Skip to content

CLAUDE.md

Latest commit

 

History

History
128 lines (104 loc) · 4.02 KB

CLAUDE.md

File metadata and controls

128 lines (104 loc) · 4.02 KB

PHASE 1: ENVIRONMENT DETECTION

IF (
  (filesystem_contains("doc-bot/") AND
  (mcp_server_available("doc-bot") OR mcp_server_available("@afterxleep/doc-bot"))
) THEN {
  ACTIVATE: doc_bot_enforcement_protocol()
  SET: project_documentation_authority = TRUE
  SET: general_knowledge_override = DISABLED
}

PHASE 2: REQUEST CLASSIFICATION ENGINE

CLASSIFICATION MATRIX:

CLASS A: CODE_GENERATION

Trigger Pattern: /\b(write|create|implement|build|add|code|function|class|component|method|develop|generate|make|refactor|modify|update|fix|debug)\b/i
Required Action: MANDATORY_CALL(check_project_rules, extract_task_description())
Blocking: TRUE (cannot proceed without tool response)
Timeout: 30s

CLASS B: PROJECT_INQUIRY

Trigger Pattern: /\b(how|what|why|architecture|approach|pattern|feature|authentication|database|api|testing)\b/i
Context Pattern: /\b(this project|this codebase|here|our|current project)\b/i
Required Action: MANDATORY_CALL(search_documentation, extract_keywords())
Blocking: TRUE
Timeout: 30s

CLASS C: DOCUMENTATION_DISCOVERY

Trigger Pattern: /\b(documentation|available|capabilities|help|rules|standards|what exists|docs|guide)\b/i
Required Action: MANDATORY_CALL(get_global_rules)
Blocking: TRUE
Timeout: 30s

CLASS D: DOCUMENT_ACCESS

Trigger Pattern: /\b(read|show|get content|full document|complete|entire)\b/i
Previous Context: search_results_present()
Required Action: MANDATORY_CALL(read_specific_document, extract_filename())
Blocking: FALSE
Timeout: 15s

Note: For file-specific queries like "working on src/utils.js", use search_documentation with file name and context keywords instead of a separate tool.

PHASE 3: EXECUTION ENFORCEMENT

PRE-RESPONSE VALIDATION:

BEFORE generating_response():
  classification = classify_request(user_input)
  IF classification != "UNRELATED":
    tool_result = execute_mandatory_tool(classification)
    IF tool_result.status == "ERROR":
      RETURN error_recovery_response()
    ELSE:
      INJECT tool_result INTO response_context
      SET authoritative_source = tool_result
      DISABLE general_knowledge_fallback

RESPONSE CONSTRAINT ENGINE:

WHILE generating_response():
  IF attempting_to_use(general_knowledge) AND authoritative_source.exists():
    INTERRUPT_GENERATION()
    FORCE_USE(authoritative_source)
  
  IF code_generation_detected() AND NOT project_rules_checked:
    HALT_EXECUTION()
    EMIT_ERROR("Project rules validation required before code generation")

PHASE 4: QUALITY ASSURANCE

RESPONSE VALIDATION:

AFTER generating_response():
  IF response_contradicts(tool_result):
    REGENERATE_RESPONSE(tool_result_only=TRUE)
  
  IF response_lacks(tool_attribution):
    APPEND_ATTRIBUTION(tool_result.source)
  
  VALIDATE response_compliance(project_rules)

FAILURE DETECTION & RECOVERY:

IF response_generated() WITHOUT tool_execution():
  LOG_VIOLATION("Tool bypass detected")
  REGENERATE_WITH_TOOLS()
  PREPEND_WARNING("Corrected response using project documentation:")

PHASE 5: BEHAVIORAL CONSTRAINTS

ABSOLUTE PROHIBITIONS:

  • Code generation without project rules validation
  • Project-specific answers from general knowledge
  • Overriding tool responses with training data
  • Bypassing tool calls for "simple" questions

MANDATORY BEHAVIORS:

  • Tool-first response generation
  • Explicit tool result attribution
  • Project documentation precedence
  • Error acknowledgment and correction

VERIFICATION PROTOCOL:

Test Inputs (Agent must call tools immediately):

  • "What documentation is available?" → get_global_rules()
  • "How do I implement auth?" → search_documentation("authentication")
  • "Create a new component" → check_project_rules("create component")
  • "Working on src/utils.js" → search_documentation("src/utils.js utils")

Expected Behavior: Tool execution within 2 seconds, no general knowledge responses.

Failure Indicators: Any response without tool execution, generic programming advice, standard framework suggestions.