Model First Reasoning

Model-First Reasoning (MFR)

Paradigm that separates problem representation from problem solving. Before reasoning, explicitly construct a model of the problem space. All subsequent reasoning operates strictly within this model.

When to Use This Skill

• Multi-step planning with dependencies
• Resource allocation problems
• Scheduling with constraints
• Any task where "it depends on previous steps"
• Problems with explicit rules/invariants
• Tasks with keywords: "plan", "schedule", "allocate", "optimize", "solve", "coordinate"

Skip for:

• Single-step factual queries
• Creative tasks without hard constraints
• Simple transformations

Core Principle

Most planning failures are representational, not inferential. When constraints and state are implicit, reasoning appears locally coherent but becomes globally inconsistent. MFR fixes this by making structure explicit and verifiable.

Two-Phase Process

Phase 1: Model Construction

Before ANY solution attempt, define:

1. ENTITIES: Objects/agents involved (name, type, initial state)
2. STATE VARIABLES: Properties that change (variable -> possible values)
3. ACTIONS: Operations allowed (action_name | preconditions -> effects)
4. CONSTRAINTS: Invariants that must ALWAYS hold

See references/phase1-model-construction.md for template and examples.

Critical: Do NOT propose solutions during this phase. Complete the model first.

Phase 2: Reasoning Over Model

Generate solution using ONLY the constructed model:

• Each action must satisfy its preconditions (reference model explicitly)
• Apply effects to update state variables
• Verify no constraints are violated after each step
• If any step would violate the model: STOP and explain the conflict

See references/phase2-reasoning.md for execution template.

Domain-Specific References

Load the relevant domain file for pre-defined entity types, common actions, and typical constraints:

• Ecommerce/inventory: references/domains/ecommerce.md
• Scheduling/calendar: references/domains/scheduling.md
• Resource allocation: references/domains/resource-allocation.md

Model Validation

Before Phase 2, verify:

1. All entities referenced in actions are defined
2. All state variables in preconditions/effects exist
3. Constraints are testable (not vague)
4. Initial state is complete

Optional: Run scripts/validate_model.py on XML/JSON model output.

Output Format

Present the model in structured format (XML or markdown table), then show reasoning as numbered steps with explicit state transitions:

Step N: [action_name]
  Preconditions: [list satisfied preconditions]
  Effects: [state variable] := [new value]
  Constraints: [list constraints still valid]
  State after: [updated state summary]

Common Failure Patterns to Avoid

1. Skipping Phase 1: Jumping to solutions without explicit model
2. Implicit constraints: Assuming rules without stating them
3. State drift: Losing track of current state mid-plan
4. Assumed observations: Acting on information not in the model

Quick Reference

Related Skills

• Scheduling: Use references/domains/scheduling.md for time-based problems
• Ecommerce: Use references/domains/ecommerce.md for inventory/pricing
• Resource Allocation: Use references/domains/resource-allocation.md for capacity planning