How do you deal with Claude Code context going stale?

[mvoutov]

This is the problem that led to building aspens. Curious how others handle it.

The journey most people seem to go through:

1. Write a CLAUDE.md — works for a while, then the codebase grows, the file gets too long, and it goes stale
2. Break it into granular files — better results, but maintaining all those files is cumbersome. Every few PRs, the docs drift from the actual code
3. Give up and let Claude search — Claude spends half its tool calls Bash/Grep-ing through the codebase, burns through tokens, and still builds components from scratch instead of reusing existing ones

The specific problems we kept hitting:

• Claude builds a full component from scratch even though one already exists in the codebase
• Claude writes a new utility function when there's already one that does the same thing
• The bigger the codebase gets, the more fragmented and inconsistent it becomes
• UI components especially — Claude creates new ones instead of using your existing design system

aspens tries to fix this by auto-generating skill files from your import graph and keeping them updated on every commit via git hook. But we're curious:

• What's your current approach? Manual CLAUDE.md? Granular docs? Something else?
• What breaks down first? Staleness? File size? Claude ignoring the context?
• What would actually solve it for you?