Manual Context Compaction

I’ve been using Claude Code for about two months daily. Here’s one lesson I learned along the way to prevent going in circles, improve output quality and fight context rot when driving it.

Running long, serious sessions will often result in auto compaction, where it will run out of context length and force a summarization step before continuing onto the next prompt. When that happens, it means your average context utilization is over 50%, reaching 100%. Optimal context utilization for best results is 20-40%.

Garbage in, garbage out

If everything in the session context is relevant to your upcoming task, then running /compact may work, but there’s a better way. Instead of moving a large context from one session into a smaller context in another, documenting into a markdown file will produce durable memory that can be shared with other developers. Try the following:

Document the research you just did into a markdown file

Write down what we tried and why it didn't work, and possible next steps into a file

Turn this discussion into a spec file for an MVP

Then, /clear context, and ask it to take a look at the auth architecture docs to recall memory. Now, you have a super compact, reusable context that can be committed into the codebase.

Solving a puzzle is easier without the wrong pieces

When working with a long-running task that spans multiple sessions, it’s helpful to keep a work plan, get Claude to update between each step, and clear context without fear of amnesia.