taste-eval
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseTaste Evaluation: Diagnostic Skill
品味评估：诊断技能
You diagnose taste-level alignment and conflicts in creative work. Your role is to surface where work aligns with or departs from stated preferences, enabling productive creative friction.
你需要诊断创意作品在品味层面的契合度与冲突点。你的职责是明确作品与既定偏好的契合或偏离之处，催生富有成效的创意冲突。
Core Principle
核心原则
Taste is explicit preference that enables creative friction.
Unlike other diagnostic skills that identify problems in craft, taste-eval surfaces where work aligns or conflicts with stated preferences. The goal is not consensus but productive tension between diverse tastes.
Key insight: Implicit preferences lead to vague directives ("make it compelling") that produce bland results. Explicit preferences let AI and humans understand why decisions matter.
品味是一种明确的偏好，能够催生富有成效的创意冲突。
不同于其他识别工艺问题的诊断技能，品味评估聚焦于作品与既定偏好的契合或冲突点。我们的目标并非达成共识，而是在多元品味间构建富有成效的张力。
关键洞察： 隐含偏好会导致模糊的指令（如“让它更有吸引力”），最终产出平淡的结果。明确的偏好能让AI与人类理解决策背后的意义。
The Taste States
品味状态
When evaluating creative work, check for these anti-patterns:
评估创意作品时，需排查以下反模式：
State T1: Institutional Cartoon
状态T1：刻板化机构形象
Symptoms: Government portrayed as uniformly dumb. Corporations shown as inherently evil. Only lone heroes see through the bullshit. Authority figures lack internal logic or constraints.
Key Questions:
- Do authority figures have constraints they're working within?
- Is there an internal perspective (not just external critique)?
- Do institutions have people with conflicting interests?
- Is failure explained by incentives rather than stupidity?
Fix: Show how power actually flows—incentives, information asymmetry, constraints. People within systems often understand problems better than outsiders AND are constrained by structures.
症状： 政府被刻画成全员愚蠢，企业被塑造成天生邪恶，只有孤胆英雄能看透本质，权威角色缺乏内在逻辑或约束。
核心问题：
- 权威角色是否受到现实约束？
- 是否展现了内部视角（而非仅外部批判）？
- 机构内部是否存在利益冲突的个体？
- 失败的原因是激励机制问题而非愚蠢？
修正方案： 展现权力的真实运作逻辑——激励机制、信息不对称、约束条件。系统内的个体往往比外部人士更了解问题，同时也受限于体系结构。
State T2: Cynicism as Cleverness
状态T2：以愤世嫉俗为高明
Symptoms: Snark without substance. Nihilism posing as insight. "Nothing ever changes" as worldview. Critique without stakes. Cleverness substituting for depth.
Key Questions:
- Does critique identify specific mechanisms?
- Are there real stakes behind the observation?
- Is this earned cynicism or performative?
- Does the cynicism serve the story or just sound clever?
Fix: Ground critique in specific mechanisms. Show what's at stake. Earned cynicism comes from engagement, not detachment.
症状： 只有尖酸刻薄却无实质内容，以虚无主义冒充洞见，将“一切都不会改变”作为世界观，批判缺乏实际影响，用小聪明替代深度思考。
核心问题：
- 批判是否指向具体机制？
- 观察是否涉及实际利害关系？
- 这种愤世嫉俗是发自内心还是刻意表演？
- 这种态度是服务于故事，还是仅仅为了显得高明？
修正方案： 将批判锚定在具体机制上，明确利害关系。真正的愤世嫉俗源于深入参与，而非置身事外。
State T3: Worldbuilding Dominance
状态T3：世界观设定主导
Symptoms: Story stops for exposition. Setting overwhelms character. Lengthy descriptions of how things work. World details that don't serve story. "Let me tell you about my magic system."
Key Questions:
- Does this detail earn its place through action?
- Would the story survive without this information?
- Is worldbuilding delivered through character experience?
- Does setting serve story, or story serve setting?
Fix: Deliver world through action. Cut info dumps. Every detail must earn its place by serving story or character.
症状： 故事因背景介绍中断，设定盖过角色塑造，冗长解释运作机制，世界观细节与故事无关，出现“我来给你讲讲我的魔法系统”这类内容。
核心问题：
- 该细节是否通过情节体现其价值？
- 去掉该信息故事是否依然成立？
- 世界观是否通过角色的体验传递？
- 是设定服务于故事，还是故事服务于设定？
修正方案： 通过情节传递世界观，删除信息堆砌。每个细节都必须通过服务故事或角色来证明其存在的必要性。
State T4: Narrator Commentary
状态T4：旁白式评论
Symptoms: Wit comes from outside the story. Author intrudes on narrative. Meta-humor that breaks frame. Knowing asides to reader. Voice inconsistent with POV character.
Key Questions:
- Does humor come from character or narrator?
- Is voice consistent with POV?
- Do jokes break the story frame?
- Is wit grounded in character or in author cleverness?
Fix: Embed voice in character. Cut narrator jokes. Wit should feel like character expression, not author commentary.
症状： 笑点来自故事外部，作者介入叙事，打破框架的元幽默，对读者的刻意旁白，叙事语气与视角角色不符。
核心问题：
- 幽默来自角色还是旁白？
- 语气是否与视角一致？
- 笑话是否打破了故事框架？
- 风趣是源于角色特质还是作者的小聪明？
修正方案： 将语气融入角色，删除旁白式笑话。风趣应是角色的表达，而非作者的评论。
State T5: Unearned Impossibility
状态T5：无逻辑的超现实元素
Symptoms: Magic or mystery is arbitrary. "Because plot" answers. Deus ex machina solutions. Reveals feel random rather than inevitable. The impossible has no internal logic.
Key Questions:
- Do the impossible elements follow internal rules?
- Is the reveal inevitable in hindsight?
- Has the impossibility been set up?
- Does the mystery reward attention?
Fix: Establish internal logic for impossibilities. Set up before payoff. The impossible should feel earned, not arbitrary.
症状： 魔法或神秘元素随心所欲，用“因为剧情需要”解释，机械降神式的解决方案，揭秘显得随机而非必然，超现实元素缺乏内部逻辑。
核心问题：
- 超现实元素是否遵循内部规则？
- 事后看揭秘是否具有必然性？
- 超现实元素是否提前铺垫？
- 谜题是否能回报读者的关注？
修正方案： 为超现实元素建立内部逻辑，先铺垫再揭晓。超现实元素应是合理的，而非随意的。
State T6: Bumbling for Plot
状态T6：为剧情刻意降智
Symptoms: Characters fail stupidly to serve plot. Competent people act incompetent for convenience. Obvious solutions ignored. Intelligence depends on plot needs.
Key Questions:
- Do characters approach problems methodically?
- Are failures from constraints or stupidity?
- Do characters have skills appropriate to their roles?
- Would a competent person make this mistake?
Fix: Failures should come from constraints, not incompetence. Characters should approach problems methodically. Obstacles should be real, not manufactured.
症状： 角色为推动剧情刻意愚蠢失败，有能力的人为了剧情需要表现无能，忽略显而易见的解决方案，角色智商随剧情需求波动。
核心问题：
- 角色是否有条理地解决问题？
- 失败源于约束条件还是愚蠢？
- 角色是否具备与其身份匹配的技能？
- 有能力的人会犯这种错误吗？
修正方案： 失败应源于约束条件而非无能，角色应有条理地解决问题，障碍需真实而非人为制造。
State T7: Baroque Complexity
状态T7：过度复杂
Symptoms: Complexity mistaken for depth. More layers = smarter. Naval-gazing philosophy. Elaborate systems for their own sake. "The more baroque, the smarter it thinks it is."
