Pattern: Prompt Engineering Has Inverted, and Labs Are Now Deleting Constraints Instead of Adding Them
Anthropic·high signal
The dominant technique for two years was accumulating rules: longer system prompts, longer CLAUDE.md files, more numbered do/don't lists. Anthropic's 80% deletion reverses that, and the stated reason is not that the rules were wrong but that they contradicted each other inside a single request and suppressed exploration where exceptions were appropriate. The replacement pattern is state the goal, let the model judge, and load context progressively. If you have been optimizing by adding, your instruction files are now a liability rather than an asset.