Context Management
Quick Reference
| Signal |
Action |
| At ~50% context usage |
Manual /compact with focus instructions |
| Approach fails, output quality drops |
/rewind to before the failed attempt |
| New unrelated task mid-session |
/clear or start new session |
| Need to explore/research |
Spawn sub-agent (keeps main context clean) |
| Over 300K tokens used |
Strongly consider new session |
| Auto-compact triggered |
Too late -- quality already degraded |
Context Rot
Context rot is the gradual degradation of output quality as the context window fills up. It typically starts around 300-400K tokens (on 1M token models), but varies by task complexity.
Symptoms
- Claude repeats previous mistakes or ignores corrections
- Output becomes more generic, less project-specific
- Instructions from CLAUDE.md start being ignored
- Code patterns drift from established conventions
Prevention
- Manual compact at 50% -- don't wait for auto-compaction
- Use sub-agents for exploration tasks (research stays in child context, only results bubble up)
- New session for new tasks -- don't carry unrelated context
- Focus compaction:
/compact "keep only the auth implementation details and test results"
Branching Strategy
Every turn in a Claude Code session is a branching point. Choose deliberately:
Current turn outcome:
├── Good result → Continue (carry all context forward)
├── Mediocre result → Two options:
│ ├── /rewind (drop the attempt, retry clean)
│ └── Correct in context (adds noise, sometimes fine for small fixes)
├── Bad result / wrong direction → /rewind (always)
├── Task complete, new task → /clear or new session
└── Need deep research → Spawn sub-agent
/rewind vs Correct In Context
| Scenario |
Use /rewind |
Correct in context |
| Wrong approach entirely |
Yes |
No |
| Small typo or missed detail |
No |
Yes |
| Generated wrong file |
Yes |
No |
| Right idea, minor adjustment |
No |
Yes |
| Multiple failed attempts |
Yes (go back further) |
No |
Compaction
When to Compact
- Good time: ~50% context usage, between logical task boundaries
- Bad time: Mid-implementation, during debugging, during a complex multi-step plan
How to Compact Well
# BAD: no guidance
/compact
# GOOD: focused compaction with retention instructions
/compact "Keep: current implementation plan steps 3-7, test results from auth module, the API schema we agreed on. Drop: exploration of rejected approaches, file reads from unrelated modules."
Compaction vs Clear vs New Session
| Method |
Preserves |
Loses |
Best For |
/compact |
Summary of work, key decisions |
Details, exact code snippets |
Mid-task context reduction |
/clear |
Nothing (clean slate) |
Everything |
Starting fresh task in same session |
| New session |
Nothing |
Everything |
Completely new task or severely degraded context |
Sub-Agent Isolation
Use sub-agents to keep the main context window clean:
| Task Type |
Use Sub-Agent? |
Why |
| File exploration / grep |
Yes |
Search results are bulky, only need the answer |
| Code review |
Yes |
Review details stay in child, summary comes back |
| Research / investigation |
Yes |
Exploration paths stay isolated |
| Implementation |
No |
Need to see and verify the changes |
| Debugging |
Depends |
Simple bug: no. Complex investigation: yes |
Pattern: Research Then Implement
1. Spawn sub-agent: "Research how auth middleware works in this codebase"
2. Get summary back (main context stays clean)
3. Implement based on summary (in main context where you can verify)
Anti-Patterns
| Anti-Pattern |
Problem |
Do Instead |
| Never compacting |
Context fills up, quality degrades silently |
Compact at ~50% usage |
| Compacting too often |
Lose important context |
Compact at natural boundaries |
| Waiting for auto-compact |
Model is least intelligent during auto-compact |
Manual compact before it triggers |
| Saying "no try again" after bad output |
Failed attempt stays in context, pollutes future attempts |
/rewind then re-prompt |
| Reading every file "just in case" |
Wastes context on irrelevant content |
Search first, read only what's needed |
| Running all research in main context |
Exploration results bloat context |
Use sub-agents for research |
| Continuing after multiple failures |
Context is poisoned with failed approaches |
/clear or new session |