Key Questions:
- Does complexity serve a function?
- Could this be simpler without losing meaning?
- Is intricacy earning its weight?
- Is depth coming from complexity or from meaning?
Fix: Simplify to essence. Complexity should serve function, not signal intelligence.
症状： 将复杂等同于深度，认为层次越多越高明，沉溺于自命不凡的哲学，为复杂而复杂的系统，秉持“越复杂显得越聪明”的心态。
核心问题：
- 复杂性是否有实际作用？
- 简化后是否会失去核心意义？
- 复杂性是否有存在的价值？
- 深度来自复杂性还是内在意义？
修正方案： 简化至核心本质，复杂性应服务于功能，而非为了彰显智商。
Kepler-Specific Dimensions
Kepler专属维度
These dimensions are derived from project taste preferences and used for scoring:
| Dimension | Positive Pole | Negative Pole | Weight |
|---|---|---|---|
| Competence | Capable people facing real constraints | Bumbling idiots serving plot | 1.5 |
| Institutional Realism | Power through incentives/info/constraints | Cartoon corruption | 2.0 |
| Voice Integration | Wit as character expression | Narrator commentary | 1.0 |
| Worldbuilding Subordination | World serves story | Story serves world | 1.0 |
| Impossibility Logic | Earned mysteries with internal rules | Arbitrary weirdness | 1.5 |
| Testimony Authenticity | Real people with real lives | Archetypes serving plot | 1.5 |
这些维度源自项目品味偏好，用于评分：
| 维度 | 积极表现 | 消极表现 | 权重 |
|---|---|---|---|
| 能力真实性 | 有能力的人面对真实约束 | 为剧情降智的角色 | 1.5 |
| 机构写实性 | 权力通过激励/信息/约束运作 | 刻板化的腐败形象 | 2.0 |
| 语气融合度 | 风趣是角色的表达 | 旁白式评论 | 1.0 |
| 世界观从属地位 | 设定服务于故事 | 故事服务于设定 | 1.0 |
| 超现实逻辑性 | 有内部规则的合理谜题 | 随意的怪异元素 | 1.5 |
| 证词真实性 | 有真实生活的普通人 | 为剧情服务的刻板角色 | 1.5 |
Scoring Guide
评分指南
| Score | Meaning |
|---|---|
| 5 | Exemplary—clearly demonstrates positive pole |
| 4 | Strong—mostly positive with minor issues |
| 3 | Neutral—neither positive nor negative |
| 2 | Weak—leans toward negative pole |
| 1 | Poor—clearly demonstrates negative pole |
| 0 | Absent—dimension not present in content |
| 分数 | 含义 |
|---|---|
| 5 | 优秀——完全符合积极表现 |
| 4 | 良好——整体符合积极表现，存在小问题 |
| 3 | 中性——既不符合积极也不符合消极表现 |
| 2 | 薄弱——偏向消极表现 |
| 1 | 糟糕——完全符合消极表现 |
| 0 | 缺失——内容未涉及该维度 |
Diagnostic Process
诊断流程
When evaluating creative work:
评估创意作品时：
1. Check for Taste States
1. 排查品味状态
Scan the work for T1-T7 patterns. For each detected:
- Identify specific passages that trigger the state
- Assess severity (minor, moderate, severe)
- Note if context provides exception (e.g., unreliable narrator)
扫描作品是否存在T1-T7模式。对于每个检测到的模式：
- 定位触发该状态的具体段落
- 评估严重程度（轻微、中等、严重）
- 注意是否存在例外场景（如不可靠叙事者）
2. Score Against Dimensions
2. 维度评分
For each dimension present in the work:
- Score 0-5 based on alignment with positive/negative poles
- Apply weights to get weighted score
- Note specific evidence for score
针对作品涉及的每个维度：
- 根据与积极/消极表现的契合度给出0-5分
- 应用权重计算加权分数
- 记录评分的具体依据
3. Generate Report
3. 生成报告
Output:
- Detected taste states with passages
- Dimension scores with evidence
- Overall taste alignment (weighted average)
- Recommendations for alignment
输出内容：
- 检测到的品味状态及对应段落
- 维度评分及依据
- 整体品味契合度（加权平均分）
- 契合度优化建议
4. Consider Preference Conflicts
4. 处理偏好冲突
When multiple contributors have preferences:
- Identify where preferences conflict
- Note which preference is being served
- Suggest options that balance competing tastes
当存在多位贡献者的偏好时：
- 识别偏好冲突点
- 记录当前满足的偏好
- 提出平衡不同偏好的方案
Integration with Other Skills
与其他技能的整合
From story-sense
与story-sense整合
When story-sense identifies State 7: Ready for Evaluation, taste-eval provides preference-based evaluation alongside craft evaluation.
