Skip to content

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

1 Commit

Folders and files

Repository files navigation

Adversarial Writing

A three-tier data journalism system that uses multi-LLM adversarial debate to produce rigorous, well-sourced longform writing.

The Three Tiers

Tier Location Purpose
Research articles/<slug>/research/ Sources, raw data, claims inventory, bias log. Ground truth for every factual claim.
Essay articles/<slug>/essay/ Longform Quarto narrative. Depth-first, unapologetic.
Interactive articles/<slug>/interactive/ Observable Framework entry point. Evidence-first, engaging.

The Adversarial Workflow

Three LLM agents critique each draft in parallel, then a synthesizer revises:

  • Research Advocate — defends factual accuracy, flags every unverified claim
  • Narrative Advocate — defends clarity and reader engagement
  • Adversary — attacks logical fallacies, unsupported causation, selection bias

Debate runs for up to N rounds until critiques converge (all agents report no major issues).

Setup

# 1. Install Python dependencies
uv sync

# 2. Configure credentials
cp .env.example .env
# Edit .env with your OpenRouter and Anthropic API keys

# 3. Start a new article
bash scripts/new-article.sh my-article-slug

# 4. Write your draft, then run the debate
uv run scripts/run-debate.py my-article-slug --max-rounds 3

Structure

adversarial-writing/
├── adversarial/          # Debate system (Python package)
│   ├── prompts/          # Agent system prompts (edit to tune behavior)
│   ├── agents.py         # Agent + persona definitions
│   ├── debate.py         # Orchestration loop
│   ├── llm.py            # OpenRouter + Anthropic abstraction
│   └── models.py         # Pydantic state models
├── articles/
│   ├── _template/        # Copy this to create a new article
│   └── <slug>/           # One folder per article
└── scripts/
    ├── new-article.sh    # Scaffold a new article
    ├── run-debate.py     # Run the adversarial debate
    └── validate-claims.py # Check claims.yaml integrity

Agent Personas

Each agent has a configurable persona that shapes its tone and emphasis. Edit articles/<slug>/debate-config.yaml to assign different personas or models per article.

Requirements

  • Python 3.12+ with uv
  • OpenRouter API key (for adversary + advocate models)
  • Anthropic API key (for synthesizer + Claude Code orchestration)
  • Quarto for essay rendering
  • Node.js + Observable Framework for interactive layer

About

No description, website, or topics provided.

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages