Stats
Links
Categories
____ ____ ___ ____ ___
/ ___| / ___|/ _ \| _ \|_ _|
\___ \| | | | | | |_) || |
___) | |___| |_| | __/ | |
|____/ \____|\___/|_| |___|
Card News · Agent Teams · VS Methodology
A Claude Code plugin for researchers and creators. Turn any paper, topic, or idea into scroll-stopping social card news — without the generic AI aesthetic.
Demo
Not one LLM running in isolation. 7 expert agents with distinct roles — they debate each other, enforce VS anti-mode-collapse, and produce output that doesn't look AI-generated.
Why Scopi
Most AI content tools call a single model once and give you one output. The result looks like every other AI-generated post: uniform layout, safe word choices, template aesthetics.
Scopi is different on two axes:
1 — Agent Teams, Not a Single LLM
┌─────────────────────────────────────────────────────────────────┐
│ /scopi:generate "topic" --teams │
└──────────────────────────┬──────────────────────────────────────┘
│
┌──────▼──────┐
│ NARA │ Content strategist
│ VS: 3 arcs │ Generates 3 story directions
│ T-scored │ with entropy T-scores
└──────┬──────┘
│ user locks direction
┌────────────────▼────────────────────┐
│ Design Team │ 5 agents · parallel
│ │ real-time messaging
│ GYEOL ◄──────────────► JURI │
│ Visual Architect Ethics │
│ VS: 3 directions License & │
│ T-scored copyright │
│ ▲ ▲ │
│ │ │ │
│ BINNA ◄──────────────► MARU │
│ Copy Surgeon Empathy │
│ Hook, tone, CTA Audience │
│ scoring │
│ ▼ │
│ GANA │
│ Slide Engineer │
│ HTML/CSS · PNG · PDF │
└─────────────────────────────────────┘
│
┌──────────▼──────────┐
│ 8 PNG slides │
│ caption.txt │
│ Posting package │
└─────────────────────┘
JURI flags copyright issues in real-time. MARU scores audience resonance mid-design. BINNA negotiates copy constraints with the layout. Agents push back on each other — the output reflects that tension.
2 — VS Methodology (Anti-Mode-Collapse)
Every LLM has a "mode" — the most probable output given a prompt. That's why AI content looks the same.
VS (Verbalized Sampling) forces the agent out of its default mode by generating multiple alternatives with T-scores (temperature-analogue entropy scores). Lower T = more unconventional.
NARA generates 3 story arcs:
T=0.71 Option A · Safe timeline narrative
T=0.38 Option B · Data-led structure
T=0.17 Option C · Reverse reveal — counterintuitive hook ← recommended
The agent must propose, score, and justify each option. You choose. The design team then applies the same VS process to visual direction, typography, and layout — ensuring no two episodes look alike.
The 7 Agents
| Agent | Role | Specialty |
|---|---|---|
| NARA | Content Strategist | VS story arcs, emotional curves, series planning |
| GYEOL | Visual Architect | Free composition, theme system, anti-AI design rules |
| GANA | Slide Engineer | HTML/CSS pipeline, Puppeteer capture, 2× retina |
| BINNA | Copy Surgeon | Hook writing, tone calibration, CTA optimization |
| DARI | Audience Strategist | Platform captions, hashtag strategy, posting schedule |
| JURI | Ethics Inspector | Copyright, citation, academic integrity — read-only |
| MARU | Empathy Tester | Scroll-stop scoring, audience reaction — read-only |
JURI and MARU are read-only — they never edit files, only message other agents with flags and scores. This prevents ethics and audience concerns from being overridden in the pursuit of speed.
Output
Each run produces: