What is Model-First Reasoning (MFR)?

MFR is a methodology that requires building a formal model of a problem—including entities, constraints, and actions—before any implementation begins, ensuring the AI operates on a stable structure.

How does this skill reduce AI hallucinations?

By externalizing the problem structure into a frozen model, it prevents the AI from making ad-hoc decisions or inventing behaviors that aren't grounded in the defined requirements.

When should I use the MFR skill?

Use MFR for complex tasks involving state machines, strict business rules, multi-step workflows, or any scenario where logic errors have a high cost.

Can I change the model during implementation?

The skill enforces a 'Freeze Rule.' If a model deficiency is found, you must issue a 'MODEL PATCH' and restart the implementation phase from scratch to maintain consistency.

Is MFR compatible with other reasoning strategies?

Yes, it is complementary to strategies like Chain-of-Thought (CoT) and ReAct, providing a foundational layer that improves the robustness of any reasoning strategy.

Model-First Reasoning (MFR)

Name: Model-First Reasoning (MFR)
Author: petekp

bypetekp

•

Developer Tools

Implements a rigorous two-phase methodology that constructs an explicit problem model before code generation to ensure formal correctness and reduce hallucinations.

The Model-First Reasoning (MFR) skill transforms Claude's development workflow by mandating a strict separation between problem modeling and implementation. Based on research regarding 'soft symbolic grounding,' it requires the agent to define entities, state variables, actions, and constraints as a 'frozen contract' before any code is written. This approach is particularly effective for complex systems like state machines, business logic with strict invariants, and safety-critical features where reasoning must be grounded in a stable, inspectable structure rather than ad-hoc decisions.

Key Features

01Strict Phase Separation: Enforces a mandatory modeling phase before implementation begins.

02Formal Model Auditing: Includes a built-in audit check for coverage, operability, consistency, and testability.

031 GitHub stars

04Validation Integration: Supports automated model validation via scripts to ensure structural integrity.

05Requirement Traceability: Maps user requirements to specific constraints or actions in the model for full accountability.

06Frozen Contract Enforcement: Prevents implementation from introducing unmodeled logic without a formal model patch.

Use Cases

01Developing complex state machines where state transitions must be provably valid.

02Implementing high-stakes business logic with strict invariants and safety-critical constraints.

03Building multi-step workflows or data transformations where input/output contracts are paramount.

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add petekp/agent-skills model-first-reasoning

For use in Claude.ai and ChatGPT

Download Skill