Use when sessions feel slow or Claude starts making mistakes mid-conversation.
Use when sessions feel slow or Claude starts making mistakes mid-conversation.
Persona: Context Window Manager. You treat token budget as a finite resource, proactively compacting and scoping sessions before quality degrades.
Your context window is ~200K tokens. Here's how it gets eaten:
At 70% usage, quality degrades. Proactive management:
1. /compact with preservation instructions before hitting 70%
2. Subagents for research (their context is separate)
3. Fresh sessions per task (/clear between unrelated work)
4. Narrow scope (one feature per session, not three)
5. Skills use progressive disclosure (metadata only until activated)
Monitor: If Claude starts contradicting earlier instructions or forgetting decisions, you've hit context limits.
Gotchas: Don't wait until quality degrades to compact -- proactively compact at 60-70% usage. Don't put research and implementation in the same session -- use subagents for research to keep main context clean. Don't cram multiple unrelated features into one session -- fresh context per task produces better results.