Context Window Limits
"I compacted but output was still terrible." Sound familiar? The hidden problem: quality degrades at 20-40% capacity, not 100%.
The Reality
200K tokens doesn't mean 200K tokens of quality work. Everything accumulates: your messages, Claude's responses, file contents, tool results. By the time you're at 50% capacity, quality is already noticeably worse.
Compaction doesn't magically restore quality. It reduces token count, but the summarized context is still noise compared to a fresh start.
Practical Techniques
Scope Conversations
One feature per conversation. Don't try to do auth, then payments, then notifications in the same thread. Start fresh for each major task.
External Memory
For complex tasks, use SCRATCHPAD.md or plan.md files.
Write important decisions and context there. Claude can read it in new sessions.
Copy-Paste Reset
When context gets bloated: copy the important stuff, run /compact,
then /clear, then paste it back. Fresh context with preserved knowledge.
Know When to Clear
/clear is better than struggling through confusion. If Claude seems
confused or keeps making the same mistakes, clear and start fresh with just the
essential context.
The Mental Model
Claude is stateless. Every conversation starts from nothing except what you give it. There's no memory between sessions. There's no learning from past interactions.
This isn't a limitation—it's a feature. It means you have complete control. But it also means you need to plan accordingly. Important context must be explicitly provided, every time.
Key Takeaways
- Quality degrades at 20-40%, not at 100%
- Compaction helps but doesn't restore full quality
- One feature per conversation keeps context fresh
- Use external files for persistent context across sessions
/clearis a feature, not a failure