Module 9: Context Management

Lesson 5: Multi-turn Conversation Handling

The most complex engineering tasks are Multi-turn. You ask a question, the model asks for a file, you provide the file, the model runs a test, it fails, it tries again. Every turn adds to the context. In a multi-turn scenario, the "Prompt" is a living, breathing object.

In this final lesson of Module 9, we look at how to manage the "Narrative Flow" of an AI session.

1. The "Current State" Block

In a multi-turn task, you should include a dedicated section in the system prompt called "CURRENT_STATUS."

At the end of every turn, update this block with: "Successfully read file A. Currently attempting to fix the bug in function B. 2 tests remain failing."

By giving the model a clear "Where am I now?" anchor, you prevent it from getting confused by old, failed attempts in the history.

2. Preventing "Repetition Loops"

In multi-turn sessions, models can get "Stuck." They try the same command 3 times and get the same error 3 times.

The Architect's Fix: Sequence Detection.
Check the history programmatically. If the model has called the same tool with the same arguments twice in a row, inject a System Instruction: "WARNING: You have attempted this 2 times already. Try a different architectural approach."

3. The "Final Turn" Summarization

When a task is complete, don't just stop. Ask for a "Success Summary."

"Great. Now summarize what you changed and why, so I can commit it to Git."
This final turn serves as a "Clean Handoff" to the next part of your human workflow.

4. Visualizing Multi-turn State

graph TD
    A[Turn 1: User Request] --> B[Turn 2: Tool Call]
    B --> C[Turn 3: Tool Result]
    C --> D[Turn 4: Reasoning + State Update]
    D --> E[Turn 5: Final Answer]
    Note right of D: State Anchor: 'Fixing bug in Line 45'

5. Summary of Module 9

Module 9 has mastered "Memory."

You defined the Context Window (Lesson 1).
You monitored the Ceiling (Lesson 2).
You optimized for SNR (Lesson 3).
You automated Compression (Lesson 4).
You managed the Multi-turn Narrative (Lesson 5).

In Module 10, we move to the most critical topic for safe deployment: Reliability and Guardrails.

Interactive Quiz

Why should you maintain a "CURRENT_STATUS" block in a 20-turn session?
What is a "Repetition Loop" and how can a system-level check break it?
Why is it important to summarize a multi-turn task upon completion?
Scenario: An agent is in Turn 15 of a debugging task. The history is 80k tokens. Describe a strategy that uses both "Summarization" and "Status Anchoring" to keep the agent focused.

Reference Video: