← Back to the overview
Local · Private
The context audit
Claude Code loads your instruction files, your memory index and every skill's description before your first prompt. It is a toll paid on every request of the session, and it is normally invisible — which is why it grows.
What it measures
- ✓Eager vs lazy — a 300 KB memory directory can cost less than one bloated
AGENTS.md, because only one of them is read before you type. - ✓Was it ever used? — skills annotated 45× or never, mined from your own transcripts rather than guessed from the file.
- ✓What-if — tick rows and watch the gauge and the cost drop, before changing anything on disk.
- ✓Grade and drift — A–F per project, and “+2.4 KB in the last 7 days”, because the number that matters is the direction.
- ✓Between projects — duplicated memory directories and dead project directories, which are invisible from inside either one.
It measures; it does not edit
Every finding comes with one plain sentence and the concrete fix, and the only action on the window is Copy cleanup prompt — which hands the paths and the numbers to Claude Code. The scan reads only sizes, names, timestamps and frontmatter from ~/.claude, never the contents of your memories, instructions or skills.