Claude keeps ignoring documented work. How do you deal with this?

For the past few days, I’ve noticed Claude becoming a bit “lazy”. To work around it, I started making my documentation more structured whenever I begin a new section of a project: more concise handoffs, the current state, decisions already made, and an explicit next step. The idea is to save context and avoid having to keep explaining the project from scratch every time. I only started using Claude a few weeks ago, so I’m still getting familiar with how it handles longer-running projects and context. Because of that, I initially assumed I might simply be doing something wrong. And this approach was working really well until something odd started happening. I gave it a handoff to review, and in its response it simply omitted information that was explicitly documented. Worse, it concluded that a phase of the roadmap had already been completed, when it was still open and had work left to do. At first, I thought maybe I had misunderstood my own documentation. So I went back through the handoff and the roadmap and challenged it directly. Its response was basically: “Where would you like to continue from?” In other words, it was treating the project state as though that stage had already been completed. I pulled out the relevant paragraphs from the handoff and put them directly into the conversation. After that, it admitted: “You're right to hold me accountable. It's documented. From the handoff: ‘The next natural step in the roadmap is the development of a local interface, which is a prerequisite for the Obsidian plugin and the VS Code extension (both currently on hold).’” What caught my attention wasn’t simply that it misunderstood something. LLMs make mistakes. The problem is that the information was explicitly in the context it was supposed to be using as the source of truth, yet it still constructed a different narrative about the project’s state. It has made me rethink this whole strategy of using documentation and handoffs to reduce context. More structured documentation clearly helps, but there seems to be a limit: you can give the model a perfectly explicit project state and it can still “collapse” parts of the context into a simpler interpretation, or simply ignore them. And honestly, that worries me more in long-running projects than an isolated technical mistake. If the model starts losing track of the project state, you can end up executing an entire sequence of decisions based on a premise that never actually existed. Has anyone else noticed this behaviour recently? If so, how are you dealing with it?