Claude Clean Context — Security skill for Claude Code

claude-clean-context

[](https://github.com/mohan-n-swamy/claude-clean-context/actions/workflows/ci.yml)


Keep Claude Code's always-loaded context lean by demoting stale memory pointers and auditing skill descriptions.

This repo gives Claude Code users a small, deterministic toolkit for cleaning the context that loads before every prompt. It focuses on the two movable layers that tend to grow over time: auto-memory indexes and skill descriptions.


*A real `memory-distill.py` dry-run on the bundled example: 15 pointers in, 6 out. Dry-run by default — it shows exactly why each pointer would be demoted and writes nothing.*

The Problem

Claude Code starts each session with standing context already loaded. Some of that context is useful: durable rules, corrections, and project-specific guard rails. Some of it quietly turns into stale status: old session snapshots, shipped work notes, and skills that are no longer relevant to the current project.

That costs tokens and dilutes signal. The more stale context is always present, the easier it is for live rules to get crowded out. See [docs/the-problem.md](docs/the-problem.md) for the full explanation.

The Principle

A thing earns standing context ONLY if it must be present before you know you need it. Everything else loads on demand.

Memory should hold guard rails, not status logs. Skills should stay eager only when the current project plausib