Do you feel the LLM chat getting dumber? You just can't point to where it turned. Someone asked me last week why my AI outputs come out sharper than theirs. We use the same model. Same tools. So why the gap? Different habits. One of them: I don't run complex work in a single chat until it falls apart. The longer a chat runs on a hard task, the worse the output gets. I saw it first in Claude Code and Codex. Turn one, the model catches an obvious error. Turn nine, it walks right past the same mistake. This is measured now, not just a feeling. Microsoft and Salesforce ran over 200,000 simulated conversations across 15 models. When the task came in piece by piece across turns instead of all at once, performance dropped 39%. The models didn't lose smarts. What broke was consistency, the spread between their best answer and worst answer on the same task more than doubled. A separate study from Chroma found a second kind of decay, this one from length alone. More tokens in, weaker output out, long before you hit any limit. Put both on an underwriting model or a financial analysis. Long session, lots of back and forth, the model's own earlier guesses piling up in the middle where it tends to forget them. The chat that nailed it at the start is quietly missing things by the end. You rarely catch the moment it turned. So I built a fix. A handoff skill. When a session gets heavy, it pulls everything that matters into a tight summary and hands it to a fresh chat. Clean window. Full context. The work picks up where it left off. Could you just open a new chat and paste a summary yourself? Sure. But you forget what mattered, and detail leaks out in the retelling. The skill does it the same way every time. Claude Code has /compact for this. Cowork doesn't. So I made my own. The models will probably solve this themselves down the road. Memory gets good enough, the problem goes away. We're not there yet. Want the skill or the prompt to build your own? Comment "handoff" below and make sure you're connected to me so I can send it over.