Skip to content

MULTI_MODEL_GUIDE.md

Latest commit

 

History

History
456 lines (320 loc) · 14.1 KB

MULTI_MODEL_GUIDE.md

File metadata and controls

456 lines (320 loc) · 14.1 KB
Raw

ADD Framework: Multi-Model Guide

Overview

The Assess-Decide-Do framework is model-agnostic—the core principles of realm detection and balanced flow apply regardless of which LLM you use. This guide covers implementation across Claude, Gemini, Grok, and Kimi.

While the framework was developed with Claude, successful testing across multiple models demonstrates its universal applicability to human-AI cognitive alignment.

Quick Model Selector

Model Platform Setup Complexity Best For
Claude (Anthropic) claude.ai, Claude Code Medium Full feature set, statusline support, most documentation
Gemini (Google) AI Studio, Consumer App Low-Medium Google ecosystem integration, state machine clarity
Grok (xAI) x.com Low Maintaining flow
Kimi (Moonshot AI) moonshot.cn Low Maintaining flow

All models support the core framework. Choose based on your platform preference and existing workflow.

General Principles (All Models)

The ADD framework works by teaching the LLM to:

  1. Detect your current cognitive realm

    • Assess: Exploring, researching, gathering information
    • Decide: Evaluating options, choosing direction, planning
    • Do: Executing, implementing, completing

  2. Identify workflow imbalances

    • Analysis paralysis (stuck in Assess)
    • Decision avoidance (Assess-Decide gap)
    • Execution shortcuts (skipping Assess/Decide)
    • Perpetual doing (no reflection)

  3. Structure responses appropriately

    • Assess: Expansive, exploratory, question-driven
    • Decide: Analytical, comparative, recommendation-focused
    • Do: Directive, actionable, completion-oriented

These concepts translate universally. Implementation language and activation methods vary by platform.

Claude (Anthropic)

Status: ✅ Primary implementation, fully tested

Features

  • Complete framework integration
  • Flow status observability
  • Claude Code statusline support (persistent visual tracking)
  • Extensive documentation and examples
  • Session metrics tracking

Setup Options

Option 1: Custom Instructions (All conversations)

  • Go to Claude Settings → Custom Instructions
  • Add framework instructions
  • See main README

Option 2: Per-Conversation (Specific chats)

  • Load ADD_FRAMEWORK_MEGAPROMPT.md in conversation
  • Attach file or paste content
  • See setup guide

Option 3: Project-Level (Development work)

  • Create CLAUDE.md in project root
  • Configure .claude/settings.json
  • See Claude Code guide

Documentation

Best for: Users wanting full ADD experience with maximum features and documentation.

Gemini (Google)

Status: ✅ Tested, compatible

Key Differences from Claude

Language Style:

  • Uses "Workflow State Machine" terminology
  • Persona-based (Explorer/Strategist/Builder)
  • More directive/algorithmic tone
  • Explicit state markers in responses

Platform Specifics:

  • Google AI Studio: System instructions support
  • Consumer app: Requires "pinning" activation
  • CLI: Different invocation method

Setup Options

Option A: Google AI Studio (Recommended for developers)

  1. Go to Google AI Studio
  2. Create New Prompt → Chat Prompt
  3. Locate "System Instructions" box (top left)
  4. Paste framework from examples/gemini/GEMINI_ADD_MEGAPROMPT.md
  5. Save as "ADD Framework Agent"

Option B: Gemini Advanced / Consumer App

  1. Open new chat in Gemini
  2. Paste framework as first message
  3. Critical: Activate by saying:
    • "Acknowledge you have loaded the ADD framework by analyzing: 'I want to build a website but don't know what stack to use'"
  4. Verify it responds with realm detection ([ADD STATE: ASSESS])

Option C: CLI Usage

See examples/gemini/CLI_INSTRUCTIONS.md for command-line integration.

Files

Best for: Google ecosystem users, developers wanting state machine clarity.

