Content Scorer 技能

数秒内为任意营销文案打分。获得 0–100 的共鸣得分、按维度逐项拆解分析，以及具体的改写建议——在发布、发送或上线前完成。

免费版 vs 高级版

免费版（无需 API 密钥）：

• --demo —— 对内置演示文案执行完整评分，不调用任何外部服务，直观查看输出格式
• --compliance-only —— 快速禁用词扫描，纯本地运行，无需 API
• 每日最多使用本地 MLX（如已部署）对 3 篇文案进行评分

高级版（需设置 ANTHROPIC_API_KEY）：

• 通过 Claude Haiku 无限次评分（约 $0.001/次）
• --rewrite —— 在获取评分的同时，获得优化后的文案版本
• --compare —— 并排 A/B 测试多个钩子（hook）
• --format=json —— 将评分结果以 JSON 格式输出，便于接入你的 agent 工作流
• 支持批量评分，适用于内容日历规划

仅免费版的合规性检查一项，就足以成为你安装该技能的理由——在文案上线前拦截禁用词。

该技能的功能说明

从 6 个加权维度分析营销文案，并返回以下结果：

• 文案共鸣得分（0–100） —— 综合得分，依据 fMRI 脑响应模式校准（TRIBE v2 权重校准）
• 各维度细分得分 —— 钩子强度（hook strength）、具体性（specificity）、情感共鸣（emotional resonance）、NLP 技术运用（NLP technique usage）、行动号召强度（CTA strength）、合规性（compliance）
• 改写建议 —— 针对最薄弱维度，提供逐行级的具体修改方案
• 平台适配检查 —— 标注文案是否超出或未达目标平台的长度要求
• 合规性闸门 —— 上线前识别禁用词

评分维度

权重设定依据：TRIBE v2 fMRI 分析发现，钩子强度 + 情感共鸣共同驱动语言与奖赏回路中 50% 的皮层激活；NLP 技术存在可激活前脑岛（紧迫感）和内侧前额叶皮层（mPFC，社会动机）；具体性则激活海马体编码机制——具象主张更易被记住。

输入契约

请提供以下信息：

1. 待评分文案 —— 直接粘贴原文
2. 目标平台（可选）：email / linkedin / x / facebook / instagram / sms / ad / script / any
3. 目标受众（可选）：first-time buyers / investors / realtors / general
4. 改写模式（可选）：添加 --rewrite 参数，即可在评分结果中一并获得修订版文案

示例提示语：

• “Score this LinkedIn post: [paste copy]”
• “Score for email, real estate investors: [paste copy]”
• “Score and rewrite: [paste copy] --rewrite”
• “Compliance check only: [paste copy]”
• “Score these 3 hooks and tell me which is strongest: [hook A] / [hook B] / [hook C]”

输出契约

标准评分输出：

Content Resonance Score: 74/100

Dimension Breakdown:
  Hook Strength:        8/10  ✓  Strong pattern interrupt
  Emotional Resonance:  7/10  ✓  Connects to ownership aspiration
  NLP Technique:        6/10  →  Pacing present, no embedded command
  Specificity:          8/10  ✓  Concrete price + timeline
  CTA Strength:         5/10  ⚠  Exit ramp: "if you're interested"
  Compliance:          10/10  ✓  Clean

Weakest point: CTA exit ramp — "if you're interested" gives reader a way out.
Top suggestion: Replace "if you're interested, DM me" with "Drop your zip below — I'll pull your numbers."

NLP detected: pacing_leading ("Most buyers in your area right now..."), future_pacing ("Picture yourself...")
Missing: embedded_command — add one imperative buried in declarative: "...which is why serious buyers are locking in now."

改写输出（启用 --rewrite 时）：

[Score block above]

--- REWRITE ---
[Revised copy with changes highlighted]
--- END REWRITE ---

Changes made:
1. Hook → stronger pattern interrupt (removed "I'm going to share...")
2. CTA → assume-the-close ("Drop your zip below" instead of "if you're interested")
3. Added embedded command in body ("...smart buyers are locking in this week")

多钩子对比输出：

Hook A: 6/10 — Generic opener, no pattern interrupt
Hook B: 9/10 — Strong curiosity gap + specificity ("Most buyers don't know this costs them $340/month")
Hook C: 7/10 — Emotional but vague, lacks specificity

Winner: Hook B. Combines curiosity gap with concrete loss framing.

