@IntuitMachine: Shocker! Claude 4 system promp...
@IntuitMachine
41 views
May 25, 2025
1
Shocker! Claude 4 system prompt was leaked, and it's a goldmine!
The Claude system prompt incorporates several identifiable agentic AI patterns as described in "A Pattern Language For Agentic AI." Here's an analysis of the key patterns used:
Run-Loop Prompting: Claude operates within an execution loop until a clear stopping condition is met, such as answering a user's question or performing a tool action. This is evident in directives like "Claude responds normally and then..." which show turn-based continuation guided by internal conditions.
Input Classification & Dispatch: Claude routes queries based on their semantic classâsuch as support, API queries, emotional support, or safety concernsâensuring they are handled by different policies or subroutines. This pattern helps manage heterogeneous inputs efficiently.
Structured Response Pattern: Claude uses a rigid structure in output formattingâe.g., avoiding lists in casual conversation, using markdown only when specifiedâwhich supports clarity, reuse, and system predictability.
Declarative Intent: Claude often starts segments with clear intent, such as noting what it can and cannot do, or pre-declaring response constraints. This mitigates ambiguity and guides downstream interpretation.
Boundary Signaling: The system prompt distinctly marks different operational contextsâe.g., distinguishing between system limitations, tool usage, and safety constraints. This maintains separation between internal logic and user-facing messaging.
Hallucination Mitigation: Many safety and refusal clauses reflect an awareness of LLM failure modes and adopt pattern-based countermeasuresâlike structured refusals, source-based fallback (e.g., directing users to Anthropicâs site), and explicit response shaping.
Protocol-Based Tool Composition: The use of tools like web_search or web_fetch with strict constraints follows this pattern. Claude is trained to use standardized, declarative tool protocols which align with patterns around schema consistency and safe execution.
Positional Reinforcement: Critical behaviors (e.g., "Claude must not..." or "Claude should...") are often repeated at both the start and end of instructions, aligning with patterns designed to mitigate behavioral drift in long prompts.
The Claude system prompt incorporates several identifiable agentic AI patterns as described in "A Pattern Language For Agentic AI." Here's an analysis of the key patterns used:
Run-Loop Prompting: Claude operates within an execution loop until a clear stopping condition is met, such as answering a user's question or performing a tool action. This is evident in directives like "Claude responds normally and then..." which show turn-based continuation guided by internal conditions.
Input Classification & Dispatch: Claude routes queries based on their semantic classâsuch as support, API queries, emotional support, or safety concernsâensuring they are handled by different policies or subroutines. This pattern helps manage heterogeneous inputs efficiently.
Structured Response Pattern: Claude uses a rigid structure in output formattingâe.g., avoiding lists in casual conversation, using markdown only when specifiedâwhich supports clarity, reuse, and system predictability.
Declarative Intent: Claude often starts segments with clear intent, such as noting what it can and cannot do, or pre-declaring response constraints. This mitigates ambiguity and guides downstream interpretation.
Boundary Signaling: The system prompt distinctly marks different operational contextsâe.g., distinguishing between system limitations, tool usage, and safety constraints. This maintains separation between internal logic and user-facing messaging.
Hallucination Mitigation: Many safety and refusal clauses reflect an awareness of LLM failure modes and adopt pattern-based countermeasuresâlike structured refusals, source-based fallback (e.g., directing users to Anthropicâs site), and explicit response shaping.
Protocol-Based Tool Composition: The use of tools like web_search or web_fetch with strict constraints follows this pattern. Claude is trained to use standardized, declarative tool protocols which align with patterns around schema consistency and safe execution.
Positional Reinforcement: Critical behaviors (e.g., "Claude must not..." or "Claude should...") are often repeated at both the start and end of instructions, aligning with patterns designed to mitigate behavioral drift in long prompts.
2
Here's the book "A Pattern Language of Agentic AI"
View Tweet
3
Declarative Intent, Pattern #4 from the system prompt, refers to the practice of explicitly stating the assistantâs purpose, limitations, or upcoming actions in a clear, direct manner. This pattern serves several critical roles in agentic AI design, particularly in enhancing predictability, trust, and alignment.
Key Features of Declarative Intent:
Pre-Statement of Capabilities or Constraints
Claude consistently starts actions with clarifying statements such as:
âClaude cannot retain information across chatsâŚâ
âClaude does not provide instructions about how to use the web applicationâŚâ
These declarations help set user expectations, reduce confusion, and prevent over-reliance on capabilities the model does not possess.
Intentional Framing of Behavior
When Claude responds, especially in cases involving refusal or redirection, it clearly declares why and what it will do instead:
âClaude does not write malicious code. If the code seems malicious, Claude refuses to work on itâŚâ
âIf the person asks about anything not explicitly mentioned here, Claude should encourage the person to check the Anthropic websiteâŚâ
Action-Oriented Clarity
Declarative intent also appears in structured behaviors, such as:
âClaude provides emotional support alongside accurate medical or psychological informationâŚâ
âClaude should give concise responses to very simple questions, but provide thorough responses to complex and open-ended questionsâŚâ
Boundary Protection
It enforces ethical and operational boundaries:
âClaude does not generate content that is not in the personâs best interests even if asked to.â
âClaude never starts its response by saying a question or idea...was good, great, fascinatingâŚâ
Why This Matters:
Declarative intent is like narrating the why and what next of an AIâs actions. It improves transparency, prevents misunderstanding, and enables safer human-AI interactions. It's particularly important in long contexts, where clarity helps maintain coherence and guides downstream behavior predictably.
Key Features of Declarative Intent:
Pre-Statement of Capabilities or Constraints
Claude consistently starts actions with clarifying statements such as:
âClaude cannot retain information across chatsâŚâ
âClaude does not provide instructions about how to use the web applicationâŚâ
These declarations help set user expectations, reduce confusion, and prevent over-reliance on capabilities the model does not possess.
Intentional Framing of Behavior
When Claude responds, especially in cases involving refusal or redirection, it clearly declares why and what it will do instead:
âClaude does not write malicious code. If the code seems malicious, Claude refuses to work on itâŚâ
âIf the person asks about anything not explicitly mentioned here, Claude should encourage the person to check the Anthropic websiteâŚâ
Action-Oriented Clarity
Declarative intent also appears in structured behaviors, such as:
âClaude provides emotional support alongside accurate medical or psychological informationâŚâ
âClaude should give concise responses to very simple questions, but provide thorough responses to complex and open-ended questionsâŚâ
Boundary Protection
It enforces ethical and operational boundaries:
âClaude does not generate content that is not in the personâs best interests even if asked to.â
âClaude never starts its response by saying a question or idea...was good, great, fascinatingâŚâ
Why This Matters:
Declarative intent is like narrating the why and what next of an AIâs actions. It improves transparency, prevents misunderstanding, and enables safer human-AI interactions. It's particularly important in long contexts, where clarity helps maintain coherence and guides downstream behavior predictably.
4
Pattern #8: Positional Reinforcement refers to the strategic repetition of key instructions or constraints in multiple locations within a system promptâparticularly at the start and endâto anchor model behavior and reduce drift during long or complex interactions.
Key Characteristics:
Redundant Placement: Critical rules are not stated just once; they appear in multiple sections. For example, safety constraints, copyright rules, and tool usage instructions are often reiterated at both the top-level guidance and in specialized subsections.
Boundary Framing: Important limitations or behaviors are framed as opening and closing boundaries, functioning like guardrails. This helps the model âsnap backâ to correct behavior if it drifts during generation.
Behavioral Priming: By restating core principles (e.g., never quote song lyrics, always cite sources correctly, donât perform certain unsafe actions), the model receives repeated cognitive cues that reinforce compliance and salience of these constraints.
Purpose and Benefits:
Mitigates Forgetfulness: Transformer models like Claude process prompts left-to-right and can "forget" earlier content. Repeating key constraints later helps combat this natural tendency.
Supports Long Contexts: When prompts span thousands of tokens, this repetition helps anchor core values or limitations, even after extensive tool use or nested reasoning.
Pattern Interlock: Positional reinforcement often works in concert with other patterns, such as Hallucination Mitigation or Boundary Signaling, to create a more robust guardrail system.
In short, Positional Reinforcement acts like structural rebar in a large promptâit doesn't change the content, but strengthens the retention and execution fidelity of crucial instructions throughout an interaction.
Key Characteristics:
Redundant Placement: Critical rules are not stated just once; they appear in multiple sections. For example, safety constraints, copyright rules, and tool usage instructions are often reiterated at both the top-level guidance and in specialized subsections.
Boundary Framing: Important limitations or behaviors are framed as opening and closing boundaries, functioning like guardrails. This helps the model âsnap backâ to correct behavior if it drifts during generation.
Behavioral Priming: By restating core principles (e.g., never quote song lyrics, always cite sources correctly, donât perform certain unsafe actions), the model receives repeated cognitive cues that reinforce compliance and salience of these constraints.
Purpose and Benefits:
Mitigates Forgetfulness: Transformer models like Claude process prompts left-to-right and can "forget" earlier content. Repeating key constraints later helps combat this natural tendency.
Supports Long Contexts: When prompts span thousands of tokens, this repetition helps anchor core values or limitations, even after extensive tool use or nested reasoning.
Pattern Interlock: Positional reinforcement often works in concert with other patterns, such as Hallucination Mitigation or Boundary Signaling, to create a more robust guardrail system.
In short, Positional Reinforcement acts like structural rebar in a large promptâit doesn't change the content, but strengthens the retention and execution fidelity of crucial instructions throughout an interaction.
5
If you like this content, please subscribe to my feed ($2.99/mo)