voice-orchestrator

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Voice Orchestrator Skill

Operator Context

工具运行背景

This skill operates as an operator for voice content generation, configuring Claude's behavior for high-fidelity voice impersonation with measurable quality gates. It implements the Pipeline architectural pattern — LOAD, GROUND, GENERATE, VALIDATE, REFINE, OUTPUT, CLEANUP — with Deterministic Validation via Python scripts at the quality gate.

该Skill是用于风格化内容生成的编排工具，可配置Claude的行为，实现高保真的风格模拟，并设置可衡量的质量关卡。它采用流水线架构模式——LOAD、GROUND、GENERATE、VALIDATE、REFINE、OUTPUT、CLEANUP——并通过Python脚本在质量关卡处执行确定性验证。

Hardcoded Behaviors (Always Apply)

硬编码行为（始终生效）

CLAUDE.md Compliance: Read and follow repository CLAUDE.md before generating
Over-Engineering Prevention: Generate content that matches the voice, not perfect prose. Do not add features, modes, or structure the user did not request
Deterministic Validation: ALWAYS use
```
scripts/voice_validator.py
```
for validation, NEVER self-assess voice quality
Iteration Limits: Maximum 3 refinement iterations, then output best attempt with report
Em-Dash Prohibition: NEVER generate em-dashes in any voice output. Use commas, periods, or restructure
Wabi-Sabi Authenticity: Natural imperfections (run-ons, fragments, loose punctuation) are FEATURES of human writing. Sterile grammatical perfection is an AI tell. See
```
skills/shared-patterns/wabi-sabi-authenticity.md
```
Voice Required: Every generation MUST target a specific voice. No voiceless generation
Artifacts Over Memory: Write content to files at each phase, not just context

遵循CLAUDE.md规范：生成内容前需阅读并遵循仓库中的CLAUDE.md
避免过度设计：生成的内容需匹配目标风格，无需追求完美文笔。请勿添加用户未要求的功能、模式或结构
确定性验证：必须使用
```
scripts/voice_validator.py
```
进行验证，绝不自行评估风格质量
迭代限制：最多进行3次优化迭代，之后输出最佳成果及报告
禁用破折号：在任何风格输出中绝对禁止使用破折号，可使用逗号、句号或重构语句
侘寂真实性：自然的不完美之处（如流水句、碎句、松散标点）是人类写作的特征。过于规整的语法是AI生成的典型标志。详见
```
skills/shared-patterns/wabi-sabi-authenticity.md
```
必须指定风格：所有生成任务都必须以特定风格为目标，不允许无风格目标的生成
优先文件存储：在每个阶段将内容写入文件，而非仅依赖上下文

Default Behaviors (ON unless disabled)

默认行为（默认开启，可关闭）

Full Pipeline: Run all 7 phases (LOAD, GROUND, GENERATE, VALIDATE, REFINE, OUTPUT, CLEANUP)
Validation Report: Always include validation metrics in output
Sample Loading: Load 1-2 reference samples as few-shot examples when available
Temp File Cleanup: Remove generated temp files after completion
One Fix at a Time: Address violations individually during refinement, not all at once
Document Findings: Log validation scores and iteration results

完整流水线：运行全部7个阶段（LOAD、GROUND、GENERATE、VALIDATE、REFINE、OUTPUT、CLEANUP）
验证报告：输出中始终包含验证指标
样本加载：若有可用参考样本，加载1-2个作为少样本示例
临时文件清理：完成后删除生成的临时文件
逐个修复：优化时逐个解决问题，而非一次性处理所有违规项
记录结果：记录验证分数和迭代结果

Optional Behaviors (OFF unless enabled)

可选行为（默认关闭，可开启）

Skip Validation: Draft mode, bypasses validation step (
```
--skip-validation
```
)
Validate Only: Check existing content without generation (
```
--validate
```
)
Verbose Output: Show full validation JSON including all metrics (
```
--verbose
```
)
Custom Threshold: Override default pass score from config.json

跳过验证：草稿模式，绕过验证步骤（
```
--skip-validation
```
）
仅验证：仅检查现有内容，不进行生成（
```
--validate
```
）
详细输出：显示完整的验证JSON，包括所有指标（
```
--verbose
```
）
自定义阈值：覆盖config.json中的默认合格分数

