{"candidate_reason": "prompt scope discovery", "chunk_end": 175, "chunk_start": 1, "chunk_summary": "The prompt defines a semantic classifier for shot relationships, establishes a brittle token-based blacklist for field validation, and enforces a contract for blind string mutation in downstream T2I prompts.", "duration_ms": 25169, "findings": [{"category": "llm_closed_list_instruction", "evidence": "ref_usage: zoom_in_detail, exact_background, atmosphere_reference", "line_end": 59, "line_start": 20, "recommended_fix": "Define these relationships in a structured schema with clear visual parameters, or move the classification logic to a dedicated vision-language model step if high precision is required.", "severity": "P1", "why_problematic": "The LLM is instructed to classify open-world visual and temporal relationships (camera movement, timing, subject identity) into a closed set of three semantic categories. This classification directly routes how reference images are used in the generation pipeline."}, {"category": "blind_string_mutation", "evidence": "반드시 확정적 지시로 작성... 조건부 표현 금지 (Must write as definitive instructions... conditional expressions forbidden)", "line_end": 92, "line_start": 87, "recommended_fix": "Instead of raw string concatenation, use a structured list of entities and actions that the prompt builder can resolve based on the final scene state.", "severity": "P1", "why_problematic": "This establishes a contract for blind string mutation. The LLM is forced to produce unconditional natural language strings because the downstream T2I prompt builder blindly consumes the text without the ability to parse or resolve conditional logic."}, {"category": "semantic_string_judgment", "evidence": "human / person / character / body / figure / man / woman / detective / prisoner / child / person silhouette", "line_end": 157, "line_start": 146, "recommended_fix": "Enforce this constraint via entity-type validation in the schema rather than substring matching on natural language labels.", "severity": "P1", "why_problematic": "The prompt defines a blacklist of natural-language tokens and states that their presence in the 'label' field will trigger a 'fail-fast' rejection. This is a brittle string-pattern approach to enforcing a semantic boundary (excluding human entities from environment/prop lists)."}, {"category": "schema_or_enum_drift", "evidence": "immobilized_character / character / pose 등 enum 외 값 절대 금지", "line_end": 161, "line_start": 161, "recommended_fix": "Centralize the enum definition and use a shared schema validator to ensure all steps use the same vocabulary without manual prompt-side warnings.", "severity": "P2", "why_problematic": "The prompt explicitly lists forbidden enum values that likely exist in other parts of the system or previous versions. This indicates that the schema enforcement is being manually supplemented by prompt-side warnings to prevent drift between different pipeline steps."}], "path": "prompts/_base/shot_dependency_t2i/7.202605151200/system.md", "scan_kind": "prompt", "sha256": "4bb9564d103d4c10a4183bbc190f58568e1564cf199cf2e4defb057a3542b378"}