Topic Analysis: Chapter 4 - During and After - Iteration

Metadata

• Syllabus Reference: Part 2, Chapter 4
• Primary Sources: Articles 2, 3, 4, 9
• Secondary Sources: Interview Q6-7, Q11
• Analysis Date: 2025-11-28
• Status: Complete

---

1. Source Materials

1.1 Primary Sources

From Article 2: "What Good Context Looks Like"

Editing prompts mid-work:

"Don't send a correction message ('sorry, I meant just the first form'). Edit the original prompt to be more specific. Continue working with new requirements. If I sent a message that I meant it differently, AI would make unnecessary intermediate steps."

Iteration revelation:

"I'm talking about some form. AI responds and shows that we have 2 implementations for that form in the code. I didn't know or forgot to mention this. My change relates to just one type. What I do: Edit the original prompt to be more specific."

From Article 3: "Why You Give Much and Get Little"

When to fix vs start fresh:

"I fix when: Output is at least 80% correct, AI understood the task just made minor mistakes, Dealing with non-technical things. I start over when: Output is completely off target, Previous steps were wrong and AI still considers them, I said 'exclude Amazon products' but it includes them again."

Explaining WHY it's wrong:

"When I correct it, I always explain why it's wrong. This is important - so it has that lesson in context and can 'learn' from it."

Context contamination:

"When AI writes an article or generates code that's wrong, it's still there as a reference taking up space. Later it can happen that this reference in context, even though it was later marked as wrong, gets used again."

Feedback with role play:

"Instead of asking 'Is my plan good?' try: 'You are a professional programmer with 10 years experience in React. What do you think about this architecture?'"

From Article 4: "Think Like an Engineer"

The 2-minute rule:

"For me, vibe coding isn't giving a prompt and going for coffee. It's: Give specific prompt, Have result in seconds, max 1-2 minutes, If it takes longer, it's an error. Usually when it takes long, result is bad too."

Clear goals, flexible path:

"You need clearly defined goals from the start. BUT! You need to expect that on the way to goals, those goals might: Expand, Deepen, Shrink, Change"

From Article 9: "Vibe Coding vs Context Engineering"

Context contamination emphasis:

"When AI responds to your latest message, context contains everything previously entered or generated. Bad content can mislead AI. Even after fixing and being satisfied, continuing work can lead AI astray."

Cleaning up:

"Better to clean up: Edit messages/prompts and delete bad content, Start new session: 'this is my new article/code but unfinished, want to continue', Add direction you don't want AI to take"

1.2 Secondary Sources

From Interview Q6 (Signals of Wrong Path)

"Usually when I see AI doing many operations (unless prompt requires it) or takes more than 1-2 minutes, result will be bad. When I rush with too general request, just seeing what AI searches or writes, I know it's wrong."

From Interview Q7 (Fix vs Restart)

"Consider whether 'bad context' outweighs the good. If yes, effort to create new session with correct context will be less than fixing it."

From Interview Q11 (Prompt Editing)

"When output contains info I didn't know or forgot to mention and would improve result - edit original prompt to be more specific."

1.3 External Citations

None specific - focus is on practical iteration patterns.

---

2. Content Extraction

2.1 Key Concepts

1. The 2-Minute Rule

   • Definition: If AI takes longer than 1-2 minutes, something's wrong
   • Signal: Long processing often means bad result coming
   • Action: Stop and adjust, don't wait
   • Source: Article 1, Article 4

2. Explain WHY It's Wrong

   • Definition: AI learns from "wrong because X" not just "wrong"
   • Example: "This is wrong because we use TypeScript, not JavaScript"
   • Source: Article 3

3. Context Contamination

   • Definition: Bad output stays in context and can mislead later
   • Solution: Start fresh session or edit/delete bad content
   • Source: Article 2, Article 3, Article 9

4. Fix vs Start Fresh Decision

   • Fix when: 80%+ correct, minor mistakes, understood task
   • Fresh when: Completely off, repeating same mistakes, ignoring instructions
   • Source: Article 3

5. Editing Prompts (Not Sending Corrections)

   • Definition: Edit original prompt instead of adding "sorry, I meant..."
   • Why: Avoids unnecessary intermediate steps
   • Tool: Zed allows this, others need new session
   • Source: Article 2, Interview Q11

6. Ask for Opinion, Not Validation

   • Definition: "Is my plan good?" gets praise, not feedback
   • Better: "What would you suggest?" / "How would you solve this?"
   • Source: Article 3

2.2 Key Examples

1. Form Implementation Discovery

   • Context: AI reveals 2 implementations exist
   • Wrong: Send "sorry, I meant first form"
   • Right: Edit original prompt to specify
   • Source: Article 2