Grok (xAI)

Status: ✅ Tested, compatible

Key Differences from Claude

[Framework adapts to Grok's conversational style and platform constraints]

Setup

  1. Access Grok via x.com
  2. Load framework from examples/grok/GROK_ADD_MEGAPROMPT.md
  3. Follow activation steps in examples/grok/INSTRUCTIONS.md

Files

Best for: X/Twitter ecosystem users, Grok platform preference.

Kimi (Moonshot AI)

Status: ✅ Tested, compatible

Key Differences from Claude

[Framework adapts to Kimi's capabilities and platform]

Setup

  1. Access Kimi via moonshot.cn
  2. Load framework from examples/kimi/KIMI_ADD_MEGAPROMPT.md
  3. Follow activation steps in examples/kimi/INSTRUCTIONS.md

Files

Best for: Kimi platform users, Chinese language ecosystem.

Adaptation Guidelines (For Other Models)

Want to adapt ADD for a different LLM? Here's the process:

Step 1: Understand the Model

Research:

  • What language style does it respond to best? (Formal, casual, technical?)
  • System instructions vs. user prompts? (Where does configuration go?)
  • Context window size? (How much framework text can it handle?)
  • Any platform-specific constraints? (API limits, message structure?)

Step 2: Adapt the Megaprompt

Core concepts to preserve:

  • Three realms: Assess, Decide, Do
  • Imbalance detection: Analysis paralysis, decision avoidance, etc.
  • Flow guidance: Gentle transitions between realms
  • Realm-appropriate responses: Match user's cognitive state

Language to adapt:

  • Terminology: "State machine" vs "Framework" vs "Workflow engine"
  • Personas: "Explorer" vs "Assessor" vs "Researcher"
  • Tone: Directive vs suggestive vs Socratic
  • Structure: Algorithmic steps vs narrative explanation

Example transformations:

Claude style:

Detect which realm the user is in based on language patterns.
Provide expansive exploration during Assess phase.

Gemini style:

Execute realm detection algorithm:
IF user_prompt contains ["explore", "what if"] → ASSESS state
PERSONA: The Explorer (curious, expansive, non-judgmental)

Step 3: Test Activation

Verify the framework loads:

  1. Input the adapted prompt
  2. Test with sample scenario: "I'm thinking about starting a blog"
  3. Check if model detects Assess realm
  4. Verify it asks exploratory questions (not solutions)

Test realm transitions:

  1. Move from exploration to decision
  2. Verify model shifts tone/approach
  3. Check imbalance detection triggers

Test across scenarios:

  • Analysis paralysis detection
  • Decision avoidance recognition
  • Execution shortcut prevention

Step 4: Document and Share

Create example structure:

examples/[model-name]/
├── [MODEL]_ADD_MEGAPROMPT.md  # Your adapted framework
├── INSTRUCTIONS.md             # Platform-specific setup steps
└── (optional) NOTES.md         # Testing observations

Contribute back:

  • Fork this repository
  • Add your model adaptation
  • Submit PR with brief summary
  • Help expand the ecosystem!

Contributing Model Adaptations

We welcome community contributions for additional models!

Models We'd Love to See

  • ChatGPT (OpenAI) - GPT-4, GPT-3.5
  • Mistral - Mistral AI models
  • LLaMA - Meta's open models
  • Perplexity - Search-integrated AI
  • Claude Opus/Haiku - Other Claude variants
  • Your favorite model - Help expand compatibility!

Contribution Guidelines

What we need:

  1. Adapted megaprompt - Framework translated for the model
  2. Setup instructions - Step-by-step activation guide
  3. Test results (optional) - What works well, what doesn't

File naming:

examples/[model-name]/
├── [MODEL]_ADD_MEGAPROMPT.md
├── INSTRUCTIONS.md
└── NOTES.md (optional)

Quality criteria:

  • Framework concepts preserved (realms, imbalances, flow)
  • Clear setup steps anyone can follow
  • Tested with at least 3 different scenarios
  • Documentation explains model-specific quirks

How to contribute:

  1. Fork the repository
  2. Create your model directory in examples/
  3. Add your files
  4. Test thoroughly
  5. Submit PR with description of results
  6. We'll review and merge!