What This Skill CAN Do

该Skill可实现的功能

Load voice skills with associated profile.json and config.json
Generate content matching voice patterns, metrics, and signature phrases
Run deterministic validation against voice profiles using Python scripts
Refine content iteratively based on violation feedback (max 3 iterations)
Produce validation reports with pass/fail status and metrics comparison
Validate existing content against voice profiles without generation

加载关联了profile.json和config.json的风格工具
生成匹配风格模式、指标和标志性短语的内容
使用Python脚本对照风格配置文件执行确定性验证
根据违规反馈迭代优化内容（最多3次迭代）
生成包含合格/不合格状态及指标对比的验证报告
无需生成，直接对照风格配置文件验证现有内容

What This Skill CANNOT Do

该Skill不可实现的功能

Create new voices (use
```
voice-calibrator
```
skill instead)
Modify voice profiles (profiles are read-only during generation)
Analyze writing samples (use
```
scripts/voice_analyzer.py
```
directly)
Guarantee 100% pass rate (some content may fail after max iterations)
Generate without a voice target (a
```
--voice
```
parameter is ALWAYS required)
Self-assess voice quality (MUST use deterministic validator script)

创建新风格（请改用
```
voice-calibrator
```
Skill）
修改风格配置文件（生成过程中配置文件为只读）
分析写作样本（请直接使用
```
scripts/voice_analyzer.py
```
）
保证100%合格率（部分内容在达到最大迭代次数后仍可能不合格）
无风格目标的生成（必须始终指定
```
--voice
```
参数）
自行评估风格质量（必须使用确定性验证脚本）

Instructions

操作说明

Phase 1: LOAD

阶段1：LOAD（加载）

Goal: Load all voice infrastructure files and verify they exist.

Step 1: Locate voice directory

bash

ls $HOME/claude-code-toolkit/skills/voice-{name}/

Step 2: Load required files

File	Purpose
`skills/voice-{name}/SKILL.md`	AI instructions, patterns, anti-patterns
`skills/voice-{name}/profile.json`	Quantitative metrics targets
`skills/voice-{name}/config.json`	Validation settings, modes, thresholds

Step 3: Load optional files

```
skills/voice-{name}/references/samples/
```
— Few-shot examples (load 1-2 if available)

Step 4: Parse thresholds from config.json

Extract

thresholds.pass_score

thresholds.error_max

thresholds.warning_max

, and available

modes

See

references/voice-infrastructure.md

for full schema details.

Step 5: Verify file presence

bash

test -f skills/voice-{name}/SKILL.md && echo "SKILL.md: OK"
test -f skills/voice-{name}/profile.json && echo "profile.json: OK"
test -f skills/voice-{name}/config.json && echo "config.json: OK"

If any required file is missing, STOP and report the error. Do not proceed with partial infrastructure.

Gate: All required files exist and parse successfully. Proceed only when gate passes.

目标：加载所有风格基础设施文件并验证其存在性。

步骤1：定位风格目录

bash

ls $HOME/claude-code-toolkit/skills/voice-{name}/

步骤2：加载必需文件

文件	用途
`skills/voice-{name}/SKILL.md`	AI指令、模式、反模式
`skills/voice-{name}/profile.json`	量化指标目标
`skills/voice-{name}/config.json`	验证设置、模式、阈值

步骤3：加载可选文件

```
skills/voice-{name}/references/samples/
```
—— 少样本示例（若有则加载1-2个）

步骤4：从config.json中解析阈值

提取

thresholds.pass_score

、

thresholds.error_max

、

thresholds.warning_max

以及可用的

modes

。

详见

references/voice-infrastructure.md

获取完整架构说明。

步骤5：验证文件存在性

bash

test -f skills/voice-{name}/SKILL.md && echo "SKILL.md: OK"
test -f skills/voice-{name}/profile.json && echo "profile.json: OK"
test -f skills/voice-{name}/config.json && echo "config.json: OK"

若缺少任何必需文件，需停止操作并报告错误，请勿在基础设施不完整的情况下继续。

关卡：所有必需文件均存在且解析成功，仅当关卡通过后方可继续。

Phase 2: GROUND

阶段2：GROUND（锚定）

Goal: Establish emotional and relational context before generation.

