Anthropic says it has made a major change to how Claude Code is guided: the system prompt has been reduced by 80 percent. The reason, according to Tariq Shihipar, a member of technical staff at Anthropic, is that the new Fable 5 models behave differently from earlier systems.
Shihipar describes the change as part of a broader shift in AI model steering. More instructions, more examples, and more restrictions are no longer automatically the best way to get better results.
A shorter prompt for a different kind of model
The central claim is simple: the new Fable 5 models, also known as the Mythos class, do not need the same kind of dense system prompt that previous models required. Anthropic says Claude Code's system prompt was cut by 80 percent because this new class responds better to a smaller guiding layer.
Shihipar puts it directly: "Most recently we found this new class of models want a smaller system prompt." That is a notable reversal from the common assumption that a stronger prompt is usually a longer one.
For Claude Code, the system prompt is part of the hidden instruction layer that shapes how the model behaves. Reducing it by 80 percent signals that Anthropic now sees excessive steering as a possible limit, not just a safeguard or improvement.
Why examples can become a constraint
One of the most important details in Shihipar's explanation is about examples. Earlier approaches often relied on many examples to show a model the desired pattern. For the Fable 5 models, Anthropic has found that examples can narrow the model's behavior too much.
Shihipar says examples "tend to constrain it because it's actually more imaginative than the examples we give it." In plain terms, the model may be capable of producing useful outputs that go beyond the patterns shown in the prompt. If the prompt contains too many examples, the model may stay too close to them.
This does not mean examples are useless. The source does not say Anthropic has abandoned them entirely. The point is narrower and more specific: for this new model class, more examples are not always better, because they can limit the range of responses the model is able to explore.
From hard rules to context
Anthropic is also changing the kind of guidance it uses. Instead of leaning on direct prohibitions such as "do not do this," Shihipar says the company now tries to steer Fable models through context.
That is a different philosophy of control. A hard rule tells the model what behavior to avoid. Context gives the model more information about what kind of response is appropriate. For a model described as more imaginative than the examples provided to it, context may leave more room for useful judgment while still shaping the output.
The source frames this as a practical adjustment, not a slogan. Anthropic appears to be responding to observed behavior in the Fable 5 models rather than applying one fixed prompting method across all model generations.
How prompting has changed over time
According to Shihipar, this did not happen all at once. The pattern moved through stages as models changed.
- Early models needed short prompts with many examples and restrictive instructions.
- As models became better at understanding prompts, those prompts became longer.
- With Fable 5, Anthropic says prompts are getting shorter again.
This history matters because it shows that prompt design is not static. A prompt that worked well for one generation of models may be poorly matched to another. The same tool can require a different steering strategy as its underlying model changes.
For teams building with AI systems, the practical lesson is not simply to write shorter prompts. The lesson is to test whether each instruction is still doing useful work. If a model already understands the task, extra rules and examples may add friction instead of clarity.
What this means for Claude Code users
The source does not describe specific product changes beyond the 80 percent reduction in Claude Code's system prompt. Still, the change points to an important direction for AI coding tools: stronger models may need less micromanagement in their hidden instructions.
Claude Code is being steered in a way that assumes the Fable 5 models can use context more effectively. That could make the prompt layer less about exhaustively listing behaviors and more about setting the right frame for the task.
The broader implication is that prompt engineering may become less about piling on examples and more about deciding what context matters. For Anthropic's Fable 5 models, the company says the smaller system prompt is not a reduction in ambition. It is an attempt to match the steering method to the model's actual strengths.