Interpretable Context Methodology (ICM) is a lightweight, filesystem-based approach to orchestrating AI agent workflows. Instead of using complex multi-agent frameworks, ICM structures workflows using numbered folders, markdown files, and local scripts. Each folder represents a stage in the workflow, with markdown files defining prompts, inputs, and outputs. The system relies on a single AI agent that reads context from the appropriate files at each step, enabling full control through plain text.

Key principles include:

• One stage, one job: Each folder handles a single, well-defined step.
• Plain text as interface: All communication occurs via markdown and JSON files.
• Layered context loading: Agents load only relevant context per stage, improving performance and reducing token usage.
• Every output is an edit surface: Humans can inspect and modify intermediate outputs before the next stage runs.
• Configure the factory, not the product: Workspace setup is done once; each run uses the same configuration.

The architecture uses a five-layer context hierarchy:

1. Layer 0: Global identity (CLAUDE.md)
2. Layer 1: Task routing (CONTEXT.md)
3. Layer 2: Stage-specific instructions
4. Layer 3: Reference material (stable across runs)
5. Layer 4: Working artifacts (per-run content)

ICM is ideal for sequential, human-reviewed workflows like content creation, research, and training material development. It enables full observability, version control via Git, and easy sharing. However, it is not suited for real-time agent collaboration, high-concurrency systems, or complex branching logic.

The method draws from Unix pipelines, Make, literate programming, and multi-pass compilation. Early practitioner use shows a U-shaped intervention pattern—frequent edits at early and final stages, less in the middle—reflecting creative direction-setting and final alignment checks. Future directions include semantic debugging, traceability tools, and source integrity improvements that learn from repeated human edits.

ICM is open source under the MIT license and includes a workspace-builder for creating new workflows across domains.