Day 4 - Session 1: Building Your Research Infrastructure

Managing state across conversations

Author

Dr Brian Ballsun-Stanton

Published

September 11, 2025

Good Morning!

Today’s Big Question

How should we work?

This Morning

Build infrastructure for complex projects
Learn why conversations need management
Practice strategic termination/restart
Test context window limits

Show Me Your Prompts! (15 min)

From Your Bibliography Work

Who found 3-5 sources with good quotes?
What annotation patterns are emerging?
Which models gave you the most trouble yesterday?
Any spectacular failures to share?

Key question: How would you maintain this bibliography across multiple work sessions?

What We’ll Learn This Session

By the end of this morning, you will be able to:

Create: System prompts that maintain consistent behavior
Execute: Strategic conversation management
Analyze: Context window overflow patterns
Apply: State management to your project

The Context Window Problem

What Happens in Long Projects?

Morning: Start bibliography
Afternoon: Add more sources
Tomorrow: Continue work
Next week: …what bibliography?

The Bucket Overflows

Old information falls out. No memory between conversations. Inconsistent behavior across sessions.

Three Types of State

1. Project Documents

Your bibliography, notes, outline - the artifacts
Notes to the LLM, context preserved across sessions, fiddly todo lists

2. User and System Prompts

Instructions for consistent AI behavior

3. Conversation History

The current discussion context

We need to manage all three.

Exercise Part 1: Document Preparation (15 min)

Create Your Research State File

# Research Project: [Your Topic]

## Current Status: Day 4


## Bibliography

[Paste your 3-5 sources with quotes from yesterday]

## Key Questions

1. [Your main research question]
2. [Sub-questions]

## Working Notes

[Any insights so far]

Save this as a text file on your computer.

Exercise Part 2: System Prompt Design (20 min)

Write Your Research Assistant Prompt

TODO

Test this with your document. Green sticky when working.

Exercise Part 3: Conversation Management (30 min)

Practice the Workflow

Start fresh conversation
Load system prompt
Paste project document
Work for 5-10 exchanges
Before overflow: Save important outputs
Terminate conversation
Start fresh with updated project document

The Key Skill

Knowing WHEN to restart, not just how.

Signs of Context Overflow

Watch For

AI “forgets” earlier instructions
Quotes become approximate
Behavior becomes generic
Response time increases dramatically

The Rule

Better to restart early than fight degraded performance.

Discussion: State Management Strategies (15 min)

In Pairs, Compare

How often did you need to restart?
What did you prioritize keeping?
What got lost despite your efforts?

Critical Insight

State management is labor. This is the fundamental judgement we need to apply.

The Infrastructure Trade-off

Convenience (ChatGPT, Claude.ai)

Projects and other convienences
But: No control over priorities or specific modes of interaction

Control (ChatCraft + manual management)

You decide what matters
But: Constant maintenance
Visible: The labor involved

Looking Ahead

This Afternoon: The Boring Important Stuff

Legal implications of AI use
Privacy and security
Terms of Service deep dive

Tomorrow: Breaking Everything

Using your infrastructure to find limits
Systematic failure exploration
Why breaking things teaches us

Key Takeaways

Context windows are finite - plan for the next conversation
State requires management - it’s not automatic
Restart strategically - before degradation
Document everything - your memory, not AI’s

Before the Break

Update your research state file with: - Today’s new insights - Refined system prompt - Questions for tomorrow’s breaking session

Save multiple versions - we’ll break them tomorrow!

See you at 11:00 for legal and ethical discussions!

--- title: "Day 4 - Session 1: Building Your Research Infrastructure" subtitle: "Managing state across conversations" author: - "Dr Brian Ballsun-Stanton" date: "11 September 2025" format: revealjs: slide-number: true chalkboard: true theme: [default, ../reveal-custom.scss] footer: "AI Summercamp 2025 - Basic Course" preview-links: auto transition: slide background-transition: fade highlight-style: github navigation-mode: vertical --- ## Good Morning! ### Today's Big Question **How *should* we work?** ### This Morning - Build infrastructure for complex projects - Learn why conversations need management - Practice strategic termination/restart - Test context window limits --- ## Show Me Your Prompts! (15 min) ### From Your Bibliography Work 1. Who found 3-5 sources with good quotes? 2. What annotation patterns are emerging? 3. Which models gave you the most trouble yesterday? 4. Any spectacular failures to share? **Key question**: How would you maintain this bibliography across multiple work sessions? --- ## What We'll Learn This Session By the end of this morning, you will be able to: - **Create**: System prompts that maintain consistent behavior - **Execute**: Strategic conversation management - **Analyze**: Context window overflow patterns - **Apply**: State management to your project --- ## The Context Window Problem ### What Happens in Long Projects? - Morning: Start bibliography - Afternoon: Add more sources - Tomorrow: Continue work - Next week: ...what bibliography? ### The Bucket Overflows Old information falls out. No memory between conversations. Inconsistent behavior across sessions. --- ## Three Types of State ### 1. Project Documents - Your bibliography, notes, outline - the artifacts - Notes to the LLM, context preserved across sessions, fiddly todo lists ### 2. User and System Prompts Instructions for consistent AI behavior ### 3. Conversation History The current discussion context We need to manage all three. --- ## Exercise Part 1: Document Preparation (15 min) ### Create Your Research State File ```markdown # Research Project: [Your Topic] ## Current Status: Day 4 ## Bibliography [Paste your 3-5 sources with quotes from yesterday] ## Key Questions 1. [Your main research question] 2. [Sub-questions] ## Working Notes [Any insights so far] ``` Save this as a text file on your computer. --- ## Exercise Part 2: System Prompt Design (20 min) ### Write Your Research Assistant Prompt TODO Test this with your document. Green sticky when working. --- ## Exercise Part 3: Conversation Management (30 min) ### Practice the Workflow 1. Start fresh conversation 2. Load system prompt 3. Paste project document 4. Work for 5-10 exchanges 5. Before overflow: Save important outputs 6. Terminate conversation 7. Start fresh with updated project document ### The Key Skill Knowing WHEN to restart, not just how. --- ## Signs of Context Overflow ### Watch For - AI "forgets" earlier instructions - Quotes become approximate - Behavior becomes generic - Response time increases dramatically ### The Rule Better to restart early than fight degraded performance. --- ## Discussion: State Management Strategies (15 min) ### In Pairs, Compare 1. How often did you need to restart? 2. What did you prioritize keeping? 3. What got lost despite your efforts? ### Critical Insight State management is labor. This is the fundamental judgement we need to apply. --- ## The Infrastructure Trade-off ### Convenience (ChatGPT, Claude.ai) - Projects and other convienences - But: No control over priorities or specific modes of interaction ### Control (ChatCraft + manual management) - You decide what matters - But: Constant maintenance - Visible: The labor involved --- ## Looking Ahead ### This Afternoon: The Boring Important Stuff - Legal implications of AI use - Privacy and security - Terms of Service deep dive ### Tomorrow: Breaking Everything - Using your infrastructure to find limits - Systematic failure exploration - Why breaking things teaches us --- ## Key Takeaways 1. **Context windows are finite** - plan for the next conversation 2. **State requires management** - it's not automatic 3. **Restart strategically** - before degradation 4. **Document everything** - your memory, not AI's --- ## Before the Break Update your research state file with: - Today's new insights - Refined system prompt - Questions for tomorrow's breaking session Save multiple versions - we'll break them tomorrow! See you at 11:00 for legal and ethical discussions!