Interpretable Context Methodology (ICM) is a filesystem-first approach to AI workflow orchestration using folders, markdown contracts, and staged human review.
This repository captures a practical and conceptual exploration of ICM, centered on the idea that:
For many sequential, human-reviewed AI workflows, the filesystem can act as the orchestration layer.
Instead of relying on complex multi-agent framework code, ICM organizes workflow stages with folders and markdown files so context, process, and handoffs are explicit and easy to inspect.
icm-core/— working ICM implementation with three production-ready workspaces:workspaces/script-to-animation/— content idea through script, animation spec, and Remotion codeworkspaces/course-deck-production/— unstructured material (PDFs, notes) into polished slide decksworkspaces/workspace-builder/— build a new ICM workspace for any domain_core/— shared conventions, templates, and placeholder syntax reference
ICM_conversation.md— detailed long-form breakdown of ICM concepts, architecture, and implications.docs/File Tree as Agent.md— exploration of file-tree-first agent design and the context wall problem.src/filetree-agent.html— interactive visual presentation of the core thesis and ICM paper walkthrough.ICM Paper.pdf— local copy of the ICM paper referenced throughout the project.
- Demonstrates architecture thinking for AI systems, not just prompt iteration.
- Focuses on interpretability, portability, and human-in-the-loop workflow design.
- Shows how simple primitives (files/folders) can replace unnecessary orchestration complexity for the right class of tasks.
git clone <this-repo>
cd icm-core
# Open in Claude Code, then navigate to a workspace:
cd workspaces/script-to-animation
# Type "setup" to run onboardingSee icm-core/README.md for full documentation on workspaces, conventions, and how to build your own.
- arXiv abstract: Interpretable Context Methodology
- arXiv PDF: https://arxiv.org/pdf/2603.16021