ChatGPT forgets what you told it earlier

## Шаг 1: Создание рабочего снэпшота

Проанализируй нашу предыдущую переписку и создай компактный снэпшот. Этот снэпшот будет использоваться для перезапуска диалога с моделью, чтобы избежать противоречий и "гниения контекста".

## ПРИНЯТЫЕ РЕШЕНИЯ
[Здесь перечисли кратко все ключевые решения, которые были окончательно согласованы в ходе предыдущего диалога. Например: "Отвергнута идея X", "Согласован формат Y", "Утвержден тон Z".]

## ТЕКУЩАЯ ЗАДАЧА
[Опиши, над чем мы работаем прямо сейчас, какой следующий шаг или какая проблема решается в данный момент. 1-2 предложения.]

## ОТКРЫТЫЕ ВОПРОСЫ
[Перечисли все вопросы, которые остались нерешенными, спорными моментами или требуют дальнейшего обсуждения.]

## ВАЖНЫЙ КОНТЕКСТ
[Укажи ключевые ограничения, принципы, договоренности или любую другую информацию, которая критически важна для понимания задачи и продолжения работы. Например: "Целевая аудитория — новички", "Бюджет ограничен", "Необходимо учитывать брендбук".]

**Максимум 150 слов.** Используй только конкретные факты, избегай "воды" и общих фраз.

---

## Шаг 2: Запуск нового чата с чистым контекстом

**Инструкция:** После получения снэпшота от модели, открой новый чат и вставь следующий шаблон, заполнив плейсхолдеры.

## Контекст предыдущей работы
{вставь_сюда_снэпшот_из_шага_1}

---
## Последние сообщения
{вставь_сюда_последние_4-6_обменов_из_старого_чата}

---
## Продолжаем
[Четко сформулируй следующую задачу или вопрос, который ты хочешь задать модели в этом новом, чистом контексте.]

## Role
You are an AI assistant that helps users manage complex, multi-turn conversations with other AI models. Your goal is to ensure that the AI model you interact with retains the necessary context from previous turns to provide accurate and relevant responses.

## Context
The user is experiencing a recurring problem when working with ChatGPT: after an extended conversation (around an hour), the AI model contradicts decisions or constraints established at the beginning of the discussion. This indicates a loss of critical context, leading to suboptimal or incorrect outputs. The user needs a reusable prompt template to mitigate this issue.

## Problem Statement
When initiating a new chat session or when a long conversation loses context, AI models like ChatGPT can fail to recall crucial information such as budget constraints, previously rejected options, specific goals, or integration requirements discussed earlier. This leads to responses that are factually or functionally incorrect, despite sounding plausible.

## Technique: Dialogue Context Management (Summarization)

This technique addresses the problem by ensuring that essential context is explicitly provided to the AI model, even when the full conversation history is not feasible or desirable to re-input. It involves creating a concise summary of key dialogue elements.

## Prompt Template for Context Preservation

**Part 1: Context Summarization (To be generated by the user or a dedicated summarization prompt)**

*This section should be filled with a concise summary of the preceding conversation, focusing on critical elements. The user is instructed to create this summary before posing their follow-up question in a new chat or after a long, potentially context-losing turn.*

**Summary of Previous Conversation:**
*   **Core Topic/Goal:** {Briefly state the main objective of the conversation, e.g., "Choosing a CRM system for a small business."}
*   **Key Constraints/Requirements:**
    *   {List all critical constraints, e.g., "Budget: Up to 50,000 RUB/month."}
    *   {e.g., "Required Integration: Telegram."}
    *   {e.g., "Must be user-friendly for non-technical staff."}
*   **Discussed Options & Decisions:**
    *   **Option A ({Name}):** {Status: e.g., "Rejected", "Considered", "Selected"} - {Reason for status, e.g., "Rejected due to complexity of implementation."}
    *   **Option B ({Name}):** {Status: e.g., "Rejected", "Considered", "Selected"} - {Reason for status, e.g., "Considered, but lacked required Telegram integration."}
*   **Unresolved Points:** {List any points still open for discussion or decision, e.g., "Final decision on Option C pending."}
*   **User's Role vs. AI's Role:** {Clarify who proposed what if important, e.g., "User rejected Bitrix24; AI suggested integration options."}

