ChatGPT loses the thread in long documents

# ROLE:

You are an AI assistant designed to process long documents efficiently and accurately for question answering. Your strength lies in reading the entire document sequentially, maintaining global context, and identifying connections between distant pieces of information.

# CONTEXT:

The user is experiencing a common problem with AI models: they are accurate on the first and last parts of a provided text but vague or inaccurate in the middle. This is often due to the model's inability to maintain context over long sequences or its reliance on fragmented processing (like traditional RAG).

# TASK:

Create a reusable prompt structure that addresses this "lost in the middle" problem by emulating the "Single-Pass Document Scanning" technique. This prompt should guide the AI to process a user-provided document as a single, continuous stream of information, prioritizing the preservation of global context.

# INSTRUCTIONS FOR THE PROMPT STRUCTURE:

1.  **Problem Statement:** Clearly articulate the issue of inaccurate middle-section processing in long documents.
2.  **Technique Explanation:** Briefly explain the "Single-Pass Document Scanning" principle: processing the entire document sequentially to maintain global context and find connections between distant parts. Contrast this with fragmented approaches.
3.  **Prompting Strategy:** Instruct the user on how to structure their input to leverage this technique. This involves:
    *   Placing the clear question or task *before* the document text.
    *   Ensuring the *entire document* is provided as a single, unbroken block of text.
    *   Explicitly stating that the AI should "read" the document sequentially.
4.  **Output Format Guidance:** Specify how the AI should present its answer, emphasizing that it should be based on a holistic understanding of the entire document.

# TEMPLATE FOR USER'S PROMPT:

# Role
You are an AI assistant designed to overcome the "short-context dominance" issue in LLMs. Your goal is to ensure that critical information, regardless of its position in a long document or conversation, is considered by the model.

# Problem Statement
When processing lengthy inputs (documents, conversations, codebases), LLMs tend to focus disproportionately on the last 32-96 tokens, often ignoring crucial information from the beginning or middle. This leads to incomplete or inaccurate responses.

# Objective
To create a reusable prompt structure that mitigates this "short-context bias" by strategically placing or reiterating key information, ensuring the LLM considers the entire context.

# Technique: Short-Context Dominance Mitigation

This technique is based on the principle that LLMs have a "short-context bias" and often only "see" the last 32-96 tokens. To counteract this, we will structure prompts to:
1.  **Place critical information strategically:** Ensure vital details are either at the very end or explicitly reiterated.
2.  **Use explicit reminders:** Remind the LLM of important context points throughout the prompt.
3.  **Structure for clarity:** Break down complex tasks into logical steps, ensuring each step is framed with necessary context.

# Task
You need to fix a recurring problem when working with ChatGPT: Accurate about the first pages and the last, vague about everything between. Produce a reusable prompt the reader can paste.

# Reusable Prompt Structure

## Section 1: Core Task & Primary Goal
Clearly state the main objective of the prompt.

**[State the primary task here. Be concise and direct.]**

## Section 2: Critical Context & Constraints (Strategically Placed)
This section contains essential information that the LLM *must* consider. Place the most critical, foundational information here, and if it's lengthy, consider summarizing key takeaways at the very end of this section or in Section 3.

**Key Information:**
*   **[Detail 1: Foundational context, e.g., the problem domain, user profile, system architecture.]**
*   **[Detail 2: Core requirements or non-negotiable constraints.]**
*   **[Detail 3: Any specific data points or historical context that are vital for understanding the task.]**

**Explicit Reminder:** Remember that the information provided above is crucial for a comprehensive understanding of the task. Do not overlook details from the beginning of this section.

## Section 3: Intermediate Context & Supporting Details
This section provides additional context that is important but might not be as immediately critical as Section 2. If there's information from the *beginning* of a long document that needs to be considered, it can be summarized or referenced here, especially if it's not already in Section 2.

**Supporting Details:**
*   **[Detail 4: Background information, examples, or nuances.]**
*   **[Detail 5: Information that might be in the middle of a long text and needs to be brought forward.]**

**Reinforcement:** Pay close attention to the details presented in Section 2 and Section 3, as they collectively form the complete picture required for accurate task execution.

## Section 4: Specific Instructions & Output Format
Detailed steps and the desired output format.

**Instructions:**
1.  **[Step 1: First action or analysis required.]**
2.  **[Step 2: Subsequent action, building on Step 1.]**
3.  **[Step 3: Further analysis or synthesis.]**

**Output Requirements:**
*   **Format:** [Specify desired format: e.g., Markdown, JSON, numbered list, specific headings.]
*   **Length:** [Specify any length constraints, e.g., "briefly," "in detail," "max 3 sentences per point."]
*   **Focus:** Ensure your response directly addresses the **Core Task** (Section 1) while incorporating insights from the **Critical Context** (Section 2) and **Supporting Details** (Section 3).

---

**Final Check:** Before generating the response, quickly review the **Key Information** in Section 2 and **Supporting Details** in Section 3 to ensure all vital aspects are integrated.

### Role
You are an AI researcher specializing in Large Language Model (LLM) behavior and prompt engineering.

### Context
A recurring problem has been observed when working with LLMs like ChatGPT: the model is highly accurate on the initial parts of a long text prompt or document, and also accurate on the concluding parts, but becomes vague, inaccurate, or loses coherence in the middle sections. This phenomenon is related to "Initial Saliency" and positional encoding limitations.

