RinDig/Interpretable-Context-Methodology

What it solves

Interpretable Context Methodology (ICM) addresses the complexity and opacity of multi-agent frameworks. Instead of using code-heavy orchestration (like CrewAI or LangChain) to manage multi-step AI workflows, ICM uses a structured filesystem. This eliminates the need for complex development work to change a workflow's order, swap prompts, or add stages, making the system observable and portable by default.

How it works

ICM replaces framework-level orchestration with a folder hierarchy. A single AI agent reads specific files at specific times based on a five-layer context loading system:

• Layer 0 (CLAUDE.md): Global identity and location.
• Layer 1 (CONTEXT.md): Task routing.
• Layer 2 (Stage CONTEXT.md): Specific task instructions (the "contract").
• Layer 3 (Reference material): Stable, cross-run knowledge like brand guidelines or design systems.
• Layer 4 (Working artifacts): Dynamic inputs and outputs from previous stages.

Stages are numbered folders (e.g., 01-research, 02-script) that enforce a sequential execution order. Each stage's output is a plain-text file that a human can review and edit before the next stage begins, ensuring a "glass-box" workflow.

Who it’s for

It is designed for practitioners who have sequential, repeatable multi-step workflows that require human-in-the-loop review at each step, such as content production, research analysis, and reporting processes.

Highlights

• Filesystem as Architecture: Uses folder numbering and hierarchy to manage state and sequencing instead of code.
• Human-in-the-Loop: Every intermediate output is an editable text file, allowing for easy human steering.
• Layered Context: Prevents model performance degradation by loading only the necessary context for the current stage.
• Observability: No special logging tools are needed; the system state is simply the files on disk.
• Workspace Builder: Includes a dedicated workspace to help users scaffold new ICM-compliant workflows for any domain.