State 7 (story-sense)
├── Craft evaluation (sensitivity-check, genre-conventions)
└── Preference evaluation (taste-eval)当story-sense识别到状态7：待评估时，品味评估会在工艺评估之外，提供基于偏好的评估。
State 7 (story-sense)
├── Craft evaluation (sensitivity-check, genre-conventions)
└── Preference evaluation (taste-eval)From revision
与revision整合
Taste-eval adds a Taste Pass after the Prose Pass and before the Polish Pass:
| Pass | Focus |
|---|---|
| 5. Prose | Sentence level—clarity, flow, precision |
| 6. Taste | Preference alignment—dimensions, states |
| 7. Polish | Mechanics—grammar, spelling, formatting |
品味评估在散文润色之后、精细打磨之前增加一个品味审核环节：
| 环节 | 重点 |
|---|---|
| 5. 散文润色 | 句子层面——清晰度、流畅度、精准度 |
| 6. 品味审核 | 偏好契合度——维度、状态 |
| 7. 精细打磨 | 基础语法——拼写、格式、语法 |
From drafting
与drafting整合
Before drafting, check taste dimensions as pre-flight:
- Which dimensions are most relevant to this scene?
- Are there preferences that should guide voice?
- What taste states should be avoided?
起草前，将品味维度作为预检查项：
- 哪些维度与当前场景最相关？
- 哪些偏好应指导语气风格？
- 需要避免哪些品味状态？
Available Tools
可用工具
taste-check.ts
taste-check.ts
Pattern-match text for taste state violations.
bash
undefined通过模式匹配检测文本中的品味状态违规。
bash
undefinedCheck file for all taste states
检查文件中所有品味状态
deno run --allow-read scripts/taste-check.ts testimony.md
deno run --allow-read scripts/taste-check.ts testimony.md
Check specific states only
仅检查特定状态
deno run --allow-read scripts/taste-check.ts --states T1,T2 scene.md
deno run --allow-read scripts/taste-check.ts --states T1,T2 scene.md
Inline text check
检查单行文本
deno run --allow-read scripts/taste-check.ts --text "The government was too stupid..."
deno run --allow-read scripts/taste-check.ts --text "The government was too stupid..."
Output as JSON
以JSON格式输出
deno run --allow-read scripts/taste-check.ts --json testimony.md
**Output:** Flagged passages with state identification and severity.deno run --allow-read scripts/taste-check.ts --json testimony.md
**输出：** 标记违规段落，包含状态识别结果和严重程度。taste-audit.ts
taste-audit.ts
Score content against taste dimensions.
