Back to ICM
About · Perspective · 4 min

About this series

Credit, sources, and what's mine vs. what's Jake Van Clief's.

Where ICM comes from

Interpretable Context Methodology (ICM) was created by Jake Van Clief (RinDig on GitHub), who developed and documented the approach across his GitHub repository, a 2026 academic paper (co-authored with David McDermott), a 19-video YouTube series, and an ongoing community on Skool called Clief Notes. Jake's central argument is that AI agent orchestration is a file-organization problem, not a code problem. That's the foundation this entire series rests on.

The framework's core constructs are Jake's:

  • The five-layer context model (Layer 0 root CLAUDE.md → Layer 4 working artifacts)
  • The map / rooms / workspace analogy used to teach it
  • The 60/30/10 rule for production AI systems
  • The Book / Movie / Video Game framework for deciding what to invest in
  • The 15 conventions in _core/CONVENTIONS.md
  • The stage-contract template (Inputs / Process / Checkpoints / Audit / Outputs)
  • The reference workspaces (script-to-animation, course-deck-production, workspace-builder, newsletter-writer)

If you read this series and want to go deeper on any of those, Jake's original work is the right destination:

What this series adds

Two things, both clearly delineated from Jake's original work.

A non-technical synthesis. Jake's audience is mostly developers and AI consultants: people who can read a folder tree and infer the methodology. Most of the people I work with at Time To Value can't, and shouldn't have to. The bulk of this series (Parts 1 through 4) is a translation of Jake's methodology into language for tech-adjacent professionals: consultants, PMs, founders, agency owners, ops leads. The structure of the argument is mine; the methodology being argued for is Jake's. Where specific Jake-isms appear in the prose (the 60/30/10 rule, the map/rooms/workspace analogy), they're attributed inline. The Book/Movie/Video Game framework is also Jake's, but it's a whether to build question rather than a how to build one, so it lives in a separate post outside this series.

A forward-looking durable-execution chapter. Part 5 (Beyond ICM) is where I diverge into my own work. It names a gap I felt when I started building ICM workspaces myself: the methodology is a clean workflow specification with no formal runtime semantics. The proposal in Part 5, adding workflow.yaml and YAML frontmatter on stage CONTEXT.md so any durable runtime (DBOS, LangGraph, Temporal) can compile from the same spec, is mine, drawn from research notes I wrote while exploring the question. Part 5 is honest about being aspirational; nothing in it is in production yet.

Source material referenced throughout

The series draws from my private working notes (the ICM and Durable Execution clusters in my Obsidian vault) plus two real ICM workspaces I run today: meeting-summarizer and content-to-guide. Both are in ~/source/scripts-and-agents/icm-workspaces/. The starter skills shipped on the landing page (icm-workspace-builder and skill-to-icm-converter) are sanitized versions of skills I wrote for a client engagement.

The starter skills are free to download. The series is free to read. If you find errors, contradictions, or places where the synthesis loses Jake's intent, tell me. I'd rather fix them than have them sit.

ICM doesn't sit alone. The broader pattern (file-based, markdown-driven agent architecture) is converging from several directions:

  • AGENTS.md: the open standard adopted by 60,000+ repositories, backed by the Linux Foundation Agentic AI Foundation. The convention layer that makes ICM workspaces portable across Cursor, Codex CLI, GitHub Copilot, Windsurf, and Devin.
  • Anthropic's official guidance: Effective Context Engineering for AI Agents and the CLAUDE.md hierarchy docs essentially endorse ICM's layering model without naming it.
  • Skills vs. MCP: recent writing on running AI agents on markdown files instead of MCP servers (a 100x token reduction in some benchmarks). Same instinct as ICM, lower in the stack.
  • The simpler-is-better movement: PydanticAI, smolagents, and GitHub's agentic primitives writing, all converging on the same point from the framework side.

ICM is the most structured and complete version of this pattern I've found. That's why it's worth a series of its own.

Contact

Errors, questions, or suggestions: contact form. For Jake's work specifically, his Skool community is the right place.

All ICM parts