If Claude starts sharp and gets sloppier the longer you work, you're not imagining it. Everything in the session sits in its context window: every message, every file it read, every command output. As that fills up, the model has more noise to sort through and the early stuff gets pushed aside.
Anthropic says this plainly in the Claude Code docs: performance degrades as context fills, and Claude may start forgetting earlier instructions or making more mistakes. That's when the guessing and hallucinating start.
Keep reading
The rest is one email away. Then every guide is open.
Put your email in once and you can read this guide and every other one on the shelf, including the ones from reels I haven't posted yet. I'll send you the link too, so you can find it again.
Still to come
- The 3-move fix
- The exact prompt
- In Claude Code
- In claude.ai