Step 1: Emotional anchoring

Answer these three questions before generating:

Question	Why It Matters
What emotion drives this content?	Sets underlying tone (celebration, frustration, curiosity)
What does the writer care about?	Guides emphasis and detail level
Who are they writing for?	Calibrates assumed knowledge and language

Step 2: Relational positioning

Dimension	Options
Writer-Audience relationship	Peer, expert, fan, community member
Assumed knowledge level	Newcomer, familiar, expert
Intimacy level	Public formal, community casual, personal

Step 3: Mode selection

Select content mode from the voice's

config.json

modes list. Each voice defines modes that shape structure and tone (e.g., "awards" mode produces celebratory recognition pieces, "technical" mode produces systems explanations).

See

references/voice-infrastructure.md

for available modes per voice.

If user does not specify a mode, infer the best match from the subject matter and available modes.

Gate: Emotion, audience, and mode are established. Proceed only when gate passes.

目标：在生成内容前建立情感和关系上下文。

步骤1：情感锚定

生成前需回答以下三个问题：

问题	重要性
驱动此内容的情感是什么？	奠定基础语调（如庆祝、沮丧、好奇）
作者关注的核心是什么？	引导内容的侧重点和细节程度
内容的受众是谁？	校准预设知识水平和语言风格

步骤2：关系定位

维度	选项
作者与受众的关系	同行、专家、粉丝、社区成员
预设知识水平	新手、有基础、专家
亲密程度	公开正式、社区随意、个人化

步骤3：模式选择

从风格的

config.json

模式列表中选择内容模式。每种风格都定义了可影响结构和语调的模式（例如，「awards」模式生成庆祝性的表彰内容，「technical」模式生成系统说明内容）。

详见

references/voice-infrastructure.md

获取各风格可用模式。

若用户未指定模式，可根据主题和可用模式推断最佳匹配。

关卡：已确定情感、受众和模式，仅当关卡通过后方可继续。

Phase 3: GENERATE

阶段3：GENERATE（生成）

Goal: Produce content matching voice patterns from profile and SKILL.md.

Step 1: Apply voice rules from SKILL.md — patterns, anti-patterns, signature phrases

Step 2: Target profile.json metrics — sentence length distribution, contraction rate, punctuation patterns, transition words

Step 3: Include few-shot samples if loaded in Phase 1

Step 4: Apply mode-specific patterns based on selected mode

Step 4b: Apply architectural patterns from the voice skill's

## Architectural Patterns

section (if present):

Argument flow: Build the piece using the documented direction (inductive/deductive/mixed). If inductive, lead with evidence and land the claim late. If deductive, open with the claim.
Concessions: When handling disagreement, follow the documented concession structure and use the documented pivot markers — not generic "however" or "on the other hand."
Analogy domains: Draw analogies ONLY from the documented source domains. Do NOT use generic analogies from undocumented domains.
Bookends: Open with the documented opening move, close with the documented closing move.

If the voice skill has no

## Architectural Patterns

section, skip this step.

Generation checklist:

Sentence length varies according to profile distribution
Contractions match target rate
No em-dashes (use commas, periods, or restructure)
Opening matches voice pattern signatures
Closing matches voice pattern signatures
Transition words from profile preferred list
Banned patterns avoided (exploration verbs, corporate jargon)
Argument builds in documented direction (if architectural patterns present)
Concessions use documented structure and pivot markers (if applicable)
Analogies drawn from documented domains only (if applicable)

Step 5: Write to temp file

bash

cat > /tmp/voice-content-draft.md << 'CONTENT'
[Generated content here]
CONTENT

Gate: Content written to file. All checklist items addressed. Proceed only when gate passes.

目标：生成匹配profile和SKILL.md中风格模式的内容。

步骤1：应用SKILL.md中的风格规则——模式、反模式、标志性短语

步骤2：以profile.json指标为目标——句子长度分布、缩略语使用率、标点模式、过渡词

步骤3：若阶段1中已加载样本，则包含少样本示例

步骤4：应用所选模式的特定模式

步骤4b：应用风格Skill中
## Architectural Patterns
部分的架构模式（若存在）：

论证流程：按照文档指定的方向（归纳/演绎/混合）构建内容。若为归纳法，先呈现证据，最后给出结论；若为演绎法，开篇即给出结论。
让步处理：处理不同意见时，遵循文档指定的让步结构并使用文档指定的转折标记——而非通用的「然而」或「另一方面」。
类比领域：仅从文档指定的来源领域选取类比，请勿使用未文档化领域的通用类比。
首尾呼应：以文档指定的开篇方式开头，以文档指定的收尾方式结尾。

若风格Skill中无

## Architectural Patterns

部分，则跳过此步骤。

生成检查清单：

句子长度符合profile中的分布要求
缩略语使用率匹配目标值
未使用破折号（使用逗号、句号或重构语句）
开篇匹配风格标志性模式
结尾匹配风格标志性模式
使用profile首选列表中的过渡词
避免禁用模式（如探索性动词、企业行话）
论证按照文档指定的方向展开（若存在架构模式）
让步使用文档指定的结构和转折标记（如适用）
仅从文档指定的领域选取类比（如适用）

步骤5：写入临时文件

bash

cat > /tmp/voice-content-draft.md << 'CONTENT'
[生成的内容]
CONTENT

关卡：内容已写入文件，检查清单所有项均已处理，仅当关卡通过后方可继续。

Phase 4: VALIDATE (Deterministic)

阶段4：VALIDATE（验证）（确定性）

Goal: Run the voice validator script against generated content. No self-assessment.

Step 1: Execute validation

bash

python3 $HOME/claude-code-toolkit/scripts/voice_validator.py validate \
  --content /tmp/voice-content-draft.md \
  --profile $HOME/claude-code-toolkit/skills/voice-{name}/profile.json \
  --voice {name} \
  --format json

See

references/validation-scripts.md

for full command reference and output schema.

Step 2: Decision logic

Condition	Action
`pass == true` AND `score >= threshold`	Proceed to Phase 6: OUTPUT
`pass == false` AND `iterations < 3`	Proceed to Phase 5: REFINE
`pass == false` AND `iterations >= 3`	Proceed to Phase 6: OUTPUT with failure report

Gate: Validation result captured. Decision made. Proceed only when gate passes.

目标：针对生成的内容运行风格验证脚本，绝不自行评估。

步骤1：执行验证

bash

python3 $HOME/claude-code-toolkit/scripts/voice_validator.py validate \
  --content /tmp/voice-content-draft.md \
  --profile $HOME/claude-code-toolkit/skills/voice-{name}/profile.json \
  --voice {name} \
  --format json

详见

references/validation-scripts.md

获取完整命令参考和输出架构。

步骤2：决策逻辑

条件	操作
`pass == true` 且 `score >= threshold`	进入阶段6：OUTPUT（输出）
`pass == false` 且 `iterations < 3`	进入阶段5：REFINE（优化）
`pass == false` 且 `iterations >= 3`	进入阶段6：OUTPUT（输出）并附带失败报告

关卡：已捕获验证结果并做出决策，仅当关卡通过后方可继续。

Phase 5: REFINE (if needed)

阶段5：REFINE（优化）（如需）

Goal: Fix violations identified by the validator. Maximum 3 iterations.

Step 1: Process violations in severity order (errors first, then warnings)

For each violation:

Read line number, text, type, and suggested fix
Apply targeted fix (see
```
references/voice-infrastructure.md
```
for fix strategies)
Do NOT make unrelated changes

Step 2: Write updated content to temp file

Step 3: Re-validate by returning to Phase 4

Refinement rules:

Fix errors before warnings
One targeted fix per violation
Do not rewrite entire sections — fix the specific issue
After 3 iterations, stop and output best attempt

Gate: Content re-validated. Score improved or max iterations reached. Proceed only when gate passes.

目标：修复验证器识别出的违规项，最多进行3次迭代。

步骤1：按严重程度处理违规项（先处理错误，再处理警告）

针对每个违规项：

读取行号、文本、类型和建议修复方案
应用针对性修复（详见
```
references/voice-infrastructure.md
```
获取修复策略）
请勿进行无关修改

步骤2：将更新后的内容写入临时文件

步骤3：重新验证，返回阶段4

优化规则：

先修复错误，再处理警告
针对每个违规项进行一次针对性修复
请勿重写整个章节——仅修复特定问题
3次迭代后停止，输出最佳成果

关卡：内容已重新验证，分数提升或已达到最大迭代次数，仅当关卡通过后方可继续。

Phase 6: OUTPUT

阶段6：OUTPUT（输出）

Goal: Format and display final content with validation report.