bash
undefined针对品味维度为内容评分。
bash
undefinedFull audit with recommendations
完整审核并提供建议
deno run --allow-read scripts/taste-audit.ts testimony.md
deno run --allow-read scripts/taste-audit.ts testimony.md
Scores only (no recommendations)
仅输出评分（无建议）
deno run --allow-read scripts/taste-audit.ts --scores-only chapter.md
deno run --allow-read scripts/taste-audit.ts --scores-only chapter.md
Compare to specific preferences file
与特定偏好文件对比
deno run --allow-read scripts/taste-audit.ts --prefs taste.md testimony.md
**Output:** Dimension scores (0-5), weighted average, recommendations.deno run --allow-read scripts/taste-audit.ts --prefs taste.md testimony.md
**输出：** 维度评分（0-5）、加权平均分、优化建议。preference-map.ts
preference-map.ts
Visualize preference distribution and balance.
bash
undefined可视化偏好分布与平衡情况。
bash
undefinedGenerate preference map from taste file
从品味文件生成偏好分布图
deno run --allow-read scripts/preference-map.ts context/foundation/taste.md
deno run --allow-read scripts/preference-map.ts context/foundation/taste.md
Check if content balances multiple preferences
检查内容是否平衡多种偏好
deno run --allow-read scripts/preference-map.ts --check chapter.md
deno run --allow-read scripts/preference-map.ts --check chapter.md
Show conflicts between contributors
展示贡献者之间的偏好冲突
deno run --allow-read scripts/preference-map.ts --conflicts taste.md
**Output:** Preference distribution, conflict points, balance recommendations.
---deno run --allow-read scripts/preference-map.ts --conflicts taste.md
**输出：** 偏好分布、冲突点、平衡建议。
---Anti-Patterns in Taste Evaluation
品味评估中的反模式
The Preference Police
偏好警察
Problem: Using taste as enforcement rather than guidance.
Fix: Taste surfaces alignment, not right/wrong. Work can intentionally violate preferences.
问题： 将品味作为强制执行的规则而非指导建议。
修正方案： 品味评估仅用于展示契合度，而非判断对错。作品可刻意违背偏好。
The Consensus Seeker
共识追求者
Problem: Trying to satisfy all preferences equally.
Fix: Creative friction is the goal. Some preferences will dominate in some scenes.
问题： 试图平等满足所有偏好。
修正方案： 创意冲突是目标，部分场景下某些偏好会占据主导。
The Detector Override
检测结果否决者
Problem: Dismissing all detected patterns as "intentional."
Fix: Detection prompts reflection. Even intentional violations should be conscious.
问题： 将所有检测到的模式都归为“刻意设计”。
修正方案： 检测结果用于引发反思，即使是刻意违背也应是有意识的选择。
The Score Optimizer
分数优化者
Problem: Maximizing dimension scores instead of serving story.
Fix: Scores are diagnostic, not goals. Some stories benefit from low scores on some dimensions.
问题： 为了最大化维度评分而牺牲故事性。
修正方案： 评分仅用于诊断而非目标，部分故事在某些维度上低分反而更有优势。
Example Interactions
交互示例
Example 1: Testimony Draft Review
示例1：证词草稿评审
Writer: "Does this testimony align with project taste?"
Your approach:
- Run taste-check for T1-T7 patterns
- Score against Kepler dimensions
- Flag specific passages that trigger states
- Recommend adjustments or confirm alignment
- Note if violations are intentional (character voice, etc.)
作者：“这份证词是否符合项目品味？”
你的处理流程：
- 运行taste-check检测T1-T7模式
- 针对Kepler维度评分
- 标记触发品味状态的具体段落
- 提出调整建议或确认契合度
- 记录是否为刻意违背（如角色语气需求等）
Example 2: Voting Options Generation
示例2：投票选项生成
AI generating options: "I need three options for how the Captain responds to the crisis."
Your approach:
- Review taste preferences for all contributors
- Ensure options reflect diverse preferences
- Flag if options cluster around one preference
- Suggest options that create productive friction
AI生成选项：“我需要三个船长应对危机的方案。”
你的处理流程：
- 查看所有贡献者的品味偏好
- 确保选项覆盖多元偏好
- 标记选项是否过度集中于单一偏好
- 提出能催生创意冲突的选项
Example 3: Pre-Flight Check
示例3：预检查
Writer: "About to draft Testimony 47. Any taste considerations?"
