ite-workflows/AGENT.md

ICM (Interpretable Context Methodology) – Process Overview

ICM is a workflow for orchestrating AI agent tasks using the filesystem as the coordination mechanism, eliminating the need for complex agent frameworks.

Core Idea

• Numbered folders represent sequential stages of a workflow.
• Each stage contains a CONTEXT.md file that defines:
  • Inputs (what files to read from reference and working layers)
  • Process (what the agent should do)
  • Outputs (what files to write)
• Plain markdown and JSON files carry prompts and context; local scripts handle non‑AI tasks.
• The agent reads the appropriate files at each stage, producing intermediate outputs that humans can inspect, edit, and approve before proceeding.

Five‑Layer Context Hierarchy

1. Layer 0 – CLAUDE.md: Workspace identity and routing.
2. Layer 1 – CONTEXT.md: Workspace‑level task routing (which stage to run).
3. Layer 2 – Stage CONTEXT.md: Stage‑specific contract (inputs, process, outputs).
4. Layer 3 – Reference material (references/, _config/): Stable rules, voice guides, design systems (the “factory”).
5. Layer 4 – Working artifacts (output/): Per‑run intermediate results (the “product”).

Workflow Characteristics

• Sequential: Stage n+1 reads the output of Stage n.
• Human‑in‑the‑loop: After each stage, a human can review and edit the output file before the next stage runs.
• Editable & observable: All prompts, context, and intermediate results are plain text files in folders—easy to inspect, version, and modify.
• Portable: A workspace is just a folder; it can be copied, versioned with Git, or shared without extra configuration.
• Focused context loading: Each stage loads only the context it needs (Layers 0‑4 relevant to that stage), keeping the model’s context window small and relevant.

Benefits

• Replaces multi‑agent orchestration frameworks with simple folder conventions.
• Enables non‑technical users to modify prompts and workflows by editing markdown files.
• Provides inherent audit trails and review gates, supporting human oversight and debugging.
• Scales token usage efficiently by avoiding irrelevant context.

Typical Use Cases

Content production pipelines (research → script → animation), slide‑deck generation, research analysis, policy workflows—any repeatable, sequential process where human review at each step adds value.

License

ICM is open source under the MIT license; a workspace‑builder tool is included to scaffold new workspaces.