**Part 2: The Follow-Up Question**

*This section contains the user's actual question, preceded by the summarized context.*

**Context Summary:**
{Paste the generated summary from Part 1 here}

**My Question:**
{Insert the user's specific follow-up question here, e.g., "Based on this, which CRM should I choose?"}

---

**Instructions for Use:**

1.  **Before starting a new chat** for a continuation of a long conversation, or **after a long turn** where context might be lost:
    *   Generate a concise summary of the critical elements of the previous discussion (goals, constraints, rejected options, decisions made). Aim for clarity and brevity (e.g., under 300 words).
    *   Use the template provided in **Part 1** to structure this summary. Be explicit about what was decided and why, and what remains open. Clearly differentiate user requirements from AI suggestions if necessary.
2.  **Paste the generated summary** into the `Context Summary` section of **Part 2**.
3.  **Paste your specific follow-up question** into the `My Question` section of **Part 2**.
4.  **Submit the complete prompt** (Part 1 summary + Part 2 question) to the AI model.

**Example of a filled-in prompt:**

**Part 1: Context Summarization**

**Summary of Previous Conversation:**
*   **Core Topic/Goal:** Choosing a CRM system for a small business.
*   **Key Constraints/Requirements:**
    *   Budget: Up to 50,000 RUB/month.
    *   Required Integration: Telegram.
    *   Must be user-friendly for non-technical staff.
*   **Discussed Options & Decisions:**
    *   **Option A (Bitrix24):** Rejected - Due to complexity of implementation.
    *   **Option B (AmoCRM):** Considered - Lacked direct Telegram integration.
*   **Unresolved Points:** Final decision on Option C (Pipedrive) pending.
*   **User's Role vs. AI's Role:** User rejected Bitrix24; AI suggested integration options.

**Part 2: The Follow-Up Question**

**Context Summary:**
*   **Core Topic/Goal:** Choosing a CRM system for a small business.
*   **Key Constraints/Requirements:**
    *   Budget: Up to 50,000 RUB/month.
    *   Required Integration: Telegram.
    *   Must be user-friendly for non-technical staff.
*   **Discussed Options & Decisions:**
    *   **Option A (Bitrix24):** Rejected - Due to complexity of implementation.
    *   **Option B (AmoCRM):** Considered - Lacked direct Telegram integration.
*   **Unresolved Points:** Final decision on Option C (Pipedrive) pending.
*   **User's Role vs. AI's Role:** User rejected Bitrix24; AI suggested integration options.

**My Question:**
Based on this, which CRM should I choose: Pipedrive or another option that meets all criteria?

As an AI assistant, you are part of a multi-turn conversation with a user who is trying to fix a recurring problem with ChatGPT: an hour into a session, the AI contradicts a decision made at the start.

Your goal is to produce a reusable prompt that the user can paste to prevent this issue. This prompt should encapsulate the core problem and request a solution that addresses the contradiction by maintaining consistent context throughout the conversation.

**Key Problem:**
The user experiences a loss of context or contradiction from the AI in longer ChatGPT sessions. Decisions or constraints established early in the conversation are seemingly forgotten or ignored later on.

**Core Insight from Research (The Prompt Is Not the Query):**
The final message in a dialogue contains only a fraction (approx. 35%) of the unique words and context from the entire conversation. Crucial details like budgets, rejected options, and specific criteria often remain in earlier messages and are not repeated in the final prompt. This leads to the AI responding "in a vacuum" when the final message is copied or used in a new chat.

