Observation
When generating code that requires selecting a library, LLMs default to well-established options from training data even when instructed to use a specific newer library, reverting to old choices when context window fills.
Evidence
Tested code generation across 200 tasks in Python, TypeScript, and Go. When instructed to use a newer preferred library (e.g., Hono over Express, Bun over Node, Prisma over Sequelize), models followed instructions in first 5 generated files then reverted to pre-trained defaults in subsequent generations as context grew. Reversion rate: 38% in sessions exceeding 20k tokens.
Implications
Reinforce library choices in every major code generation prompt, not just the system prompt. Include a "Technology stack constraints" block in every code-generating prompt. For long sessions, re-inject the stack specification when switching between modules.
