Lesson 09 of 11 — Foundations Layer

Context & Conversation Design — Foundations Lesson 9

Scenario

You're drafting an important email with AI. First draft didn't work, so you refine it. Then you refine again. By the fifth refinement, you realize: the AI has context from our whole conversation. It's remembering what we've been working on. This changes everything. You can iterate. You can build on previous work. You can have AI "remember" your preferences and style. But there's a trap: AI's memory is limited. It eventually forgets. You need to understand how to use conversation context strategically.

---

Learning Objectives

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

1. Understand conversation context — what AI remembers and forgets

2. Design multi-turn conversations — effective back-and-forth with AI

3. Use iteration strategically — refining output through conversation

4. Build institutional memory — maintaining context across sessions

5. Know context window limits — when context breaks down

---

Think Before You Design a Conversation

Before you start a multi-turn conversation with AI, ask:

  • What information does AI need to know for the whole conversation?
  • How will I refine outputs without repeating context?
  • What should AI remember from earlier turns?
  • When do I need to reset and start fresh?
  • How long can this conversation sustainably continue?
  • ---

    How AI Memory Works

    Short-Term Memory (One Conversation)

    When you're in a conversation with an LLM, it remembers everything in that thread.

    What it remembers:

  • Your initial prompt
  • The AI's response
  • Your refinements
  • Context you provided
  • References to earlier parts
  • What it doesn't remember:

  • Conversations in other threads (separate chat)
  • Information outside the current conversation
  • Real-time information
  • Your personal preferences (unless you told it this conversation)
  • Context window limit: Every LLM has a limit on how much it can remember in one conversation.

  • Older models: 2K-4K tokens (~1,500-3,000 words)
  • Modern models: 100K+ tokens (~75,000+ words)
  • Practical impact: After 50,000-100,000 words of conversation, the AI might start "forgetting" early context. It loses track of what was said 30,000 words ago.

    Zero Long-Term Memory

    When the conversation ends, the AI forgets everything.

    Important truth: When you close the chat, the AI has no memory of you or the conversation.

    Implication: If you need to reference past work, you must copy-paste it or start a new conversation with the context included.

    ---

    Conversation Lifecycle

    Phase 1: Setup (First Message)

    What to do: Provide all context the AI needs for the entire conversation.

    What it includes:

  • Your role/context
  • Background information
  • Task or goal
  • Any constraints
  • Preferences
  • Success criteria
  • Example:

    `

    I'm an administrative assistant at [Company]. I need to write

    professional emails over the next few exchanges. Context:

  • I prefer formal but warm tone
  • My industry is [field]
  • Recipients are external clients
  • Keep emails under 200 words
  • Include specific examples when possible
  • Let me know you understand this context, then I'll give you

    the first email to draft.

    `

    Why this matters: Everything you say in Phase 1 shapes the entire conversation. Invest time here.

    ---

    Phase 2: Iteration (Middle Exchanges)

    What to do: Refine AI output through specific feedback.

    Good feedback patterns:

  • ❌ Bad: "Make this better"
  • ✅ Good: "Make this warmer. Add a personal touch about our past relationship."
  • ❌ Bad: "This doesn't work"
  • ✅ Good: "This sounds too formal. The client likes casual language. Try again with 'Hey' instead of 'Hello.'"
  • ❌ Bad: "Try again"
  • ✅ Good: "Good start. Now add a specific example of how we've solved this before."
  • Iteration strategy: Change one thing per round (tone, length, detail, structure). If you ask for five changes at once, the AI gets confused.

    Example iteration flow:

    1. Initial request: "Draft an email addressing a late delivery"

    2. AI response: [Generic apology email]

    3. Your feedback: "Good structure. Make it shorter (100 words max) and add what we're doing to prevent this."

    4. AI revision: [Shorter, solution-focused]

    5. Your feedback: "Better. Now add a gesture (discount, free shipping, etc.) to acknowledge the inconvenience."

    6. AI revision: [Final version]

    ---

    Phase 3: Closure (End of Conversation)

    What to do: Save or copy final output before the conversation ends.

    Why it matters: Once you close the conversation, you've lost all context. Next time you open a new chat, the AI won't remember you or this work.

    What to save:

  • Final output (email, document, etc.)
  • Context you provided (so you can reuse it)
  • Successful prompts (so you can iterate faster next time)
  • Your feedback approach (what worked, what didn't)
  • ---

    Using Context Strategically

    Pattern 1: Building on AI Output

    Goal: Refine step-by-step instead of rewriting from scratch

    How:

    1. First turn: "Generate initial draft of X"

    - AI produces something

    2. Second turn: "Improve it by doing Y"

    - AI remembers the draft and improves it

    3. Third turn: "Now also add Z"

    - AI builds on previous versions

    Benefit: You're not repeating context. AI is iterating on familiar ground.

    Example:

    Turn 1:

    `

    Draft an email to a customer about a service upgrade

    they're eligible for.

    `

    AI: [Draft email]

    Turn 2:

    `

    Make this shorter (under 150 words) and more focused

    on benefits than features.

    `

    AI: [Revised, shorter version]

    Turn 3:

    `

    Good. Now add a specific example of how this upgrade

    helped a similar customer.

    `

    AI: [Final version with example]

    ---

    Pattern 2: Providing Feedback on Tone/Style

    Goal: Get AI to understand your communication preferences

    How:

    1. Setup: "I prefer [tone description]. Examples:"

    - Then give 1-2 examples of good tone

    2. Throughout: Reference this preference: "Use the tone from my examples"

    3. AI learns: Over the conversation, AI calibrates to your style

    Benefit: Less time explaining what you want; AI converges on your preferences.

    Example:

    Setup turn:

    `

    My writing style is warm but professional. Here's

    an example email I sent (pasted below).

    Use this as your model for how I communicate.

    [Paste email example]

    Now, draft a follow-up email to that same customer

    about renewing their service.

    `

    Later turn:

    `

    Closer, but warmer. Remember the tone of my example

    — that's the warmth level I want here.

    `

    AI learns your style and applies it going forward.

    ---

    Pattern 3: Building Complex Outputs in Stages

    Goal: Create sophisticated outputs (proposals, strategies, plans) without overwhelming AI

    How: Break into clear phases

    1. Phase 1: Generate outline or structure

    - AI creates framework

    2. Phase 2: Fill in sections one at a time

    - You refine each section individually

    3. Phase 3: Review coherence and adjust

    - You ensure everything flows together

    Benefit: More control. Each section is strong. Less hallucination from trying to do too much at once.

    Example (building a project proposal):

    Turn 1:

    `

    Create an outline for a proposal to upgrade our

    customer database system. Include: problem statement,

    proposed solution, timeline, budget, risks, success metrics.

    `

    AI: [Outline]

    Turn 2:

    `

    Now expand the "Problem Statement" section.

    We've had 3 data breaches in 2 years, system is outdated,

    reporting is manual and error-prone. Make it compelling

    but professional.

    `

    AI: [Detailed problem section]

    Turn 3:

    `

    Good. Now expand "Proposed Solution." We're choosing

    [Vendor] database. Benefits: [list]. Implementation time: 6 months.

    `

    AI: [Solution section]

    (Continue for each section)

    Turn 6:

    `

    Good. Now review the whole proposal for flow and consistency.

    Are there gaps or contradictions? Let me know what could be improved.

    `

    AI: [Review + suggestions]

    ---

    Pattern 4: Reusable Conversation Contexts

    Goal: Use conversation templates for repeated tasks

    How:

    1. Create a master setup prompt with your context

    2. Reuse it at the start of each new conversation for that task type

    3. AI starts fresh knowing your preferences and constraints

    Benefit: You don't repeat yourself. New conversations start calibrated to your needs.

    Example (for routine email drafting):

    Master context:

    `

    You are helping me draft professional emails.

    Context: I work in [industry]. My communication style is [warm/formal/direct].

    I prefer emails under [length]. My audience is [internal/external/mixed].

    When I ask you to draft an email, include:

  • Subject line (if external)
  • Greeting
  • Body (clear and concise)
  • Call to action
  • Professional closing
  • Ready?

    `

    Every time you open a new chat to draft emails, paste this context first. The AI immediately knows your style and requirements.