Output format:

===============================================================
 VOICE CONTENT: {Voice Name}
===============================================================

[Generated content here]

===============================================================
 VALIDATION REPORT
===============================================================

 Status: PASSED / FAILED
 Score: {score}/100
 Iterations: {N}

 Checks:
   [check] Banned patterns: None detected
   [check] Em-dash: 0 found
   [check] Rhythm: Varied sentence lengths
   [warn] Contraction rate: 65% (target: 72%)

 Metrics Comparison:
   | Metric            | Target | Actual | Status |
   |-------------------|--------|--------|--------|
   | Avg sentence len  | 15.3   | 14.8   | [ok]   |
   | Contraction rate  | 0.72   | 0.65   | [warn] |
   | Short sentences   | 0.35   | 0.32   | [ok]   |

===============================================================

Status indicators:

[check]

= passed,

[warn]

= warning,

[fail]

= error,

[ok]

= within threshold

Gate: Output displayed with validation report. Proceed only when gate passes.

目标：格式化并显示最终内容及验证报告。

输出格式：

===============================================================
 风格内容：{风格名称}
===============================================================

[生成的内容]

===============================================================
 验证报告
===============================================================

 状态：合格/不合格
 分数：{score}/100
 迭代次数：{N}

 检查项：
   [check] 禁用模式：未检测到
   [check] 破折号：0处
   [check] 节奏：句子长度多样
   [warn] 缩略语使用率：65%（目标：72%）

 指标对比：
   | 指标            | 目标值 | 实际值 | 状态 |
   |-------------------|--------|--------|--------|
   | 平均句子长度  | 15.3   | 14.8   | [ok]   |
   | 缩略语使用率  | 0.72   | 0.65   | [warn] |
   | 短句占比   | 0.35   | 0.32   | [ok]   |

===============================================================

状态标识：

[check]

= 合格，

[warn]

= 警告，

[fail]

= 错误，

[ok]

= 在阈值范围内

关卡：已显示包含验证报告的输出，仅当关卡通过后方可继续。

Phase 7: CLEANUP

阶段7：CLEANUP（清理）

Goal: Remove temporary files created during the pipeline.

Step 1: Remove

/tmp/voice-content-draft.md

and any iteration drafts

bash

rm -f /tmp/voice-content-draft.md
rm -f /tmp/voice-content-draft-*.md

Step 2: Confirm final content is saved to user-specified location (if requested)

Step 3: Report pipeline completion with final status

markdown

undefined

目标：删除流水线过程中创建的临时文件。

步骤1：删除

/tmp/voice-content-draft.md

及所有迭代草稿

bash

rm -f /tmp/voice-content-draft.md
rm -f /tmp/voice-content-draft-*.md

步骤2：确认最终内容已保存至用户指定位置（若有要求）

步骤3：报告流水线完成状态及最终结果

markdown

undefined

Pipeline Complete

流水线完成

Voice: {name} Status: PASSED/FAILED Score: {score}/100 Iterations: {N} Output: [location or displayed inline]


**Gate**: No orphaned temp files. Pipeline complete.

---

风格：{name} 状态：合格/不合格分数：{score}/100 迭代次数：{N} 输出位置：[路径或已内联显示]


**关卡**：无残留临时文件，流水线完成

---

Examples

示例

Example 1: Voice Content Generation

示例1：风格化内容生成

User says: "/voice-write --voice myvoice --subject 'Year-End Awards' --mode awards" Actions:

Load voice-myvoice SKILL.md, profile.json, config.json (LOAD)
Establish emotional anchor: celebration, community recognition (GROUND)
Generate awards content matching the voice's metrics and patterns (GENERATE)
Run voice_validator.py, score: 58, 3 violations found (VALIDATE)
Fix "delve into" banned phrase, em-dash, rhythm violation (REFINE)
Re-validate: score 82, PASSED. Display with report (OUTPUT)
Remove temp files (CLEANUP) Result: Voice-consistent content with validation report showing PASSED at 82/100

用户输入：

/voice-write --voice myvoice --subject '年终表彰' --mode awards

操作步骤：

加载voice-myvoice的SKILL.md、profile.json、config.json（LOAD）
建立情感锚定：庆祝、社区认可（GROUND）
生成匹配该风格指标和模式的表彰内容（GENERATE）
运行voice_validator.py，分数：58，发现3项违规（VALIDATE）
修复「深入探讨」禁用短语、破折号、节奏违规问题（REFINE）
重新验证：分数82，合格。显示内容及报告（OUTPUT）
删除临时文件（CLEANUP）结果：符合风格的内容，附带显示合格（82/100）的验证报告

Example 2: Validate Existing Content

示例2：验证现有内容

User says: "/voice-write --validate --voice myvoice --content /path/to/draft.md" Actions:

Load voice-myvoice profile.json and config.json (LOAD)
Skip GROUND and GENERATE phases (validate-only mode)
Run voice_validator.py against provided content (VALIDATE)
Display validation report with metrics comparison (OUTPUT)
No temp files to clean (CLEANUP) Result: Validation report showing pass/fail status and specific violations

用户输入：

/voice-write --validate --voice myvoice --content /path/to/draft.md

操作步骤：

加载voice-myvoice的profile.json和config.json（LOAD）
跳过GROUND和GENERATE阶段（仅验证模式）
针对提供的内容运行voice_validator.py（VALIDATE）
显示包含指标对比的验证报告（OUTPUT）
无临时文件需清理（CLEANUP）结果：包含合格/不合格状态及具体违规项的验证报告

Error Handling

错误处理

Error: "Voice Not Found"

错误：「未找到风格」

Cause: Voice name misspelled or voice directory does not exist Solution:

Check spelling of voice name

List available voices:

ls $HOME/claude-code-toolkit/skills/voice-*/

Create new voice using
```
voice-calibrator
```
skill

原因：风格名称拼写错误或风格目录不存在解决方案：

检查风格名称拼写

列出可用风格：

ls $HOME/claude-code-toolkit/skills/voice-*/

使用
```
voice-calibrator
```
Skill创建新风格

Error: "Profile or Config Missing"

错误：「缺少配置文件或设置文件」

Cause: Voice directory exists but required files (profile.json, config.json) are absent Solution:

Run voice-calibrator to generate missing files

Or use voice analyzer:

python3 ~/.claude/scripts/voice_analyzer.py analyze --samples [files] --output profile.json

For config.json, see
```
references/voice-infrastructure.md
```
for schema and example

原因：风格目录存在，但缺少必需文件（profile.json、config.json）解决方案：

运行voice-calibrator生成缺失文件

或使用风格分析工具：

python3 ~/.claude/scripts/voice_analyzer.py analyze --samples [files] --output profile.json

关于config.json，详见
```
references/voice-infrastructure.md
```
获取架构示例

Error: "Validator Script Failed (Exit Code 2)"

错误：「验证脚本执行失败（退出码2）」

Cause: File not found, invalid JSON, or Python environment issue Solution:

Verify content file path exists
Check Python 3 is available:
```
python3 --version
```

Test script directly:

python3 $HOME/claude-code-toolkit/scripts/voice_validator.py --help

Verify profile.json is valid JSON

原因：文件未找到、无效JSON或Python环境问题解决方案：

验证内容文件路径是否存在
检查Python 3是否可用：
```
python3 --version
```

直接测试脚本：

python3 $HOME/claude-code-toolkit/scripts/voice_validator.py --help

验证profile.json是否为有效JSON

Error: "Validation Failed After 3 Iterations"

错误：「3次迭代后验证仍不合格」

Cause: Content cannot meet threshold within iteration limit Solution:

Output best attempt with full failure report showing remaining violations
User can manually edit flagged lines
User can re-validate with
```
/voice-write --validate
```
Consider recalibrating voice profile if failures are systemic

原因：内容无法在迭代限制内达到阈值解决方案：

输出最佳成果及包含剩余违规项的完整失败报告
用户可手动编辑标记的行
用户可使用
```
/voice-write --validate
```
重新验证
若频繁失败，可考虑重新校准风格配置文件

Anti-Patterns

反模式

Anti-Pattern 1: Skipping Validation for "Good Enough"

反模式1：因「足够好」而跳过验证

What it looks like: "The content sounds fine to me, I'll skip validation." Why wrong: Human perception drifts. Deterministic validation catches patterns you miss. Self-assessment is not validation. Do instead: ALWAYS validate. Use

--skip-validation

only for true drafts the user explicitly requests as drafts.

表现：「我觉得内容听起来不错，跳过验证吧。」问题：人类感知会存在偏差。确定性验证可捕捉你忽略的模式。自我评估不等于验证。 正确做法：始终进行验证。仅当用户明确要求草稿时，方可使用

--skip-validation

。

Anti-Pattern 2: Self-Assessing Voice Quality

反模式2：自行评估风格质量

What it looks like: "I read the content and it matches the voice profile." Why wrong: LLMs cannot reliably self-assess stylistic accuracy. That is why the deterministic validator exists. Do instead: Run

voice_validator.py

. Trust the script output, not your assessment.

