CC4E #3: Why AI Gets Dumber The Longer You Talk To It
Author: Hannah Stulberg | Published: 2026-01-20
Context management is “the plumbing” – the unsexy fundamental that nothing else works without.
Context 101 (Filing Cabinet Metaphor)
Each chat session is one drawer in a filing cabinet:
| Concept | What it means |
|---|---|
| Context | The information available to the AI at any point. If it’s not in the drawer, Claude doesn’t know about it |
| Context window | The size of the drawer (~200K tokens / ~200K words / ~3 novels). Fills faster than expected |
| Compaction | At 70-80% full, the AI summarizes to make room. Original text replaced by compressed summary. Nuance is lost each time. In web tools (ChatGPT), this happens silently |
| Thinking room | AI uses the same space for reasoning and storage. More context loaded = less room to think = lower quality output |
Why Claude Code Is Different
Web-based AI tools (ChatGPT, claude.ai, Gemini) give almost no control over context. Claude Code provides:
- Visibility – status line shows context usage
- Control – manual compaction with
/compact - Structure – parallel sessions, session persistence, background agents
Five Context Management Techniques
1. Status Line
- Shows tokens used vs. total (e.g.,
47K / 200Kor24%) - Three zones:
- 0-50% – work freely
- 50-70% – getting crowded, consider clearing before complex work
- 70%+ – compact manually NOW or auto-compaction decides for you
- Enable via
/statusline - Only available in the CLI, not VS Code/Cursor extensions
2. Manual Compaction
Auto-compaction lets Claude decide what to keep. Manual puts you in control.
- Ask Claude: “What do you currently have in context?”
- Decide what to keep (current docs, active requirements, decisions) vs. drop (old drafts, abandoned tangents)
- Have Claude draft the
/compactcommand for you - Run
/compactwith specific instructions. Context typically shrinks to ~15-20% - Use Ctrl+O to verify what was kept
3. Parallel Sessions
- Each session gets its own drawer – prevents context contamination
- Three separate employees each focused on one task beats one employee juggling three tasks
4. Session Management
- Name and resume sessions to preserve context across time
- Without this, every new session starts with an empty drawer
5. Background Agents
- For related side tasks that need your current context but shouldn’t clutter the main session
- Key distinction: parallel sessions = unrelated tasks (each starts empty). Background agents = related tasks that need current context
Key Insight
AI output quality is directly tied to how much context space is available, not just what you prompt. Proactive context management separates mediocre AI output from consistently great output.