Decision Lattice: Building an AI-Assisted Decision-Making Framework with 38 Thinking Models

Decision Lattice is an open-source project that aims to help AI agents select appropriate cognitive frameworks and output interpretable recommendations in complex decision-making scenarios through the structured combination of 38 thinking models. Addressing the cognitive biases and information overload issues faced by human decision-makers in high-stakes, high-uncertainty scenarios, this project introduces the thinking model approach to AI agents, providing knowledge base and routing workflow support.

When facing complex strategic choices, human decision-makers often fall into the dilemmas of cognitive bias and information overload. Psychological research shows that when decisions are high-stakes, highly uncertain, and involve multiple stakeholders, the reliability of intuitive judgments decreases significantly. Structured thinking models can force a review of problems from different angles, reducing the risk of blind spots. The Decision Lattice project introduces this concept to AI agents, providing Codex agents with a knowledge base of 38 thinking models and routing workflows to help determine which models are applicable to a situation, why, and how to translate them into specific recommendations.

The core innovation of the project lies in its decision classification system and model routing mechanism. The system classifies decisions through five dimensions: high/low stakes, reversibility, degree of uncertainty, number of actors, and system complexity, forming a multi-dimensional coordinate for decision scenarios. Based on the classification results, a streamlined combination of models is selected from the 38 models (instead of stacking all models) to avoid cognitive overload and focus on relevant frameworks. The selected combination covers four perspectives: cognitive, system, social, and time, forming a multi-dimensional analytical lens.

The output of Decision Lattice includes final recommendations and complete decision support information: evidence gaps (areas where current information is insufficient), guardrails (decision boundaries and risk warnings), and re-review trigger conditions (when to re-evaluate). This structure focuses on decision quality, helping to understand the basis of decisions, identify risk boundaries, and establish dynamic adjustment mechanisms. It is applicable to scenarios such as strategic selection, architecture/product bets, roadmap prioritization, recruitment and organizational design, market competition analysis, uncertainty planning, and risk reviews.

The project contains 38 thinking models, divided into four categories: cognitive (first principles, reverse thinking, probabilistic thinking, etc.), system (feedback loops, bottleneck theory, emergent properties, etc.), social (game theory, network effects, consensus mechanisms, etc.), and time (compound interest effect, path dependence, option value, etc.). The documents in the references directory detail model definitions, applicable scenarios, and usage methods: model-catalog.md provides an overview, models-*.md elaborate by category, and routing-matrix.md defines the mapping between decision dimensions and model selection.

Usage Guide: Clone the repository to the Codex skills directory and call it via the $decision-lattice command. Example uses include: analyzing build-internal vs. buy-from-vendor decisions, stress-testing product strategies (uncertainty/tail risks), and selecting models for recruitment and organizational design decisions. Core principles: choose the minimal useful set of models, introduce challenger models when the framework may be wrong, prioritize survival when tail risks are high, and separate reasoning quality from evidence quality.

Limitations: Suitable for scenarios with high structure and clearly definable decision dimensions; over-reliance on models for highly creative decisions may limit thinking; model selection still requires judgment, and the routing mechanism cannot completely eliminate subjectivity. The project uses the MIT license, encouraging free use, modification, and distribution. Insights: It represents a path for AI-assisted decision-making—enhancing human decisions rather than replacing them, providing a reference for building complex decision support systems; the modular design allows the model library to be tailored, the routing mechanism can be extended into a complex engine, and the output structure can interface with downstream execution systems.