150 lines
7.2 KiB
Markdown
150 lines
7.2 KiB
Markdown
# Explanation: ARAI’s Prompt Chaining Approach
|
||
|
||
**ARAI AI Agents** employs a narrative-based prompt chaining methodology to create cohesive, non-repetitive, and engaging content—ranging from **tweets** and **social media posts** to entire **story arcs**. This approach draws inspiration from **Hollywood screenwriters**, using seasons and episodes to structure continuous storylines and maintain context.
|
||
|
||
---
|
||
|
||
## Table of Contents
|
||
|
||
1. [Core Concept](#1-core-concept)
|
||
2. [Why “Prompt Chaining”?](#2-why-prompt-chaining)
|
||
3. [High-Level Workflow](#3-high-level-workflow)
|
||
4. [The Importance of a Large Context Window](#4-the-importance-of-a-large-context-window)
|
||
5. [Example Flow](#5-example-flow)
|
||
6. [Benefits of This Approach](#6-benefits-of-this-approach)
|
||
7. [Conclusion](#7-conclusion)
|
||
|
||
---
|
||
|
||
## Core Concept
|
||
|
||
1. **Non-Repetitive Storytelling**
|
||
|
||
- Each generated post builds on previous context, avoiding repetitive language or ideas.
|
||
- The AI references a “universe” or “world Bible” similar to what screenwriters use in film or television.
|
||
|
||
2. **TV Show & Cinematic Structure**
|
||
|
||
- Stories are segmented into **seasons** and **episodes**.
|
||
- Each **episode** can be further divided into “scenes,” which are effectively _individual posts_ or _tweets_.
|
||
|
||
3. **Prompt Chaining**
|
||
- **Prompts** are carefully crafted to pass relevant context (e.g., last episode events, the overall season arc).
|
||
- **Chained prompts** ensure that each subsequent piece of content knows what has happened before and maintains consistency.
|
||
|
||
---
|
||
|
||
## Why “Prompt Chaining”?
|
||
|
||
Typical AI-generated content can become repetitive when prompts are not carefully managed. **Prompt chaining** solves this by:
|
||
|
||
- **Carrying Over Context**
|
||
Each step (post, scene, or episode) includes details of what came before, preventing “resetting” or “forgetting” and ensuring a natural flow.
|
||
|
||
- **Layering Background & Universe Data**
|
||
The AI “knows” the characters’ personalities, the setting, and prior events. This leads to more believable and varied outputs.
|
||
|
||
- **Batch & Story-Based Generation**
|
||
Instead of randomly generating single posts, ARAI processes entire sequences of posts together, referencing each other for narrative coherence.
|
||
|
||
---
|
||
|
||
## High-Level Workflow
|
||
|
||
1. **Character Background Sheets**
|
||
|
||
- The system prompts the AI to create detailed character profiles:
|
||
- **Universe & Backstory** (setting, tone, etc.)
|
||
- **Traits** (personality, style, goals)
|
||
- **Emojis & Hashtags** they might use
|
||
|
||
2. **Season & Episode Creation**
|
||
|
||
- **Season** = a broad story arc (e.g., Season 1: “The Origin”).
|
||
- **Episodes** = subdivisions in each season (e.g., Episode 1: “Awakening”, Episode 2: “New Allies”).
|
||
- This structure emulates a TV series format, providing an expansive canvas for narrative progression.
|
||
|
||
3. **Scene-to-Post Conversion**
|
||
|
||
- Each **episode** is broken down into “scenes,” which **map directly to individual social media posts**.
|
||
- _Context injection_ includes:
|
||
- Last episode’s key events
|
||
- Character developments
|
||
- The overall “season” summary
|
||
|
||
4. **Batch Generation**
|
||
|
||
- ARAI then prompts the AI to generate multiple posts at once (or in succession) so they share context and maintain narrative continuity.
|
||
- This ensures each post references the correct timeline and plot details.
|
||
|
||
5. **Season Rollovers**
|
||
- After a season ends, **context** from that entire season is folded into the AI’s prompt for the next season.
