lovelab

LoveLab

Paste a couple's conversation. Get a structured analysis based on Gottman, attachment theory, NVC, and cognitive behavioral frameworks.

Not therapy. A mirror.

Install

# Claude Code plugin (recommended)
/plugin marketplace add thc1006/lovelab-skills
/plugin install lovelab@lovelab-marketplace

# Or test locally
git clone https://github.com/thc1006/lovelab-skills.git
claude --plugin-dir ./lovelab-skills

Skills-only (no plugin):

git clone https://github.com/thc1006/lovelab-skills.git
cp -r lovelab-skills/skills/lovelab-* ~/.claude/skills/

Cross-Platform Support

Follows the Agent Skills open standard.

Platform	Installation
Claude Code	Plugin or ~/.claude/skills/
Codex CLI	~/.codex/skills/
Gemini CLI	Reference in GEMINI.md
Cursor	.cursor/rules/<name>.mdc
GitHub Copilot	.github/copilot-instructions.md
Windsurf	.windsurfrules.md

Usage

> Analyze this couple dialogue: [paste transcript]
> Use lovelab-horsemen to detect the Four Horsemen in this transcript: [paste]

See examples/ for sample input, output, and the JSON schema for longitudinal tracking.

Skills

Skill	Analyzes	Based on
lovelab-analyze	Full analysis orchestrator	All
lovelab-horsemen	Four Horsemen + repair attempts	Gottman
lovelab-attachment	Anxiety/avoidance dimensions	ECR-R, EFT
lovelab-nvc	NVC compliance scoring	Rosenberg
lovelab-cognitive	7 cognitive distortion types	CBCT
lovelab-dynamics	Pursuer-withdrawer, power, pronouns, LSM	Gottman, Pennebaker
lovelab-report	Report with strengths, growth tracking	All + cultural layer

Longitudinal Tracking

Results follow a standardized JSON schema so you can compare across sessions:

Session 1 (baseline)  -->  Session 2  -->  Session 3
  Anxiety: 6/10            Anxiety: 5/10     Anxiety: 4/10
  Repair rate: 60%         Repair rate: 75%  Repair rate: 80%

Cultural Calibration

Includes a calibration layer for Chinese/Mandarin communication. Examples:

Pattern	Western default	Calibrated reading
Low emotional expression	Avoidant	May be culturally adaptive (Song et al. 2024)
Indirect communication	Passive-aggressive	May be face-saving with care signals
Silence	Stonewalling	May be reflective processing

Generational differences noted. Post-1995 urban Taiwanese populations often diverge from traditional norms.

What This Is Not

Not a therapist. Not a divorce predictor. Not a replacement for professional help. Not a judge of who is right.

Crisis resources: Taiwan 1925 / 1995 / 1980. International: https://findahelpline.com

Methodological Limitations

Every skill discloses its limits. The big ones:

• Gottman's >90% divorce prediction was not cross-validated. Dropped to 29% in Heyman (2001). Four Horsemen are useful for pattern recognition, not prediction.
• The 5:1 ratio is a heuristic from specific samples, not a universal standard.
• NVC has a growing but limited evidence base. No meta-analysis exists yet.
• Zero Gottman studies and only 2 EFT studies have been conducted with Chinese populations.

References

• Gottman (1994). What Predicts Divorce?
• Fraley et al. (2015). Categorical or Dimensional? JPSP 109(2).
• Rosenberg (2003). Nonviolent Communication.
• Ireland et al. (2011). Language Style Matching. Psych Science 22(1).
• Song, Chan & Ryan (2024). Expressive Suppression in East Asian Societies.
• Heyman (2001). Observation of Couple Conflicts. J Marriage & Family 63.
• Kim, Capaldi & Crosby (2007). Generalizability of Gottman. J Marriage & Family 69.

Privacy

All analysis runs locally in the LLM context window. These are plain Markdown files. No external servers, no API keys, no telemetry.

Contributing

Open an issue before submitting a PR. Help wanted:

• Language calibration layers (Korean, Japanese, Spanish, Arabic)
• Validation against trained coder annotations
• Longitudinal tracking tooling

Project Structure