💭 in your head · nothing to type
Picture a witness in a courtroom delivering a confession: "Yes, I gave too little context. Yes, I buried the critical instruction in the middle. Yes, I dropped three old file versions into the context window."
Run it in your head: you're that witness right now. Picture a moment when Claude started forgetting early details mid-conversation, or answered confidently but wide of the mark. Which of the prompting sins are you ready to confess? Hold it — no self-judgment, just the facts of the case.
💭 in your head · nothing to type
Now picture a different scene: you're a ship's navigator looking at a chart and seeing that the vessel has drifted off course.
Hold that moment — where you are now versus where you should be. A navigator doesn't panic: they measure the drift, name it, and apply a correction. Which drift — context rot, lost-in-the-middle, distractors — knocked you off your heading? And what correction would you plot right now?
The longer the conversation, the worse the model handles early information.
Fix: make a summary and start a new session (/compact or a new chat with the summary).
The model over-weights the beginning and end of the context. Critical stuff in the middle gets ignored.
❌ Bad: [Intro] → [CRITICAL — in the middle] → [Conclusion]
✅ Good: [CRITICAL — up front] → [Details] → [Repeat CRITICAL]
Similar fragments in the context throw the model off the correct answer.
❌ Bad: Drop 5 files, 2 of which are old versions
✅ Good: Only relevant files + "Ignore old versions"