Notes ·
Context Engineering Has to Evolve with the Model
Thariq Shihipar writes for Anthropic about how the company changed its approach to context engineering for the Claude 5 generation of models.
Anthropic says it removed more than 80 percent of Claude Code’s system prompt for its newer models without any measurable loss on its coding evaluations. The older prompt had accumulated rules, repeated instructions and defensive constraints that were once useful but increasingly forced the model to work through conflicting guidance.
The newer approach is much lighter. Instead of prescribing behavior for every situation, Anthropic recommends allowing the model to use judgment, designing expressive tool interfaces, loading specialized guidance only when it is needed and keeping files such as CLAUDE.md focused on repository-specific gotchas rather than obvious information.
That is useful for anyone working with multiple LLMs. We use these models heavily at work, and it is easy to assume that a prompt or agent configuration that performs well with one model should work equally well with another. It often does not.
Different models respond differently to examples, repetition, rigid rules, tool descriptions and the amount of context loaded upfront. Even newer generations from the same provider may require a different approach than their predecessors.
The durable skill is not memorizing one perfect prompt format. It is understanding how each model interprets context, then testing and adapting the surrounding system as the models change.