AI Facilitation Approaches

Two complementary approaches to AI-assisted facilitation have emerged from OFL research and discussions.

Approach 1: Fine-Tuning Models

Advocates: Joseph Low (Cooperative AI), Public AI researchers

Description

Train language models on facilitation data to improve their inherent facilitation capabilities at the model level.

How It Works

1. Collect labeled facilitation datasets
2. Define feedback signals (using Why-How-Who framework)
3. Fine-tune foundation models via reinforcement learning
4. Release improved models (e.g., on Hugging Face)

Benefits

• Capabilities baked into the model
• Works across different applications
• Contributes to open-source AI ecosystem
• Decentralizes AI development beyond large labs

Challenges

• Requires significant compute resources
• Needs high-quality labeled data
• Longer iteration cycles
• Changes apply broadly (less context-specific)

Best For

• Creating general-purpose facilitation capabilities
• Research into what makes facilitation effective
• Democratizing access to facilitation AI

Approach 2: Skills at Inference Time

Advocates: Harmonica team, practitioners building tools

Description

Enhance existing language models with prompts, tools, and context to enable facilitation without changing the model itself.

How It Works

1. Design prompt templates for different facilitation methods
2. Build tool integrations (memory, retrieval, external APIs)
3. Create orchestration logic for multi-step processes
4. Configure at deployment time

Benefits

• Works with any capable LLM
• Rapid iteration and customization
• Context-specific adaptation
• Lower technical barrier

Challenges

• Dependent on base model capabilities
• Prompt engineering limitations
• May require more tokens/compute per session
• Less fundamental improvement

Best For

• Building production facilitation tools
• Adapting to specific use cases
• Rapid prototyping and testing
• Organizations needing customization

These Approaches Are Complementary

A key insight from OFL discussions: these approaches attack facilitation from different angles and can work together.

┌─────────────────────────────────────────────────────┐
│                                                     │
│   Fine-tuned models ←─── Better foundation          │
│         │                                           │
│         ▼                                           │
│   Skills & prompts ←─── Method-specific behavior    │
│         │                                           │
│         ▼                                           │
│   Better facilitation outcomes                      │
│                                                     │
└─────────────────────────────────────────────────────┘

• Fine-tuning improves the foundation
• Skills/prompts provide method specificity
• Together they enable better outcomes than either alone

Practical Implications

For Tool Builders (like Harmonica)

• Build with skills/prompts for flexibility
• Design to work with multiple LLMs
• Prepare for improved base models over time
• Contribute evaluation data back to fine-tuning efforts

For Researchers (like Joseph Low)

• Use tools as testbeds for fine-tuning experiments
• Share evaluation frameworks across approaches
• Publish datasets for community benefit
• Collaborate with practitioners on what matters

For OFL

• Document patterns usable by both approaches
• Create evaluation frameworks that work across methods
• Build community bridging research and practice
• Maintain open standards for interoperability

Current State (2026)

• Fine-tuning: Early research, datasets being developed
• Skills/prompts: Production tools exist (Harmonica, others)
• Integration: Beginning collaboration between approaches
• Standards: OFL working to create shared infrastructure

References

• OFL Substack blog post on AI facilitation approaches
• Joseph Low’s Cooperative AI fellowship research
• Harmonica technical architecture discussions