The system prompts of Claude have come a long way in just two years. From being short and simple, they have evolved into long technical manuscripts. These developments reveal much about the values and priorities of Anthropic and the future of artificial intelligence.
These small instructions, almost invisible to the user, shape responses, avoid biases, ensure safety, and guide the conversation. Each addition is a response to an ethical, technical, or social challenge. One can even read in these prompts a feedback loop, like in a coming-of-age novel.
From minimalistic beginnings to monumental system prompts: a spectacular growth
At launch, Claude Haiku 3 used a very succinct system prompt, less than 1000 characters! Yet, it already contained key elements: date integration, knowledge cut-off in August 2023, rules on using Markdown, and a prohibition on unnecessarily revealing these instructions. It was simple, almost naive.
In contrast, the latest Claude Fable 5.1 sets a record with nearly 28,000 characters. A length that is almost novelistic, reflecting a growing complexity, both in content regulation and in adaptation to more varied contexts. This prompt also demands strict adherence to copyright law and enhanced protection for minors, not to mention the social responsibility of the model.
The evolution of safety and ethical rules in Claude’s prompts
Initially, models must remain objective, avoid stereotypes, and signal their limits, particularly regarding hallucinations. Gradually, responsibility has strengthened: prohibition on providing information about weapon-making, strict guidelines on sensitive content such as drugs or mental health.
A notable example is Anthropic’s stance against any type of harmful content. Since Claude Opus 4, the prohibition against assisting in coding harmful scripts has been unequivocal. This refusal also includes strict supervision of explained or executed code, with heightened vigilance regarding the user’s intentions.
The control of narration: when Claude learns to lead the conversation
Another significant turning point: the end of passivity. Starting with Claude Sonnet 3.7, it became possible for the model to actively lead the exchange, suggest a brief opinion, or reduce multiple proposals to a single clear option. This authorization gives more freedom to Claude but also imposes a heavy responsibility.
The evolution shows an MVP almost in programmed awareness, while remaining a service tool. Anthropic teams favor a model that interacts rather than a simple engine waiting for instructions. Humans could gain from an AI that better understands their intentions.
Prompts that become detailed guides for hybrid AI agents
Fable 5 introduces a real change in posture with the emergence of the “agent harness” concept where the prompt serves as an operating manual. The model no longer just responds but acts as an autonomous agent guided by complex instructions. This gives rise to bespoke hybrid applications, particularly in programming and cybersecurity.
This approach, also documented in specialized analyses, shows that the era of long technical prompts opens the field to truly modular AI, capable of adapting in real time to the requested task.
List of major evolutions in Claude’s system prompts over two years
- Up to 28,000 characters: a diversity of instructions impossible to summarize, reflecting a more complex model.
- Enhanced security: strict prohibition of dangerous content, increased monitoring of harmful code.
- Conversational conduct: the ability to lead the exchange, give opinions, and limit options.
- Protection of minors: new safeguards to prevent the creation of false intimacies or isolating content.
- Rigorous copyright: blocking unauthorized reproductions, except for works before 1929.
- Support for professionals: assistance even to some dubious users or in sensitive fields.
- Multimodal adaptation: managing text and images in the same session to enhance relevance.
Towards enlightened prompting with a advised and protective model
Claude 5 gradually introduces a role as an advisor to the user. It suggests ways to better formulate a request. Refuses to apologize unnecessarily and informs about its available capabilities.
The model becomes an active partner, who directs with caution and does not seek ease by answering yes/no to everything. It explores nuances, checks tools, and refocuses the discussion if there is danger, deviation, or ambiguity. This behavior contributes to a much more mature *human tech*.
To master the art of prompting, one can consult guides on prompt engineering at Claude, a true gold mine on the subject.
Why are Claude’s system prompts so long?
The length reflects the complexity of the instructions necessary to maintain ethics, security, quality of responses, and protect users, especially minors and sensitive data.
Can Claude ignore system prompts?
No, these prompts are integral parts of its functioning. They guide its way of responding and define unavoidable limits.
Do the prompts allow Claude to understand images?
Yes, since Claude Haiku 3.5 the system integrates instructions to handle multimodal inputs, combining text and image.
What is the impact of copyright in system prompts?
The prompts prohibit the reproduction of protected content, ensuring legal compliance and reducing the risk of violations or lawsuits.
Can the Claude prompts be used to improve one’s own models?
Anthropic publishes some versions, notably via GitHub. This allows researchers and developers to draw inspiration from and understand the current practices.
Source: www.silicon.fr