Anthropic says it cut 80 percent of Claude Code's system prompt because Fable 5 models "want a smaller system prompt"

Details

Anthropic has reduced the system prompt for its Claude Code coding assistant by 80%, revealing a meaningful shift in how frontier models are best directed. Tariq Shihipar, a member of technical staff at Anthropic, explained the change is tied to the new Fable 5 models (also called Mythos class) — a generation that processes instructions fundamentally differently than predecessors. The core finding: more examples and tighter rules no longer improve performance and can actively degrade it. Shihipar noted examples 'tend to constrain [the model] because it's actually more imaginative than the examples we give it.' Anthropic has shifted away from prescriptive 'do not do this' instructions toward context-driven steering for these newer models. Shihipar described an interesting evolution pattern: early models needed short prompts with many examples and restrictive rules. As models improved, prompts grew longer to leverage better comprehension. Now, with Fable 5/Mythos, prompts are shrinking again — because the models' reasoning capacity exceeds what explicit examples can demonstrate. For practitioners building on Claude: this suggests that densely specified system prompts for the latest model generations may be setting a capability ceiling rather than raising a floor. The article is thin on technical benchmarks and is primarily a paraphrase of Shihipar's public post, so treat specifics as directional rather than verified.