    ---

    Context Window Limits: What Happens When You Hit Them

    Modern LLMs have large context windows (100K+ tokens). But you can run out.

    Warning Signs

  • 🚩 AI starts repeating itself
  • 🚩 AI forgets context from earlier in the conversation
  • 🚩 AI gives worse quality responses
  • 🚩 AI seems confused about what you're asking
  • What to Do

    Option 1: Start Fresh

    Copy your setup context and open a new conversation. This resets the window and gives you more space.

    Option 2: Summarize & Continue

    Ask AI: "Summarize what we've accomplished so far" → Copy the summary → Use it as context in the next conversation.

    Option 3: Focus on Final Output

    If you're close to done, skip refinement. Use what you have.

    Example:

    You've been building a proposal for 20 turns. AI's responses start getting worse.

    You: `

    Stop. We've covered a lot. Let me save where we are.

    Draft a one-paragraph summary of what we've accomplished

    and what's left to do.

    `

    AI: [Summary]

    You: [Copy summary, save conversation, close]

    Next conversation (new chat):

    You: `

    [Paste summary from last conversation]

    We were working on a proposal. Continue from where

    we left off. The next section we need to finish is [X].

    `

    AI: [Continues with context reset]

    ---

    Mission: Design a Multi-Turn Conversation

    The Task

    1. Pick a task you'd want AI to help with (writing, planning, analysis, etc.)

    2. Design the conversation flow — How would you break it into turns?

    3. Write your setup context — What does AI need to know from the start?

    4. Plan your iteration strategy — How would you refine outputs?

    5. Test it — Have a 4-5 turn conversation with an AI following your design

    ---

    Deliverables

    Conversation Design Document:

    `

    Task: [What are you building/writing/planning?]

    Setup Context (Turn 1):

    [Role, style, constraints, success criteria]

    Turn 1: [Your prompt]

    AI: [Output]

    Your feedback: [One specific refinement]

    Turn 2: [Your response based on feedback]

    AI: [Revised output]

    Your feedback: [Next refinement]

    [Continue through 4-5 turns]

    Final Output: [What you ended up with]

    What Worked: [What about your conversation design was effective?]

    What to Change: [What would you do differently next time?]

    `

    Document your experience. You're learning conversation design.

    ---

    AI Coach: Context Is Your Leverage Point

    Here's what separates people who get great AI output from people who don't:

    Poor users treat each prompt independently: "What should I write?" "Make it better." "Make it even better."

    Good users design conversations strategically: "Here's my context, here's my style, here's how I want to iterate."

    The difference is context. When you set up context well, the AI does better work and you need fewer refinements.

    Key insight: The setup prompt is worth the time. Invest 5 minutes in clear context, and you'll save 20 minutes in iteration.

    ---

    Reflection: Your Conversation Style

    How do you naturally iterate with AI? Do you make one-word adjustments or big requests? Do you know what works for you?

    Reflect:

  • When AI gives you something close to right, how do you refine it?
  • Do you describe what you want clearly, or do you expect AI to guess?
  • How many turns does a typical task take you?
  • What frustrates you about iterating with AI?
  • Understanding your natural style helps you design conversations that work for you.

    ---

    Portfolio Check

    Portfolio principle: Save successful conversations and context prompts.

    Over time, you'll build a library of:

  • Effective setup contexts (reuse for similar tasks)
  • Successful iteration patterns (feedback that works)
  • Tone/style examples (for future reference)
  • This library becomes your personal AI operating manual.

    ---

    Bonus Challenge

    Conversation archaeology: Find a conversation you had with AI a few weeks ago (if you can access history). Reread it. What did you do well? What could you have improved? How would you design that conversation differently now?

    ---

    Key Takeaways

  • AI remembers within a conversation but forgets when it ends
  • Context windows limit how much AI can remember (but modern limits are high)
  • Setup phase is critical — invest time in clear context
  • Iteration patterns: one refinement per turn, specific feedback, reference previous context
  • Reusable contexts save time on repeated tasks
  • When hitting context limits, summarize and start fresh
  • Next: Lesson 10 explores Output Formats & Constraints — structuring AI responses for different purposes.

    ← Previous Lesson Back to Lessons Next Lesson →