Smell #12: Over-Iteration

Symptoms

• Your chat session is longer than 15 turns for a single small feature.
• You find yourself saying "No, that didn't work either" to the AI multiple times.
• Each "fix" introduces a new regression in a feature that was previously working.
• The code is worse after Turn 10 than it was at Turn 2.

The "Death Spiral" Pattern

1. Turn 1: "Add sorting." (AI works)
2. Turn 3: "Wait, sort is broken when I filter." (AI patches sort)
3. Turn 7: "Sorting works but it's very slow now." (AI adds "optimization")
4. Turn 12: "Sorting is fast but it crashes on empty lists." (AI adds more patches)
5. Turn 15: Code is a 200-line "Spaghetti Monster" that no one understands.

Debt Impact

This smell contributes to:

| Debt Category | Impact | |---------------|--------| | 🧠 KNOW | Reasoning is buried in a messy, unusable chat history (Prompt Drift). | | 💸 FIN | Massive token waste from long, redundant chat sessions. | | ⚙️ OPS | Fragile, "over-patched" code fails unpredictably in production. |

How to Fix

1. The Hard Stop: When you realize you're in a loop, STOP.
2. Learnings Audit: Identify why the AI is struggling (e.g., bad architecture, missing context).
3. Reset: Start a fresh chat session with a more precise Clean Prompt.

How to Prevent

• The 3-5 Turn Rule: If it's not working by turn 5, your prompt (or the architecture) is the problem. Reset.
• Manual Intervention: If the AI can't fix a bug in 2 tries, fix it yourself. You are the engineer.
• Small Batches: Break the task into smaller, atomic sub-tasks.

Related Smells

• Smell #2: Prompt Drift — Over-iteration is the process; drift is the result.
• Smell #1: Ambiguous Goal — Vague goals lead to more iterations.

Book Reference

• Chapter 2: Chat Until It Works — the anatomy of the iteration death spiral.
• Chapter 13: Clean Prompts — how to achieve "First Prompt Magic" and avoid loops.