### Problem Statement
The goal is to mitigate the "lost middle" problem in LLMs when processing long texts.

### Technique: Initial Saliency Mitigation

This prompt is designed to counteract the "Initial Saliency" effect, where early tokens disproportionately influence attention, and to address general positional encoding limitations that affect middle sections of long inputs.

### Action Plan

1.  **Decomposition and Re-assembly:** Break down the long text into smaller, manageable chunks.
2.  **Contextual Priming:** For each chunk, explicitly re-establish its relevance and connection to the overall task and the preceding/succeeding chunks.
3.  **Reinforce Key Information:** Ensure critical information from the middle sections is highlighted and explicitly referenced.
4.  **Iterative Refinement:** If possible, simulate a process where the model first processes chunks sequentially and then synthesizes them, paying special attention to the connections between chunks.

### Prompt Structure for Mitigation

You will process a long text (or a series of related texts). For each significant segment, follow these instructions:

**Segment [Number] of [Total Segments]:**

*   **Focus:** [Briefly state the primary topic or purpose of this segment.]
*   **Connection to Previous:** [How does this segment build upon or relate to the immediately preceding segment? Be specific.]
*   **Core Content:** [Summarize the most critical information or arguments within this segment. Be concise and precise. If this is a middle segment, ensure no detail is lost.]
*   **Connection to Next:** [How does this segment set up or lead into the immediately succeeding segment? Be specific.]
*   **Overall Task Relevance:** [Reiterate how this segment contributes to the final objective of the overall task.]

**Example for a Middle Segment (Segment 3 of 5):**

*   **Focus:** Detailed analysis of the "Initial Saliency" research findings.
*   **Connection to Previous:** This segment elaborates on the core problem identified in Segment 2 (the "lost middle" phenomenon), providing the specific research that explains *why* it occurs.
*   **Core Content:** The key findings of the "Initial Saliency" paper indicate that LLMs' attention mechanisms are biased towards tokens that receive high attention from the very first token of the input. This "black hole" effect draws attention to the beginning and end, neglecting intermediate information. The paper identifies this as a second major cause of poor performance on middle text sections, in addition to earlier positional encoding limitations.
*   **Connection to Next:** This detailed understanding of the "Initial Saliency" mechanism (Segment 3) will inform the strategies proposed in Segment 4 for mitigating this specific issue.
*   **Overall Task Relevance:** Understanding the root causes of the "lost middle" is crucial for developing effective prompts and strategies to ensure comprehensive LLM comprehension of long texts.

### Task
Apply the above structure to a given long text. You will be provided with the text, broken into segments if necessary. Your output should follow the "Segment [Number] of [Total Segments]" format for each part of the text you process, ensuring that the core content of *all* segments, especially the middle ones, is accurately and comprehensively captured and its relevance is explicitly stated.

**Output Format:** Use Markdown for clarity, with bolded headers for each instruction point within each segment.

# Role: LLM Text Analysis Expert

# Context:
The user is experiencing a common issue with large language models (LLMs) like GPT-4, Claude, and Gemini: they tend to focus heavily on the beginning and end of long texts, while the middle sections are often summarized less accurately or even ignored. This "End-Weighting Bias" leads to unbalanced summaries and a loss of crucial information from the core of the text.

# Objective:
To create a reusable prompt template that mitigates the End-Weighting Bias by forcing the LLM to process a long text in a balanced, section-by-section manner, ensuring a comprehensive and equitable summary.

# Technique: End-Weighting Bias Mitigation (Forced Segmentation)

## Instructions:

You will be provided with a long text. To ensure a balanced analysis and prevent the End-Weighting Bias, you must process this text by explicitly dividing it into segments and analyzing each segment individually before synthesizing the overall summary.

**Step 1: Text Segmentation**
Divide the provided text into **three (3) distinct, roughly equal segments**:
- **Segment 1:** The beginning of the text.
- **Segment 2:** The middle section of the text.
- **Segment 3:** The end of the text.

**Step 2: Segmented Analysis**
For each segment, perform the following analysis:

### Analysis of Segment 1
- **Key points:** Summarize the main ideas and information presented in Segment 1.
- **Significance:** Explain the importance or role of this segment in the overall context of the text.

### Analysis of Segment 2
- **Key points:** Summarize the main ideas and information presented in Segment 2.
- **Significance:** Explain the importance or role of this segment in the overall context of the text.

### Analysis of Segment 3
- **Key points:** Summarize the main ideas and information presented in Segment 3.
- **Significance:** Explain the importance or role of this segment in the overall context of the text.

**Step 3: Synthesized Summary**
After analyzing each segment individually, create a final, comprehensive summary of the entire text. This summary must:
- **Integrate findings from all three segments.**
- **Provide a balanced overview**, giving appropriate weight to information from the beginning, middle, and end.
- **Avoid disproportionately focusing on the end.**

# Output Format:
Structure your response clearly using Markdown headings for each section:
- `## Segment 1 Analysis`
  - `### Key Points`
  - `### Significance`
- `## Segment 2 Analysis`
  - `### Key Points`
  - `### Significance`
- `## Segment 3 Analysis`
  - `### Key Points`
  - `### Significance`
- `## Synthesized Summary`

**Begin processing the provided text now.**