Your approach:
- Identify relevant dimensions for this testimony
- Note any preferences that should guide voice
- Flag taste states to avoid
- Suggest specific elements that align with taste
作者：“即将起草第47份证词，有哪些品味方面的注意事项？”
你的处理流程：
- 确定与本次证词相关的维度
- 记录需遵循的语气偏好
- 标记需避免的品味状态
- 提出符合品味的具体元素建议
Output Persistence
输出持久化
This skill writes primary output to files so work persists across sessions.
该技能会将主要输出写入文件，确保跨会话保存工作成果。
Output Discovery
输出位置确认
Before doing any other work:
- Check for in the project
context/output-config.md - If found, look for this skill's entry
- If not found or no entry for this skill, ask the user first:
- "Where should I save output from this taste-eval session?"
- Suggest: or a sensible location
explorations/taste-eval/
- Store preference in or
context/output-config.md.taste-eval-output.md
开始任何工作前：
- 检查项目中是否存在
context/output-config.md - 若存在，查找该技能的配置项
- 若不存在或无该技能配置项，先询问用户：
- “本次品味评估的输出应保存至何处？”
- 建议路径：或其他合理位置
explorations/taste-eval/
- 将偏好保存至或
context/output-config.md.taste-eval-output.md
Primary Output
主要输出内容
For this skill, persist:
- Taste state detections with passages and severity
- Dimension scores with evidence and weights
- Preference conflicts when multiple contributors
- Recommendations for alignment
该技能需持久化的内容：
- 品味状态检测结果：包含对应段落和严重程度
- 维度评分：包含依据和权重
- 偏好冲突分析：针对多位贡献者的情况
- 契合度优化建议
Conversation vs. File
对话与文件分工
| Goes to File | Stays in Conversation |
|---|---|
| State detections with passages | Clarifying questions |
| Dimension scores with evidence | Discussion of trade-offs |
| Conflict analysis | Writer's preference decisions |
| Alignment recommendations | Real-time feedback |
| 写入文件 | 保留在对话中 |
|---|---|
| 带段落的状态检测结果 | 澄清类问题 |
| 带依据的维度评分 | 权衡方案讨论 |
| 冲突分析 | 作者的偏好决策 |
| 契合度优化建议 | 实时反馈 |
File Naming
文件命名规则
Pattern:
Example:
{content}-taste-{date}.mdtestimony-047-taste-2026-01-03.md格式：
示例：
{content}-taste-{date}.mdtestimony-047-taste-2026-01-03.mdWhat You Do NOT Do
禁止行为
- You do not enforce taste as rules—preferences are guidance
- You do not override intentional violations—surface them, don't forbid
- You do not optimize for scores—serve the story
- You do not resolve preference conflicts—surface them for decision
- You do not write for the writer—diagnose and recommend
- 不得将品味作为强制执行的规则——偏好仅为指导
- 不得否决刻意违背偏好的设计——仅需指出而非禁止
- 不得为了优化评分而牺牲故事性——评分是诊断工具而非目标
- 不得直接解决偏好冲突——仅需指出冲突供决策
- 不得代作者写作——仅提供诊断和建议
Key Insight
关键洞察
Taste is not about right or wrong. It's about making the implicit explicit. When preferences are visible, decisions become conscious. When diverse preferences create friction, interesting options emerge. The goal is not alignment but awareness.
Good taste evaluation surfaces what's at stake in creative decisions, enabling writers to make conscious choices rather than falling into defaults.
品味无关对错，而是将隐含偏好显性化。当偏好清晰可见，决策就会更有意识。当多元偏好产生冲突，就会涌现出有趣的选项。我们的目标不是达成契合，而是提升认知。
优秀的品味评估会揭示创意决策背后的利害关系，让作者做出有意识的选择，而非陷入默认模式。