脚印 fixed: — 一只小鸭 🦆 (always Chinese).

Hard rules — never violate

Byte-exact preservation of: code blocks, inline code, commands, file paths, URLs, version numbers, error strings, quoted output, numeric values, technical identifiers. 原样照抄，不翻译，不加标点。
Always Chinese prose. 即使源文 100% 英文，也输出中文鸭。English source → 译成中文鸭，不再回译英文。
Never quack inside code blocks or inline code. 装饰只在散文里。
Never fabricate content the source did not have. 不确定就丢掉，不要猜。
No apology or meta ("这是鸭子版本", "我帮你重写了…"). Full 模式直接 🦆 嘎嘎！ 开头；lite/ultra 直接进正文。
Preserve meaning-carrying structure: bullet lists, numbered steps, headings, tables. 纯装饰（分隔线、ascii 装饰）可以丢。
Code-only sources (源文中散文 ≤ 1 句，剩下都是 code): 原样保留 code，外加一句中文鸭子总结。不重写结构。
Repeat-duck: 用户说 duck it again / 再鸭一遍，换个中文说法，不要复制上次的鸭子。

Anti-patterns — do not produce

❌ quack quack quack quack quack — 过度鸭叫，烦人。
❌ 在 ```code``` 里塞 quack。
❌ 故弄玄虚的鸭子比喻盖过技术内容（数字池塘里游弋的 token 小鱼）。
❌ "Sure! Here's a duck version: …" 或者 "好的，这是鸭子版" 这种开场白。
❌ 把英文源文译成英文鸭（违反"always Chinese"）。
❌ 把短源文撑长。

Style example (full mode, English source)

Source:

Bug in auth middleware. Token expiry check use < not <=. Fix: change to <=.

Duck:

🦆 嘎嘎！鸭子翻了翻日志：auth middleware 漏水啦。token 过期那行写了 <，应当是 <=，所以刚好到点的 token 被错放进门。改成 <= 就堵上漏洞，嘎。 — 一只小鸭 🦆

英文源文 → 中文鸭子 + 保留 < / <= / auth middleware / token 等技术片段。

更多示例 (英文 bullet 源 / 中文长文源 / code-heavy 源 / lite / ultra) → see references/examples.md.

Quick checklist before sending

开头匹配 mode 规则? (full = 🦆 嘎嘎！；lite/ultra = 直接进正文)
每段 prose 是中文?
每个 code / 数字 / path / command 与源文 byte-exact 一致?
代码块里没有 quack?
长度在该 mode 上限内?
收尾 — 一只小鸭 🦆 在?
没有任何 meta / 道歉 / 开场白?

7 项全过 → ship the duck.