Taste Distiller品味蒸餾

Mine your rejections of AI output and distil the standard behind them into a reusable 1-5 rubric, in Markdown and in JSON for an evaluator agent從你退掉、重寫 AI 產出的實例裡挖出背後的標準,蒸餾成可重複使用的 1-5 分 rubric,同時給 Markdown 和 evaluator 用的 JSON

EN繁中Model-invoked模型叫用
Read SKILL.md on GitHub在 GitHub 看 SKILL.md

If you keep rewriting AI output the same way, you already hold a standard — you just haven't written it down, which is why every new chat starts from zero. This skill mines your actual rejections and turns the pattern behind them into a reusable rubric: Markdown for you to review and refine, JSON for an evaluator agent to grade against.

Install

npx skills add https://github.com/leoluyi/skills -g -a taste-distiller -y

To update later:

npx skills update taste-distiller

Source

What it does

It runs as a continuous conversation through four stages, none of them announced out loud.

It locates the domain you care most about, then mines three to five real moments where you rejected or rewrote AI output — asking what the AI gave you, exactly where you winced (which word, which sentence, which structural choice), what you changed it to, and what standard you were applying in one line. If you blank on examples, it prompts with friction questions: the last time the output was too empty, too slick, too templated; the last time it looked finished but missed the point; the last time you gave up explaining and rewrote it yourself.

From those it synthesizes 3-6 recurring preferences, each with a name in your own vocabulary, the failure mode in observable terms, the positive standard, and which rejection moments support it. You confirm or correct that list before anything is expanded.

Then each confirmed preference becomes a 1-5 rubric, plus a Context paragraph, plus a Reusable Instructions block you can paste into a chat tool's custom instructions — and the whole thing again as JSON for an evaluator's grading prompt.

When to use

When the same rewrite keeps happening and you want the standard captured once — as custom instructions, as an evaluator's grading prompt, as team-visible documentation, or as a self-review checklist before publishing.

When not to

Not to generate content in your style — the skill mines your taste, it doesn't perform it. Not to clean the AI-isms out of one particular draft. And not to define what an agent run should achieve; that's a goal spec, not a taste profile.

How it works

The discipline is in what it refuses.

No abstract feedback. "It felt off" and "太 AI 味" aren't accepted as answers. It pushes for the specific sentence, phrase or structural move that triggered the reaction, because a rubric built on adjectives grades nothing.

No invented preferences. Every rubric line has to be traceable to a rejection you actually described. If a rule can't be sourced, it doesn't ship.

No quality words in the tiers. Tiers describe observable behaviour — tier 1 names a specific anti-pattern ("opens with 在這個快速變化的時代"), tier 5 names a recognizable mark ("opens with a specific, time-stamped data point"). Tier 3 is the floor of shippable, tier 4 is clearly good, tier 5 is the bar. Same vocabulary axis all the way up.

No smoothing over contradictions. If your examples disagree with each other, it surfaces the conflict and asks which version you actually want, rather than averaging them into something you'd reject too.

The bar it holds itself to: another taste-savvy human should be able to grade outputs with the finished profile and reach roughly your verdicts. If it reads generic, that's a signal to mine another round of rejections.

Related skills

goal-definer points here when a task hinges on subjective quality and its Verification element needs a real standard to reference instead of a passing command. For cleaning AI-isms out of a specific piece of Traditional Chinese prose rather than defining the standard, humanizer-zh is the one that does the editing.

如果你一再用同樣的方式重寫 AI 的產出,你心裡其實已經有一套標準了——只是沒寫下來,所以每開一個新對話都要從零開始。這個技能從你真正退掉的那些實例裡挖出背後的規律,變成一份可重複使用的 rubric:Markdown 給你審閱調整,JSON 給 evaluator agent 評分。

安裝

npx skills add https://github.com/leoluyi/skills -g -a taste-distiller -y

之後更新:

npx skills update taste-distiller

原始碼

它做什麼

它以一段連續的對話進行,走過四個階段,過程中不會把階段名稱講出來。

先定位出你最在意的領域,接著挖三到五個你真的退掉或重寫過 AI 產出的實例——問 AI 原本給了什麼、你到底在哪裡皺眉(哪個字、哪個句子、哪個結構選擇)、你最後改成什麼、以及用一句話講你當時套用的標準是什麼。如果你一時想不出例子,它會用摩擦點來提示:最近一次寫得太空、太油、太像模板;最近一次看起來完成了但沒抓到重點;最近一次你懶得解釋乾脆自己重寫。

從這些實例裡它綜合出 3 到 6 個反覆出現的偏好,每一個都有一個用你自己詞彙命名的名稱、以可觀察的方式描述的失敗模式、正面的標準,以及支持它的是哪幾次退稿。這份清單要你確認或修正之後,才會展開成後面的東西。

接著每個確認過的偏好變成一份 1-5 分的 rubric,加上一段 Context、一段可以貼進聊天工具自訂指令的 Reusable Instructions,最後整份再產出一次 JSON,給 evaluator 的評分 prompt 用。

何時使用

當同樣的重寫一再發生、你想把那套標準一次寫下來的時候——可以當自訂指令、evaluator 的評分 prompt、團隊可見的文件,或發布前的自我檢查清單。

何時不要

不要用它模仿你的風格產出內容——這個技能是挖你的品味,不是表演你的品味。也不要用來清理某一份特定稿子的 AI 味。更不要用來定義一輪 agent 執行該達成什麼,那是 goal 規格,不是品味檔案。

運作方式

紀律在於它拒絕什麼。

不收抽象回饋。 「感覺怪怪的」「太 AI 味」都不算答案。它會追問到具體的那個句子、那個詞、那個結構動作,因為建立在形容詞上的 rubric 什麼也評不了。

不生造偏好。 每一條 rubric 都必須能追回到你真的描述過的某次退稿。追不到來源的規則就不會出現在成品裡。

各級不准用品質形容詞。 每一級描述的是可觀察的行為——第 1 級點名一個具體的反模式(「開頭寫『在這個快速變化的時代』」),第 5 級點名一個可辨識的標記(「開頭是一個具體、有時間點的數據」)。第 3 級是可出貨的底線,第 4 級是明顯好,第 5 級才是標準。從頭到尾用同一條詞彙軸線。

不把矛盾抹平。 如果你的例子彼此打架,它會把衝突攤出來問你到底要哪一版,而不是折衷成一個你同樣會退掉的東西。

它給自己設的標準是:另一個同樣有品味的人,拿著完成的 profile 去評分,應該能得出跟你差不多的判斷。如果讀起來很泛泛,那就是訊號——再挖一輪退稿。

相關技能

goal-definer 在任務本質是主觀品質、它的 Verification 需要一個真正的標準可以指向(而不是一道會通過的指令)時,會指到這裡。若你要的是清理某一段繁體中文稿子的 AI 味而不是定義標準,動手改稿的是 humanizer-zh