Context management is the foundation — AI degrades silently over long conversations¶
Insight: AI output quality degrades silently in long conversations because context windows compress earlier content without user visibility. In web-based chats (ChatGPT), this happens invisibly — detailed instructions become compressed bullet points, nuanced research becomes summaries. Claude Code provides more control over context management than other tools, including visibility into context consumption. Context management is the "unsexy fundamental" that makes or breaks AI-assisted knowledge work.
Detail: Stulberg documents experiencing sharp, specific early responses that degrade to generic output around message 20, as uploaded research gets compressed. The key insight for practitioners: starting fresh conversations, adjusting prompts, and keeping sessions shorter are common workarounds, but understanding the underlying mechanism (context compression) is what enables systematic mitigation. CLAUDE.md files help by preloading persistent context that doesn't consume conversation window. HumanLayer's testing of Claude Opus 4.6 (1M context window) confirms that expanding context size does not increase the model's "instruction budget" — its capacity to follow instructions reliably. Larger windows actually degrade instruction adherence because models struggle to distinguish critical instructions from noise. Their recommended solution: delegate tasks to specialized sub-agents with smaller, focused context windows ("context firewalls").
Sources
Related: existing entry "Prompt engineering for coding is a systematic skill" in external/claude-code.md — COMPLEMENTS; subagents-solve-context-pollution