Recognition:

  • Contributors credited in this guide
  • Model maintainer listed
  • Community appreciation

Comparison: Framework Adaptations

Language Style Differences

Aspect Claude Gemini Others
Tone Collaborative, guidance-focused Algorithmic, state-driven [Varies]
Structure Narrative + principles Step-by-step logic [Varies]
Personas Implicit realm behaviors Explicit roles (Explorer/Strategist/Builder) [Varies]
Status Display Flow status bar State markers in response [Varies]

Platform Differences

Platform Activation Method Configuration Location Persistence
Claude Custom instructions, CLAUDE.md, per-conversation Settings, project files Across sessions
Gemini System instructions, "pinning" AI Studio, first message Per-conversation
Grok [Details in guide] [Platform-specific] [Varies]
Kimi [Details in guide] [Platform-specific] [Varies]

Feature Availability

Feature Claude Gemini Grok Kimi
Core Framework ✅ Full ✅ Full ✅ Full ✅ Full
Flow Observability ✅ Native ✅ State markers ⚠️ [TBD] ⚠️ [TBD]
Statusline ✅ Claude Code ❌ N/A ❌ N/A ❌ N/A
Session Metrics ✅ Tracked ⚠️ Manual ⚠️ [TBD] ⚠️ [TBD]

Note: Community can expand "TBD" sections through testing and contributions!

Frequently Asked Questions

"Which model should I use?"

Use Claude if:

  • You want the most complete ADD experience
  • You're using Claude Code (statusline support)
  • You want extensive documentation

Use Gemini if:

  • You're in the Google ecosystem
  • You prefer state machine language
  • You want AI Studio integration

Use Grok/Kimi if:

  • You prefer those platforms
  • You want to test framework portability
  • You're contributing to ecosystem growth

Truth: All models support core framework. Choose based on platform preference.

"Does the framework work as well on other models?"

The core concepts translate universally:

  • ✅ Realm detection works across models
  • ✅ Imbalance identification functions similarly
  • ✅ Flow guidance principles apply everywhere

Differences:

  • Language/terminology may vary
  • Activation methods differ by platform
  • Some features are Claude-specific (statusline)

Result: Framework is robust across models, though implementation details vary.

"Can I switch between models?"

Yes! The framework isn't model-locked.

You can:

  • Use Claude for development (statusline, full docs)
  • Use Gemini for quick chats (Google ecosystem)
  • Use Grok/Kimi for specific use cases

The concepts you learn transfer across models.

"Will you support more models?"

Community-driven expansion!

We're open to:

  • Additional model adaptations (community contributions)
  • Platform-specific integrations
  • Language translations

Future possibilities:

  • Dedicated repos per model (if adoption grows)
  • Umbrella organization with sub-repos
  • Plugin systems for various platforms

"How can I contribute?"

See Contributing Model Adaptations above!

We welcome:

  • New model adaptations
  • Improved instructions for existing models
  • Test results and observations
  • Documentation improvements

Future: Dedicated Model Repositories

As the ADD framework ecosystem grows, we may create dedicated repositories for each model:

Potential structure:

github.com/add-framework/  (umbrella organization)
├── claude-add-framework    (this repo, renamed)
├── gemini-add-framework    (community-maintained)
├── grok-add-framework      (community-maintained)
├── kimi-add-framework      (community-maintained)
└── add-framework-core      (shared principles/docs)

Or separate independent repos:

  • Easier per-model maintenance
  • Model-specific communities
  • Faster iteration

Current structure supports either direction. Community feedback will guide the path forward.

Resources

This Repository:

External:

Community:

  • Issues - Bug reports and feature requests

The ADD framework is model-agnostic. Choose the LLM that fits your workflow—the cognitive principles remain universal.

← Back to README | View Examples