Why Claude gets worse the longer you use it.

Claude gets vaguer the longer you run a single chat, and it isn't your prompting. It's the context window, the model's fixed working memory. Three two-minute fixes keep it sharp.

It's not your prompting. It's the context window.

Claude has a fixed working memory. Everything connected and everything typed has to fit inside it. The longer a chat runs, the more of that space is already spoken for, and the careful context from message one is the first thing to blur as the model compresses to make room.

The model has a fixed amount of room. Spend it on your problem, not on clutter you forgot was switched on.

Three fixes. Two minutes each.

01

Turn off unused connectors

Every connector switched on reserves working memory the moment it activates, before you've typed anything. Keep only what this task needs.

02

One Project per topic

A single mega-chat blends everything together. Use a separate Project per client or deal so context stays about one thing.

03

Save your context once

Put standing context into Project instructions instead of retyping who you are and what you're working on at the start of every chat.

The Project instructions template

Standing instructions for this Project. Who I am: [role]. Play this role: act as a [expert type] who has done this many times. Always assume these facts: [key specifics]. Never guess: if uncertain, say so. Ask first if something is missing.

What it still gets wrong.

i.

None of this is about prompting harder

It's about protecting the working memory the model thinks in, so message fifty is as clear as message one.

ii.

Clutter compounds silently

You won't notice the window filling up until the answers start getting vaguer.

Next Playbook