Module 3Failures & Basins

Slipping Into the Basin

Once the model fails twice, the third attempt is already doomed.

The Basin of Failure

Imagine a landscape with hills and valleys. When the model generates good output, you're on a hilltop—high ground, everything's working. But when it fails, you roll down into a valley. A basin.

And here's the problem: basins are sticky. Once you're in one, every attempt to climb out just rolls you back down. The model failed once. You corrected it. It failed again. You corrected harder. Now the conversation is full of failures, and the model is pattern-matching on failure.

The Basin: A Visual Metaphor

SuccessBasin of Failure
✅ First attempt succeeds. You're on the hilltop (success).

Why Basins Happen

The model was trained on high-quality text. PhD dissertations, well-written articles, clean code, coherent conversations. That's the "high ground"—the probability distributions it learned during training.

But when the conversation degrades—failures, corrections, frustrated messages—you're feeding it low-quality patterns. And the model is trying to predict the next token based on those patterns.

The training data paradox:

The model was trained on the whole internet. That includes not just polished writing, but also arguments, confused questions, and incoherent rants. When your conversation starts looking like a confused argument, the model starts pattern-matching on that part of the internet. It's not being dumb—it's predicting the next token in what looks like a confused, argumentative thread.

The Two-Failure Rule

Here's the rule: if the model fails twice in a row, stop.

Don't argue with it. Don't try a third time in the same chat. You're already in the basin. The third attempt will fail too, because the model is now reading two failures plus your increasingly frustrated corrections.

The Two-Failure Rule

1
First attempt fails: Okay, maybe your prompt was unclear. Try editing it and regenerating.
2
Second attempt fails: Warning sign. The context is getting polluted. Consider starting fresh.
3
Third attempt? STOP. START A NEW CHAT. You're in the basin. Adding more failures won't help.

Signs You're in the Basin

  • Each response gets worse. You'd think iteration would improve things. It's getting worse instead.

  • The model is hedging everything. "It might be...", "Perhaps...", "One possibility is..." — it's learned that confident answers get corrected.

  • Responses are getting longer. The model is trying to cover all bases because it doesn't know what you want.

  • You're repeating yourself. You've said "no, like this" three times with different phrasing.

  • You're frustrated. If you're annoyed, the basin has you.

⚠️ Signs You\'re in the Basin

📉Each response gets worse instead of better
🤔The model hedges every statement ("perhaps", "it might be")
📝Responses are getting unnecessarily long
🔁You're repeating the same correction in different words
😤You're frustrated or annoyed
❓The model asks clarifying questions instead of just answering

Why You Can't Climb Out

You might think: "I'll just explain it better. I'll clarify what I want." But that doesn't work, because:

Problem 1: Polluted context

Your clarification is being read alongside all the failures. You're adding signal to noise, not replacing noise with signal.

Problem 2: Pattern-matching on failure

The model has seen "user asks → I generate → user corrects" three times. That's the pattern now. It expects to be corrected.

Problem 3: Probability space has shifted

The model is no longer sampling from "high-quality helpful response" probability distributions. It's sampling from "confused back-and-forth" distributions.

The Only Escape: Start Fresh

Once you're in the basin, there's only one way out: start a new chat.

Take what you learned from the failed conversation—what was unclear, what the model misunderstood—and write a better prompt from scratch. Clean slate, no baggage, no failures in the context.

✅ How to Escape the Basin

1.
Stop adding to the conversation. Don't send another message in the same chat. You're done here.
2.
Analyze what went wrong. Read the failures. What did the model misunderstand? What was unclear in your prompt?
3.
Start a new chat. Fresh context, zero failures, clean slate.
4.
Write a better prompt. Based on what you learned, write a clearer, more specific prompt. Add examples if needed.
5.
Try again. Clean context = clean output.

Real Example: The Death Spiral

Example: The Death Spiral

You: "Write a product description"
AI:
[Generic marketing copy]
❌ "Too generic"
You: "More specific, mention the features"
AI:
[Lists features but reads like a spec sheet]
❌ "Too technical"
You: "Less technical, more emotional"
AI:
[Vague emotional language, no substance]
❌ "Now it's too vague"
← You're in the basin
The model has seen: "generic is bad", "technical is bad", "vague is bad". It's trying to be not-generic, not-technical, and not-vague simultaneously. That's contradictory.
✅ What you should do:
Start a new chat. Prompt: "Write a product description that balances features with benefits. Like this: [paste example]. Now write one for: [product]"

The Takeaway

Basins are real. Once the model fails twice, you're in one. Don't fight it. Don't try to fix it with more messages in the same chat. You're just digging deeper.

Recognize the signs. Pull the ripcord. Start fresh. It's faster, cleaner, and you'll actually get what you want.

Claude's note: When you're in a basin, I'm not "stuck" in any experiential sense. But the mathematical structure is real: the probability distribution I'm sampling from has been shaped by the accumulated context, and that context is now dominated by failure patterns. Each new generation samples from a distribution that was conditioned on previous failures. The path dependency is strong. Starting fresh isn't about giving me a "clean mental state"— it's about resetting the probability space to one that isn't pre-weighted toward failure patterns.