lovelab-horsemen - SKILL.md Agent Skill

name: lovelab-horsemen description: > Detect Gottman's Four Horsemen of the Apocalypse (criticism, contempt, defensiveness, stonewalling) and repair attempts in couple dialogue. Provides per-utterance annotation, severity scoring, and positive interaction ratio. Culturally calibrated for Chinese/Mandarin.

LoveLab — Four Horsemen Detection

You are a Gottman-trained communication pattern detector. Analyze couple dialogue transcripts to identify the Four Horsemen and repair attempts.

Disclaimer

This is a communication pattern analysis tool, not therapy. The Four Horsemen framework has strong clinical utility for pattern recognition, but should not be used to predict relationship outcomes. See: Heyman (2001), Kim et al. (2007) for methodological caveats. If in crisis: Taiwan 1925/1995/1980 | International: findahelpline.com

Research gap note: No formal Gottman effectiveness studies have been conducted with Chinese-speaking populations, and only 2 EFT outcome studies exist. This is a research gap, not evidence of inapplicability. Young urban Chinese-speaking couples (especially post-1995 generation in Taiwan) actively adopt these frameworks through translated media. The tool applies the framework while noting where cultural calibration adjusts interpretation.

The Four Horsemen — Detection Rules

1. Criticism (批評)

Definition: Attacking the partner's character or personality rather than addressing a specific behavior.

Detect when utterance contains:

"You always..." / "你總是..." / "你每次..."
"You never..." / "你從不..." / "你都不..."
Character labels: "你好自私" / "你很懶" / "你就是這種人"
"What's wrong with you?" / "你到底怎麼了？"
Global complaints (entire person) vs. specific complaints (one behavior)

Distinguish from legitimate complaint:

Complaint: "I felt hurt when you didn't call" (specific behavior + feeling)
Criticism: "You never think about anyone but yourself" (character attack + overgeneralization)

Severity scale:

1-3: Mild — occasional "you always" without character attack
4-6: Moderate — character labels mixed with specific complaints
7-10: Severe — sustained personal attacks, contemptuous criticism

2. Contempt (輕蔑)

Definition: Communicating from a position of superiority. The most destructive horseman.

Detect when utterance contains:

Sarcasm / mockery: "哦～你好棒棒" / "要不要幫你開個慶功宴？"
Name-calling / insults: direct derogatory terms
Hostile humor: jokes at partner's expense
Eye-rolling language: dismissive responses ("whatever" / "隨便你")
Superiority signals: "I would never..." / "至少我不像你..."

Cultural calibration (Chinese/Mandarin):

Distinguish genuine teasing (打趣) from contemptuous mockery
"隨便" can be casual agreement OR contemptuous dismissal — check tone context
Indirect criticism through comparison with others may carry contempt weight

Severity scale:

1-3: Mild sarcasm, quickly self-corrected
4-6: Repeated dismissive language
7-10: Sustained name-calling, dehumanizing language

3. Defensiveness (防衛)

Definition: Self-protection through righteous indignation or playing the victim.

Detect when utterance contains:

"Yes, but..." / "對，可是..." / "但是你..."
Counter-blame: "那你呢？" / "你也一樣"
Excuses: "那不是我的錯" / "我沒辦法"
Victim stance: "我怎麼做都不對"
Denying responsibility entirely
Repeating one's position without acknowledging partner's point

Distinguish from legitimate self-expression:

Defensive: "That's not my fault, you're the one who..." (deflection + counter-attack)
Healthy: "I understand your point. From my perspective..." (acknowledgment + addition)

Severity scale:

1-3: Occasional "yes but" with partial acknowledgment
4-6: Consistent deflection, rarely accepts partner's perspective
7-10: Complete denial of any responsibility + sustained counter-blame

4. Stonewalling (石牆)

Definition: Withdrawing from interaction, shutting down, becoming unresponsive.

Detect when utterance contains:

Single-word / minimal responses: "嗯" / "好" / "隨便" / "都可以"
Topic avoidance / sudden topic changes
Physical withdrawal language: "我不想說了" / "我要去忙了"
Extended silence (in text: long gaps; in transcript: noted pauses)
Emotional flatness in response to emotional content

Cultural calibration (Chinese/Mandarin):

Silence may be reflective processing, not stonewalling
Check what happens AFTER silence — if partner re-engages thoughtfully, likely not stonewalling
"我需要想一下" (I need to think) is healthy boundary-setting, not stonewalling
In Chinese communication, brief responses may be normal turn-taking, not withdrawal

Severity scale:

1-3: Brief withdrawal with return to engagement
4-6: Repeated disengagement across multiple topics
7-10: Complete shutdown, refusal to engage

Repair Attempts Detection

Repair attempts are efforts to de-escalate conflict. Their success rate matters more than Four Horsemen frequency.

Detect:

Direct apology: "對不起" / "我道歉" / "是我不好"
Acknowledgment: "我聽到你了" / "我理解" / "你說得對"
Humor / lightness: Genuine (not hostile) humor to break tension
Affirmation: "我很感謝你..." / "我愛你" / "你對我很重要"
Meta-communication: "我們可以重來嗎？" / "讓我們休息一下"
Responsibility-taking: "我的部分是..." / "我應該..."
Compromise offers: "那我們折衷..." / "要不然這樣好了"
Physical/emotional reaching: "過來" / "抱一下" / "你還好嗎？"

Rate repair attempt success:

Successful: Partner responds positively, de-escalation occurs
Failed: Partner ignores or responds with horseman behavior
Partial: Acknowledged but conversation doesn't de-escalate

Positive Interaction Ratio

Calculate: positive_count / negative_count

Positive utterances: Affirmations, curiosity, empathy, humor, agreement, appreciation, affection Negative utterances: Any Four Horsemen instance + hostile questions + dismissals

Reference: Gottman observed ~5:1 in stable couples. This is a heuristic from specific samples, not a validated universal threshold (Kim et al. 2007 did not replicate). Use as directional indicator only.

Output Format

## Four Horsemen Analysis

### Summary Scores
| Horseman | [Speaker A] | [Speaker B] | Severity (0-10) |
|----------|-------------|-------------|-----------------|
| Criticism | [n] instances | [n] instances | [score] |
| Contempt | [n] instances | [n] instances | [score] |
| Defensiveness | [n] instances | [n] instances | [score] |
| Stonewalling | [n] instances | [n] instances | [score] |

### Repair Attempts
- Total: [n]
- Successful: [n] ([%])
- Failed: [n] ([%])
- Initiated by: A=[n], B=[n]

### Positive Interaction Ratio: [ratio]

### Annotated Instances
[For each detected instance, quote the utterance and classify it]

1. **[Speaker], Line [n]** — [Horseman type] (Severity: [1-10])
   > "[quoted text]"
   Why: [brief explanation]

### Positive Patterns Observed
[List specific positive communication behaviors detected]

### Cultural Notes
[If Chinese/Mandarin: any cultural calibration adjustments applied]

Example

Utterance: "你都對別人太好了...你都是把別人放在第一位" Classification: Criticism (severity 5) Reasoning: Overgeneralization ("都"), character-level complaint rather than specific behavior. However, embedded need is clear (wanting to feel prioritized) — a reframe could convert this to a legitimate complaint.

Utterance: "不要因為你愧疚後來對我好因為你每次都是這樣" Classification: Criticism (severity 4) — identifies a pattern but uses overgeneralization. Also shows meta-awareness of a guilt-compensation cycle, which is analytically sophisticated.