Skip to content

AI Anti-pattern

The Context Cliff

Long conversations degrading until the model loses the thread.

Why it matters

Large language models degrade over long conversations. The user re-explaining the original goal is a symptom, not a fix. Every re-statement costs tokens and signals the conversation has already exceeded its effective length.

  • Extended context degradation documented across frontier models (Anthropic technical reports)
  • Mid-conversation re-stating as a fluency signal (VibeWorks v14 analyzer)

What this looks like

Conversations exceeding 15 turns where the user re-states the original question, the model contradicts its own earlier outputs, or the user asks 'what did we decide?'

This is a generic example. Run your own analysis to see instances from your actual AI conversations.

See this in your own conversations

Paste a conversation with Claude, ChatGPT, or Gemini to find out if you show signs of The Context Cliff.

Run your own analysis →