XNews

Prompt Engineering: Less Is More for Advanced AI Models

Prompt Engineering: Less Is More for Advanced AI Models

Claude Code creator Boris Cherny recommends deleting CLAUDE.md, Skills, and Hooks every six months to test what the model can do unaided, especially after upgrading to Opus 5, as old rules may become limitations. The Claude Code team revealed they cut the system prompt by 80% for Fable 5 and Opus 4.8. Key findings: removing few-shot examples improved creativity because models now surpass their examples; prohibitions like 'don't do X' create strong conflicting impulses and should be replaced with context; every instruction must be 100% correct with no exceptions, or it should be softened. This approach relies on frontier model judgment and does not apply to smaller models like Haiku, which still need detailed prompts. The team confirmed that Claude now writes prompts for subagents, and that frontier intelligence is almost always worth the token cost even on hard problems. Additional discussion covered Claude Tag's 65% PR automation rate, automated code review for non-core changes, evaluation suites ensuring new models are strictly better, and a shift toward fewer, non-overlapping tools.

Key insights

Few-Shot Examples Have Become Counterproductive for Frontier Models

The Claude Code team found that removing examples was 'extremely effective' because models are now more creative than the examples given. This contradicts years of prompt engineering best practice and suggests a fundamental shift in how frontier models process instructions.

Prohibitions Create Conflict Rather Than Constraint

Hard 'do not do X' rules act as strong impulses that conflict with user instructions, causing confusion. The team moved toward fewer hard constraints, more background context, and fewer total instructions, trusting the model's judgment.

Prompt Minimalism Requires Regular Auditing

Boris Cherny's advice to periodically strip all custom instructions and test bare performance reflects the reality that prompts optimized for older models may actively hinder newer ones. This turns prompt maintenance into an ongoing process rather than a one-time setup.

Model-Dependent Prompting Is Now a Necessity

The approach works for Opus/Fable-level models but not Haiku. This necessitates maintaining separate system prompts per model tier, adding complexity to multi-model deployments.

Source posts