I built a prompt compressor that cuts token usage and help prevents Prompt Decay in long sessions

If you use Claude (or any LLM) for long sessions you've probably hit this at some point. You're deep into a coding project or research session, halfway through finishing what you started, and you notice the AI starting to take shortcuts, answers that are clearly rushed, constraints you set early being ignored, and you end up sending more messages just to correct it, which eats into your limits for no reason. This is Prompt Decay. It's not a Claude bug, it's just what happens when your original...