Anthropic Shares Guide to Optimizing Claude Prompts

As businesses become heavily reliant on generative AI, many users have found that the popular Claude model appears to get “dumber” over time. AI developer Anthropic says the problem is not a lack of model capability, but developers piling too many conflicting rules into prompts and creating contextual confusion. The company has recently stepped up efforts to promote better collaboration practices, emphasizing reorganized structures and XML tags to separate information rather than simply making prompts longer.

Anthropic formally released a practical guide at its Code with Claude technical event in July 2026. The company said caching can cut costs by 90%, although three engineering failures in April, including a caching bug, had caused performance to decline. Those issues have since been fixed. The latest guide recommends that developers build evaluation systems with test cases and restructure information using XML tags.