How to stop AI context rot using the 60% rule

Executive overview

AI tools like ChatGPT, Claude, and Gemini degrade within a single conversation as their context window fills. Instructions get ignored, facts change, and the model contradicts itself — all predictable, all preventable.

The 60% rule: performance starts degrading noticeably past 60% of the model's context window. Past 95%, degradation is rapid. Four tactics extend the useful window.

Start a new conversation at the 60% mark rather than fighting a degrading one.

The four warning signs

• AI stops following instructions it was given earlier in the thread
• AI contradicts a decision or recommendation it made earlier
• Facts you provided (prices, stats, names) are changed or dropped
• Claude shows an automatic compaction notice ("organizing my thoughts") — a signal the thread hit ~90–95% capacity

Knowing your token budget

• ChatGPT UI: ~60k tokens (~120 pages / half a novel)
• Claude: 200k tokens (~1.5 novels)
• Gemini: 1M tokens (~7–8 novels)
• Use Google AI Studio (ai.studio.google.com) to measure token usage — drag a file or paste a conversation to see its exact token count

Tactic 1: the handoff summary

• When approaching 60% capacity (or seeing warning signs), open a fresh chat tab
• Ask the degrading conversation to summarize four things: what was covered, key decisions made, where the to-do list stands, and what the next AI should do first
• Paste that summary into the new tab — the new AI picks up exactly where the old one left off
• This gives a clean memory and avoids compounding degradation

Tactic 2: strategic file choice

• Text/Word files: cheap — low token cost, safe to upload freely
• PDFs: moderate — cost rises with size and complexity (images, diagrams); check anything over 120 pages in AI Studio
• Images: expensive — assume high token cost by default
• Excel files: variable — complex multi-tab sheets strain AI; export only the relevant tab as CSV
• Video: very expensive — anything over 2–5 minutes fills most context windows; use Gemini if required
• Don't upload the full file if only a section is needed — slice it to reduce noise and token use

Tactic 3: experimentation and task sizing

• Push AI with complex tasks to learn its current capability ceiling
• If the AI succeeds, update your intuition of what's possible
• If it fails, break the task into subtasks and run each in a separate tab with a fresh AI
• Parallel tabs with separate AIs avoid cross-contamination and share the token load

Tactic 4: strategic in-thread summaries

• For users who won't switch threads (ChatGPT and Gemini primarily)
• Every 5–10 exchanges, ask the AI to write a brief summary: achievements so far, key decisions, current to-do status, next steps
• Keep the summary brief — a verbose summary wastes the memory it's trying to conserve
• Ask the AI to continue from the summary; repeat the cycle throughout the conversation
• Extends the useful thread life but doesn't match the clean reset of the handoff approach