What happened
- Anthropic advises dropping phrases like 'think hard' from prompts
- Opus 5.5 performs better at Medium effort level than previous models at High
- Users should describe desired results in a single message rather than giving instructions one at a time
Why it matters
Anthropic's advice on prompting habits for Claude Opus 5.5 highlights how outdated practices can hinder model performance. By adjusting effort levels and prompt structures, users can unlock better efficiency and results from the new model, which is now the default on paid plans.
The Elephant take
๐ ๐ฆ Anthropic's 'think hard' advice is a relic of older models. Opus 5.5 thrives on simplicity, not overthinking. The company's tests show that dropping those phrases can speed things up without sacrificing quality.
Who should care
- AI developers
- Business users
- Model administrators
What to do next
- Drop 'think hard' phrases from prompts
- Start at Medium effort level for Opus 5.5
- Describe desired results in a single message
- Use diagrams and tables as images instead of typing them out
Keep in mind
Anthropic's recommendations are based on internal tests, which may not reflect real-world usage across all user scenarios.