Context contamination ("Poison.md") as a major risk in agentic coding sessions¶
Insight: Context contamination is a significant but underappreciated risk in agentic coding. Every action pairing in context creates potential training patterns — e.g., telling Claude to "update code" followed by "deploy" can cause Claude to associate all code updates with immediate deployment. Practitioners must actively scan for dangerous associations before they take root, especially in agentic tasks where the consequences of misinterpreted patterns compound.
Detail: This is the negative side of CLAUDE.md's instruction hierarchy: while persistent context is powerful for consistent behavior, it also means contaminated context persists and compounds errors throughout a session.
Sources
- ClaudeLog by InventorBlack — "Poison.md" (2026-01-15)