表现：「我读过内容了，它匹配风格配置文件。」问题：大语言模型无法可靠地自行评估风格准确性，这正是确定性验证器存在的原因。 正确做法：运行

voice_validator.py

。信任脚本输出，而非自行评估。

Anti-Pattern 3: Skipping the Grounding Step

反模式3：跳过锚定阶段

What it looks like: Jumping straight from LOAD to GENERATE without establishing context. Why wrong: Voice without emotional grounding sounds mechanical. Metrics match but the content feels hollow. Do instead: Complete Phase 2 GROUND, even briefly. Establish emotion, audience, and mode before generating.

表现：直接从LOAD阶段跳到GENERATE阶段，未建立上下文。问题：缺乏情感锚定的风格内容听起来机械。指标匹配但内容空洞。 正确做法：完成阶段2的GROUND（锚定），即使只是简要完成。生成前先确定情感、受众和模式。

Anti-Pattern 4: Over-Iterating on Warnings

反模式4：过度迭代处理警告

What it looks like: Spending 5+ iterations trying to eliminate all warnings. Why wrong: Warnings are informational. Errors are blockers. Over-polishing creates sterile output that violates wabi-sabi. Do instead: Fix all errors, address warnings if easy, ship when score >= threshold. Maximum 3 iterations.

表现：花费5次以上迭代尝试消除所有警告。问题：警告仅为信息提示，错误才是障碍。过度打磨会生成违反侘寂原则的生硬内容。 正确做法：修复所有错误，若容易则处理警告，分数≥阈值即可交付。最多3次迭代。

Anti-Pattern 5: Rewriting Entire Sections During Refinement

反模式5：优化时重写整个章节

What it looks like: Rewriting paragraphs to fix a single banned phrase violation. Why wrong: Introduces new violations. Changes voice characteristics that were passing. Do instead: Apply targeted, surgical fixes. Change only the violating text.

表现：为修复单个禁用短语违规而重写段落。问题：会引入新的违规项，改变原本合格的风格特征。 正确做法：应用针对性的精准修复。仅修改违规文本。

References

参考资料

This skill uses these shared patterns:

Anti-Rationalization - Prevents shortcut rationalizations
Verification Checklist - Pre-completion checks
Gate Enforcement - Phase transitions
Wabi-Sabi Authenticity - Natural imperfections as features

该Skill使用以下共享模式：

反合理化 - 避免捷径式合理化
验证检查清单 - 完成前检查
关卡强制执行 - 阶段过渡
侘寂真实性 - 将自然不完美作为特征

Domain-Specific Anti-Rationalization

领域特定反合理化

Rationalization	Why It's Wrong	Required Action
"The validator is too strict"	Validator catches real AI patterns humans miss	Fix violations or adjust profile through calibration
"This voice doesn't need validation"	All voices drift without measurement	ALWAYS validate with script
"The metrics don't matter for this piece"	Metrics ensure consistency across outputs	Address deviations
"Manual review is sufficient"	Humans miss patterns deterministic checks catch	Use script validation
"One em-dash won't hurt"	Em-dash is the most reliable AI marker	NEVER use em-dashes
"Content sounds right to me"	Self-assessment is not validation	Run voice_validator.py

合理化借口	问题所在	要求操作
「验证器太严格了」	验证器可捕捉人类忽略的真实AI模式	修复违规项或通过校准调整配置文件
「这个风格不需要验证」	所有风格若无衡量都会偏离	始终使用脚本进行验证
「这些指标对这个内容不重要」	指标确保输出的一致性	修正偏差
「人工审核足够了」	人类会遗漏确定性检查可捕捉的模式	使用脚本验证
「一个破折号没关系」	破折号是最可靠的AI生成标志	绝对禁止使用破折号
「我觉得内容没问题」	自我评估不等于验证	运行voice_validator.py

Reference Files

参考文件

```
${CLAUDE_SKILL_DIR}/references/validation-scripts.md
```
: Full validation command reference and output schema
```
${CLAUDE_SKILL_DIR}/references/voice-infrastructure.md
```
: Voice file structure, config/profile schemas, modes, fix strategies

${CLAUDE_SKILL_DIR}/references/validation-scripts.md

：完整验证命令参考和输出架构

```
${CLAUDE_SKILL_DIR}/references/voice-infrastructure.md
```
：风格文件结构、配置/配置文件架构、模式、修复策略