solo-sgr

/sgr

Design and implement Schema-Guided Reasoning (SGR) pipelines. Translate domain expert mental checklists into structured reasoning schemas for LLMs.

Source: Rinat Abdullin — Schema-Guided Reasoning

Core Principle

SGR = guide LLM reasoning through predefined steps via constrained decoding. Instead of free-form text → enforce a schema that defines what steps, in which order, where to focus attention.

Domain expert mental checklist → Pydantic/Zod schema → Constrained decoding → Deterministic dispatch

When to Use

• Designing agent tool dispatch (NextStep pattern)
• Building structured analysis pipelines (compliance, code review, evaluation)
• Replacing prompt chains with single structured call
• Any place where LLM output must be parseable and actionable

Steps

1. Parse task from $ARGUMENTS:

   • If "audit": scan project for existing Pydantic/Zod schemas, evaluate against SGR patterns
   • If task description: design SGR pipeline from scratch
   • If empty: ask "What domain/task should the SGR pipeline handle?"

2. Identify the reasoning cascade — interview the domain:

   • What decisions does a human expert make? In what order?
   • What information does each step need from previous steps?
   • Where does the expert need to "look before deciding"?
   • What are the possible actions at the end?

   This is the critical step. SGR quality = how well you translate the expert's mental checklist.

3. Design the schema following SGR patterns:

   The NextStep Pattern (agent loop)

   class NextStep(BaseModel):
       current_state: str                    # thinking space
       plan_remaining_steps: list[str]       # 1-5 steps, only first used
       task_completed: bool                  # routing gate
       function: Union[Tool1, Tool2, ..., ReportCompletion] = Field(
           ..., description="execute first remaining step"
       )
   

   The Analysis Cascade Pattern (single-shot)

   class Analysis(BaseModel):
       preliminary: str                      # initial assessment
       classification: Literal["a", "b", "c"]  # force categorization
       evidence: list[str]                   # cite sources
       gaps: list[GapItem]                   # structured findings
       verdict: Literal["pass", "partial", "fail"]  # final decision
       reasoning_for_verdict: str            # explain after deciding