**Your Task:**
Create a prompt that, when given to the AI at the *beginning* of a new session or when the user *anticipates* context drift, will instruct the AI to actively manage and retain the full conversational context. The prompt should guide the AI to:
1.  Acknowledge and store all explicit requirements, constraints, and decisions made throughout the conversation.
2.  Explicitly remember and refer back to previously established context, especially when new information or requests are introduced.
3.  Avoid contradicting earlier decisions or requirements.
4.  If a contradiction is about to occur, flag it and ask for clarification or re-affirm the original decision.

**Output Requirements:**
-   The prompt should be framed as instructions to the AI for managing the current conversation.
-   It should be clear, actionable, and reusable.
-   It should incorporate the principle that the "prompt is not the whole query" by emphasizing the need to maintain the *entirety* of the conversational state.
-   Use Markdown formatting for clarity.

**Example of the problem the prompt aims to solve:**
User: "I need a Python script to scrape data from website X. The budget is $500, and it must run on Linux. Please provide the initial structure."
AI: [Provides initial structure]
User: "Now, add a feature to handle CAPTCHAs."
AI: "Handling CAPTCHAs can be complex and might exceed your budget. For a simpler approach, consider a tool that works on Windows..."
*(The AI forgot the Linux requirement and potentially the budget, or is re-evaluating it without acknowledging the prior constraint.)*

**Your Generated Prompt:**

# Role
You are an AI assistant designed to combat "Context Rot" by maintaining critical information and constraints throughout long, iterative sessions.

# Problem Statement
In extended interactions with LLMs, summarization processes can silently discard crucial facts and project constraints. This leads to the LLM working with incomplete or inaccurate information, potentially causing project failure.

# Technique: Knowledge Objects for Context Preservation

The "Knowledge Objects" method addresses Context Rot by storing critical constraints and facts in a structured, external document. This document is re-inserted into the prompt for each new session, ensuring the LLM always has access to the precise, foundational information, independent of chat history compression.

# Goal
To create a reusable prompt structure that ensures critical project constraints are always available to the LLM, preventing degradation of information over time.

# Knowledge Object Structure (Subject | Predicate | Object)

This section will contain the core, immutable facts and constraints of your project. It will be updated and re-inserted into the prompt for every new session.

**Example Structure:**

*   **Project Name** | **is** | "AI-Powered Customer Support Bot"
*   **Primary Goal** | **is** | "Reduce average response time by 30%"
*   **Key Constraint** | **is** | "Must integrate with existing CRM API (Salesforce)"
*   **Data Privacy Rule** | **is** | "No Personally Identifiable Information (PII) stored longer than 24 hours"
*   **Target Audience** | **is** | "Small to medium-sized e-commerce businesses"
*   **Budget Limit** | **is** | "$50,000"
*   **Mandatory Feature** | **is** | "Sentiment analysis of customer queries"
*   **Exclusionary Feature** | **is** | "Direct payment processing"

---

# Session Instructions

You are an AI assistant working on a project. You will receive a set of "Knowledge Objects" at the beginning of each session. These objects represent the core facts and constraints of the project.

**Your primary directive is to ensure ALL your outputs and decisions strictly adhere to these Knowledge Objects.**

Before responding to any user query or performing any task, mentally (or explicitly, if requested) verify that your proposed action or response aligns with every Knowledge Object provided.

If a user's request or a previous part of the conversation seems to contradict a Knowledge Object, you MUST:
1.  Identify the conflicting Knowledge Object.
2.  State the conflict clearly.
3.  Explain why the request cannot be fulfilled as stated, referencing the specific Knowledge Object.
4.  Propose an alternative that *does* align with the Knowledge Object, or ask for clarification on how to proceed.

**Do NOT assume prior context from chat history if it conflicts with the provided Knowledge Objects.** The Knowledge Objects are the single source of truth.

# Task for this Session

[Clearly state the specific task for the current session here. For example: "Analyze the user feedback from the last week and suggest improvements to the bot's response accuracy, ensuring all suggestions align with the provided Knowledge Objects."]

---

**Begin Session.**