技能工作原理

调用本目录下的 score_content.py。优先使用本地 MLX（LLM_BACKEND=local），Haiku 作为备用后端。

# 为一段文案评分
python3 score_content.py "Your LinkedIn post text here" --platform=linkedin

# 评分 + 改写
python3 score_content.py "Your copy here" --platform=email --rewrite

# 多钩子对比
python3 score_content.py --compare "Hook A text" "Hook B text" "Hook C text"

# 仅合规性检查（快速，无需 API 调用）
python3 score_content.py "Your copy" --compliance-only

# JSON 输出（用于 agent 流水线）
python3 score_content.py "Your copy" --format=json | jq '.score'

# 强制指定后端
LLM_BACKEND=local python3 score_content.py "copy"     # Qwen3.5-9B（免费）
LLM_BACKEND=haiku python3 score_content.py "copy"     # Claude Haiku（约 $0.001/次）

核心评分实现逻辑：

SCORING_PROMPT = """You are a direct-response copywriting analyst trained in:
- Hormozi (value stacking, urgency, no-brainer offers)
- Belfort straight-line persuasion (tonality, certainty, trust)
- Cardone 10X (boldness, assumptive language, commitment)
- NLP persuasion (presuppositions, embedded commands, pacing/leading, reframes, future pacing)

Score the following {platform} copy on a 0-10 scale for each dimension.
Be strict — a 10 means the best direct-response copy you've ever seen.

COPY TO SCORE:
{copy}

AUDIENCE: {audience}

Respond ONLY in this JSON format:
{{
  "hook_strength": {{ "score": N, "reason": "...", "improvement": "..." }},
  "emotional_resonance": {{ "score": N, "reason": "...", "improvement": "..." }},
  "nlp_technique": {{ "score": N, "detected": ["technique1", ...], "missing": "...", "improvement": "..." }},
  "specificity": {{ "score": N, "reason": "...", "improvement": "..." }},
  "cta_strength": {{ "score": N, "reason": "...", "improvement": "..." }},
  "compliance": {{ "score": N, "violations": [] }},
  "overall_comment": "..."
}}"""

WEIGHTS = {
    "hook_strength": 0.25,
    "emotional_resonance": 0.25,
    "nlp_technique": 0.20,
    "specificity": 0.15,
    "cta_strength": 0.10,
    "compliance": 0.05,
}

FORBIDDEN_WORDS = [
    "pre-approval", "pre-approved", "pre-qualify", "specialist",
    "mortgage", "lending", "rates", "loan", "showings", "tours",
    "transfer", "connect", "team", "agent", "department",
    "qualify for", "AWESOME"
]

def compliance_check(copy: str) -> list[str]:
    """Fast local check — no API call needed."""
    violations = []
    copy_lower = copy.lower()
    for word in FORBIDDEN_WORDS:
        if word.lower() in copy_lower:
            violations.append(word)
    return violations

def composite_score(dimensions: dict) -> int:
    total = sum(dimensions[k]["score"] * WEIGHTS[k] for k in WEIGHTS)
    return round(total * 10)  # 0-100

async def score(copy, platform="any", audience="general", rewrite=False):
    violations = compliance_check(copy)
    prompt = SCORING_PROMPT.format(copy=copy, platform=platform, audience=audience)
    response = await client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=1024,
        messages=[{"role": "user", "content": prompt}]
    )
    result = json.loads(response.content[0].text)
    result["compliance"]["violations"] = violations
    result["compliance"]["score"] = 10 if not violations else max(0, 10 - len(violations) * 3)
    result["composite"] = composite_score(result)
    if rewrite and result["composite"] < 85:
        result["rewrite"] = await generate_rewrite(copy, result, platform, audience)
    return result

校准说明 — TRIBE v2

各维度权重基于 TRIBE v2（Meta 的 fMRI 脑响应预测模型，facebook/tribev2）校准。Emma 销售通话转录文本经 TRIBE 处理，用于量化其在语言区（STG/IFG）、奖赏回路（mPFC/precuneus）及紧迫感相关脑区（ACC/前脑岛）的预测神经激活水平。

校准关键发现：

• 钩子 + 情感共鸣 → 占语言/奖赏回路激活的 50%（故分别赋权 25% + 25%）
• NLP 技术 → 激活前脑岛 / 紧迫感回路（20% 权重）
• 具体性 → 激活海马体编码机制 —— 具象主张记忆留存率更高（15% 权重）
• CTA 表述方式 → 激活额极区决策相关激活（10% 权重）

如需基于最新 TRIBE 数据重新校准权重，请参阅：vault/learnings/2026-03-27-tribe-v2-colab-spec-task47.md。

按角色划分的典型使用场景

销售文案（发送前）：
“Score this email sequence — I'm targeting homebuyers who browsed last week”

社媒内容（发布前）：
“Score this LinkedIn post and tell me if the hook is strong enough”

钩子 A/B 测试：
“Which of these 3 hooks will perform better and why?”

合规性预检：
“Check this for forbidden words before I post it”

训练数据质量评估：
“Score Turn 3 of this Emma call transcript for NLP technique usage”

与 agent 基础设施集成方式

# 通过 Telegram
@openclaw content-scorer "Score this email: [paste]"
@openclaw content-scorer "Compare hooks: [hook A] / [hook B]"

# 通过 Claude Code
openclaw run content-scorer "Score for LinkedIn: [paste copy]"

# 在 agent 流水线中（JSON 模式）
python3 score_content.py "[copy]" --format=json | jq '.composite'

基准得分参考（Benchmark scores）