Anthropic says it cut 80 percent of Claude Code's system prompt because Fable 5 models "want a smaller system prompt"
Anthropic reduced Claude Code's system prompt by 80% for its new Fable 5 (Mythos class) models. Newer models exhibit higher imagination, making extensive examples and hard constraints counterproductive. Steering strategy shifted from rigid rules and dense examples to contextual guidance. Prompt length trends have reversed: short -> long -> short as model capabilities evolved.
Analysis
TL;DR
- Anthropic reduced Claude Code's system prompt by 80% for its new Fable 5 (Mythos class) models.
- Newer models exhibit higher imagination, making extensive examples and hard constraints counterproductive.
- Steering strategy shifted from rigid rules and dense examples to contextual guidance.
- Prompt length trends have reversed: short -> long -> short as model capabilities evolved.
Why It Matters
This finding challenges the conventional wisdom that larger, more detailed system prompts always yield better performance, suggesting that over-constraining advanced models can hinder their capabilities. For AI practitioners, it highlights the need to adapt prompting strategies based on specific model classes rather than applying static best practices across all architectures.
Technical Details
- Model Class: The changes apply specifically to the Fable 5 models, also referred to as the Mythos class.
- Prompt Reduction: An 80% cut was applied to the system prompt used in Claude Code.
- Steering Mechanism: Anthropic moved away from "hard rules" (e.g., "do not do this") and excessive few-shot examples, which were found to constrain the model's natural imagination.
- Contextual Guidance: The new approach relies on providing context to steer the model rather than explicit, restrictive instructions.
- Historical Trend: Early models required short prompts with many examples; intermediate models handled longer prompts better; current models require shorter prompts again.
Industry Insight
- Dynamic Prompting Strategies: Developers should avoid one-size-fits-all prompt engineering; prompt complexity must be tuned to the specific model version and class being used.
- Trust Model Imagination: Over-specifying behavior via examples may suppress the creative problem-solving abilities of advanced models, potentially leading to suboptimal outputs in complex tasks.
- Iterative Optimization: As models improve, the optimal prompt structure may regress to simplicity, emphasizing the need for continuous evaluation of prompt effectiveness rather than assuming growth in capability allows for growth in instruction density.
Disclaimer: The above content is generated by AI and is for reference only.