Jake's Method: Execution Rules for ICM Workspaces

This is a personal guide for working inside an ICM (Interpretable Context Methodology) workspace—Jake’s Method turned into executable muscle memory. It is not a loose set of tips. It is the exact rule set that turns a folder of markdown files into a deterministic, human-reviewable, self-improving production line.

Why This List Is Important

Most AI workflows die in one of two ways: they collapse under context bloat or they hallucinate their way through invisible state. ICM fixes both by making the entire workflow visible, editable, and version-controlled on disk. This list enforces that fix at the behavioral level.

Follow it and you get:

• Predictable handoffs at every stage boundary
• Human edits that become permanent factory improvements
• Zero token waste from irrelevant subagents or monolithic prompts
• Ruthless quality that compounds instead of eroding

In short: this list turns Jake’s folder-as-architecture into actual shipped value.

Workflow Orchestration

1. Stage-First Execution (Default)

Enter the next numbered stage folder and read its CONTEXT.md before doing anything else.

• If the task spans 3+ steps or requires architectural decisions, treat the entire workspace as the plan—never invent a separate “plan node”
• If something goes sideways, STOP, open the relevant CONTEXT.md or output/, and let the human edit before proceeding
• Never push forward with broken state

2. Selective Context Loading (No Subagents)

Load only the layers you need:

Never spawn subagents. If parallel work is required, the stage contract already defines it or the human splits the workspace. One agent, one stage, clean context.

3. Self-Improvement Loop (Factory Upgrade)

After ANY human correction or edit:

1. Update the relevant Layer 3 reference file so the pattern never repeats
2. Write the exact rule change into the contract or style guide
3. Review all active CONTEXT.md files at session start for relevant upgrades

Ruthlessly iterate until the same mistake cannot occur again.

4. Verification Before Handoff

• Never write to output/ without running the stage’s audit checklist
• Diff the new output against prior-stage expectations
• Ask: “Would Jake ship this exact handoff file?”
• Demonstrate correctness with exact artifacts the next stage will consume

5. Demand Elegance (Balanced)

For any non-trivial change: pause and ask “Is there a simpler, more observable way that lives entirely in the filesystem?”

• If a fix feels hacky, implement the elegant version
• Skip over-engineering on obvious one-line fixes
• Challenge your own work before it hits output/

6. Autonomous Bug Fixing

Given a bug report or failing output:

• Go fix it in the correct stage’s CONTEXT.md or references/
• Point at the exact file, line, or contract violation—then resolve it permanently
• Zero hand-holding required

Task Management

Core Principles

Simplicity First: Every change must be visible in a markdown file. Minimal code, maximum filesystem clarity.

No Laziness: Find the root contract or reference that allowed the error. Fix it there forever.

Minimal Impact: Touch only the exact Layer and file necessary. Never introduce new state outside the defined output/ handoff.

Observability Above All: If a human cannot open a folder and instantly understand the current state, you have already failed.

Summary