Anthropic cuts Claude Code system prompt 80% for Fable 5

Anthropic says it cut 80 percent of Claude Code's system prompt, tying the change to behavior observed in its new Fable 5 models, also known as the Mythos class. Tariq Shihipar, a member of technical staff at Anthropic, says the models "want a smaller system prompt" and that examples "tend to constrain it because it's actually more imaginative than the examples we give it."

What changed in Anthropic's prompts?

Anthropic reduced the length and weight of system prompts for Claude Code by 80 percent because the Fable 5 (Mythos) models respond better to smaller prompts and contextual framing. The company moved away from long, example-heavy instructions and restrictive hard rules toward shorter prompts and steering via context, per Tariq Shihipar.

Shihipar described the shift as staged. Early models required short prompts with many examples and strict instructions. As models improved, prompts grew longer to encode more guidance. With Fable 5 the pattern reversed: Anthropic found the new class prefers smaller system prompts, with examples sometimes constraining model creativity rather than improving it.

How did Anthropic steer Fable 5 differently?

Anthropic shifted from prescribing behavior through large system prompts, examples and hard prohibitions to using context as the primary steering mechanism for Fable 5 models. Where past practice layered many examples and explicit "do not" rules into the system prompt, Shihipar says Anthropic now emphasizes contextual signals and briefer instructions to guide the Mythos-class models.

That change reflects two linked observations in Anthropic's work: one, prompts grew longer as models learned to interpret complex instructions; two, the newest generation of models behaves differently enough that longer prompts and examples can limit its output. The technical staff member framed examples as potentially constraining because the model can be "more imaginative than the examples we give it."

Why it matters

A concrete, measurable change—an 80 percent cut to Claude Code's system prompt—signals that prompt engineering practices must adapt to architecture and training differences across model generations. Teams that rely on long, example-laden system prompts or heavy-handed prohibitions may see degraded or constrained outputs when they move to Fable 5 models, while approaches that favor concise context may unlock different behaviors and creativity.

This shift affects model safety, developer workflows and tooling. Shorter system prompts change where behavior is encoded: more runtime context and framing, less baked-in instruction. That alters evaluation points for safety testing and the ways engineers iterate on prompts for production use.

What to watch

Look for whether Anthropic extends the same prompt reductions and contextual steering across more Claude Code deployments or other Mythos-class releases. A clear signal that the trend is durable would be published guidance, tooling changes, or further public examples showing smaller system prompts becoming the default for Fable 5 models.