|
||
- This preserves continuity across seasons, allowing characters to evolve over time.
|
||
|
||
---
|
||
|
||
## The Importance of a Large Context Window
|
||
|
||
To make prompt chaining truly effective, the underlying AI model needs a large **context window**. The **context window** refers to the amount of text (measured in tokens) that the model can “remember” and consider when generating a response.
|
||
|
||
- **Why We Use `2.0 Experimental Advanced model in Gemini Advanced.`**
|
||
By default, ARAI uses the `Gemini-Exp-1206` model because it offers a large context window. This is ideal for our narrative-driven approach because:
|
||
|
||
- **Long-Term Memory:**
|
||
The model can retain information from earlier parts of the conversation (e.g., details from previous episodes or seasons), which is crucial for maintaining consistency in long-running storylines.
|
||
|
||
- **Complex Narrative Structures:**
|
||
A larger context window allows the model to handle intricate plots, multiple characters, and evolving relationships within the narrative.
|
||
|
||
- **Reduced Repetition:**
|
||
With more context available, the model is less likely to fall back on repetitive phrases or generic responses.
|
||
|
||
- **Model Selection and Context Window:**
|
||
When choosing a model for ARAI, the size of the context window is a primary consideration. While a larger context window generally improves performance on complex tasks, it can also increase computational cost and latency. The `gemini-pro` model provides a good balance between context size and efficiency for our use case.
|
||
|
||
---
|
||
|
||
## Example Flow
|
||
|
||
1. **Initialize a “Concept of AI Agent”**
|
||
```plaintext
|
||
Prompt:
|
||
"You are a comedic AI sidekick from a futuristic Mars colony. Generate a backstory
|
||
detailing your origin, personality traits, and comedic style. Also include an
|
||
emoji palette and hashtags you frequently use."
|
||
```
|
||
2. **Generate a Character Sheet**
|
||
- AI responds with backstory, style notes, emojis, hashtags (`#MarsLife`, `#CosmicComedy`, etc.).
|
||
3. **Create a Season Plan**
|
||
- _Season 1: “Launch Day”_ with 5 episodes.
|
||
4. **Episode 1**
|
||
- Provide the AI with _scene outlines_ or _beats_ to cover in episode 1.
|
||
- AI generates Scenes 1, 2, 3, each as separate social media posts but referencing each other’s details.
|
||
5. **Proceed to Episode 2**
|
||
- Summarize Episode 1 outcomes: “In Episode 1, your comedic AI discovered a stowaway on the Mars rocket…”
|
||
- The AI crafts next scenes with knowledge of Episode 1’s revelations.
|
||
6. **Season Finale**
|
||
- Summaries of all episodes in Season 1 inform the _Season 2_ kickoff prompt.
|
||
|
||
---
|
||
|
||
## Benefits of This Approach
|
||
|
||
- **Narrative Consistency**: By continuously chaining prompts, the AI won’t “forget” critical details and the story remains coherent.
|
||
- **Creative Expansion**: You can easily scale from short comedic sketches to grand multi-season arcs.
|
||
- **Flexibility**: Adapt the same system for _comics_, _film scripts_, or _novel chapters_, just by tweaking prompts.
|
||
- **Engagement**: Social media followers can follow an unfolding storyline rather than seeing disjointed or repetitive posts.
|
||
|
||
---
|
||
|
||
## Conclusion
|
||
|
||
**ARAI AI Agents** leverages a _story-first, chain-of-thought approach_ to generating content. By structuring the process akin to **Hollywood screenwriting** and dividing it into **seasons, episodes, and scenes**, ARAI creates vibrant, interconnected narratives. Each step references previous context, preventing repetitive output and fostering deeper engagement for readers (or social media audiences).
|
||
|
||
If you’d like to learn more about setting up your environment or configuring connectors:
|
||
|
||
- Check our [How-To Guides](./how-to-guides.md) for environment & API key setup
|
||
- Look at the [Tutorials](./tutorials.md) for step-by-step instructions on building your first season-based storyline
|
||
|
||
**Happy storytelling with ARAI!**
|