2. Amazon Products Keep Appearing

   • Context: Told to exclude, but they keep showing
   • Decision: Bad context outweighs good, start fresh
   • Source: Article 3

3. Bug Location Discovery

   • Context: Solving "user can't save form"
   • Iteration: Problem not in saving but validation → only certain emails → found 2 more bugs → common cause
   • Result: Goal changed from "fix saving" to "fix email validation"
   • Source: Article 4

4. Role Play Feedback

   • Context: Need real feedback on architecture
   • Before: "Is my plan good?" (gets praise)
   • After: "You are a 10-year React expert. What do you think?"
   • Source: Article 3

2.3 Key Quotes

1. "If it takes longer than 1-2 minutes for first usable output, I know something's wrong." - Article 1

   • Use for: 2-minute rule

2. "Never ask AI 'Is my plan good?' Current models are set to praise you." - Article 3

   • Use for: Validation trap

3. "When AI writes wrong code, it stays as reference taking space. Later this reference can be reused even after being marked wrong." - Article 9

   • Use for: Context contamination

4. "I always explain why it's wrong - so it has that lesson in context." - Article 3

   • Use for: Teaching AI

2.4 Data/Statistics

• 80% correct threshold for fixing vs restarting
• 1-2 minute rule for AI response
• Every 8-12 interactions: Summary checkpoint

---

3. Gap Analysis

3.1 Content Gaps

• [x] 2-minute rule covered
• [x] Explain WHY covered
• [x] Context contamination covered
• [x] Fix vs fresh decision covered
• [x] Prompt editing covered
• [x] Opinion not validation covered
• [ ] Could add more concrete iteration examples

3.2 Clarity Issues

• None - concepts practical and clear

3.3 Depth Assessment

• Strong decision framework
• Clear action points
• Good balance theory/practice

---

4. Structure Proposal

4.1 Chapter Outline

Chapter 4: During and After - Iteration

Section 4.1: The 2-Minute Rule

• Main point: Long processing is a signal, not patience required
• Content from: Article 1, Article 4
• Include: What to do when it takes too long

Section 4.2: Explain WHY It's Wrong

• Main point: AI learns from reasons, not just corrections
• Content from: Article 3
• Include: "wrong because X" pattern

Section 4.3: Context Contamination

• Main point: Bad output pollutes future responses
• Content from: Article 2, Article 3, Article 9
• Include: When to clean up, how to start fresh

Section 4.4: Ask for Opinion, Not Validation

• Main point: Models are tuned to praise, get real feedback differently
• Content from: Article 3
• Include: Better question patterns, role play technique

Section 4.5: Editing Prompts Mid-Work

• Main point: Edit originals, don't send correction messages
• Content from: Article 2
• Include: Zed approach, general tool advice

4.2 Opening Hook

"AI gave you something wrong. Do you send 'sorry, I meant...' and hope for the best? Or do you know when to fix, when to restart, and how to actually teach AI what you need?"

4.3 Key Takeaways

1. The 2-minute rule: Long processing = signal to stop and adjust
2. Explain WHY something is wrong, not just that it's wrong
3. Bad output contaminates context - sometimes starting fresh is faster
4. Don't ask "Is this good?" - ask "What would you do?" for real feedback
5. Edit original prompts instead of sending correction messages

4.4 Transition

"Now that you know how to prepare context and iterate effectively, let's look at the specific tools that make this workflow practical."

---

5. Writing Notes

5.1 Tone/Voice

• Problem-solving oriented
• Decision frameworks
• "When X, do Y" patterns

5.2 Audience Considerations

• Universal iteration principles
• Technical examples (code) and general (articles, research)
• Clear decision points

5.3 Potential Visuals

1. Fix vs Fresh Decision Tree

   • 80% correct? → Fix
   • Repeating mistakes? → Fresh
   • Context contaminated? → Fresh

2. Good vs Bad Feedback Questions

   • "Is this good?" vs "What would you suggest?"

3. Context Contamination Diagram

   • Visual of bad output persisting

---

6. Prepared Citations

Internal

• [A2] Article "What Good Context Looks Like"
• [A3] Article "Why You Give Much and Get Little"
• [A4] Article "Think Like an Engineer"
• [A9] Article "Vibe Coding vs Context Engineering"
• [I6] Interview Q&A, Question 6 (Signals)
• [I7] Interview Q&A, Question 7 (Fix vs restart)
• [I11] Interview Q&A, Question 11 (Prompt editing)

External

• None specific to this chapter

---

7. Open Questions

1. How much detail on Zed-specific features?

   • Decision: Mention as example, keep principles tool-agnostic

2. Include role play examples for different professions?

   • Decision: Yes, 2-3 examples (programmer, marketer, writer)

3. Summary checkpoint (every 8-12 interactions) - include here?

   • Decision: Yes, as practical tip for long sessions