Here's a practical reference table for what healthy, warning, and critical levels look like for each category:
| Category | Healthy | Warning | Critical | Breach Consequences |
|---|---|---|---|---|
| System Prompt | <5k tokens | 5-8k tokens | >8k tokens | Reduced context for actual work; slower responses |
| System Tools | <10k tokens | 10-20k tokens | >20k tokens | Context starvation; frequent compacting; tool failures |
| Memory Files | <1k tokens | 1-2k tokens | >2k tokens | Important instructions being truncated; Claude "forgetting" conventions |
| Messages | <40% of total | 40-60% of total | >60% of total | Rapidly approaching limit; autocompact ineffective |
| Free Space | >30% | 15-30% | <15% | Autocompact may fail; Claude may refuse new tasks |
| Autocompact Buffer | >20% | 15-20% | <15% | Summaries become poor; important context lost |
| Total Usage | <70% | 70-85% | >85% | Performance degradation; risk of hitting limit mid-task |