AI Business Maturity Model
Certifications
Find a CoachFind a SpeakerSign In

Planning System

A multi-agent planning framework that decomposes complex projects into self-contained tasks for AI execution. Uses three roles: Architect (plans), Build Agent (executes), and Test Agent (verifies).

How the System Works

This is a file-based coordination system for AI agents. The human user acts as a dispatcher — they describe what they want, and the system produces structured files that different AI agents can execute independently. No shared memory or API integration is needed; everything flows through markdown files on disk.

The Workflow
  1. User describes the goal in plain language (one sentence is enough).
  2. User pastes ARCHITECT.md into a high-capability AI model (Claude Opus, GPT-4, etc.). The AI reads the instructions and creates the planning file structure: master-plan.md, individual task files, and test files.
  3. User hands task files to Build Agents. Each build agent gets ONE task file and the shared context from master-plan.md. It claims the task, builds the code, and logs its work.
  4. User hands test files to Test Agents. Each test agent gets ONE test file and verifies the corresponding task output. It documents pass/fail results.
  5. Architect reviews. The user goes back to the Architect model with the completed plan and test results. The Architect signs off or creates fix-up tasks.
Why Three Separate Documents?

Each role has different capabilities and cost profiles. The Architect needs deep reasoning (expensive model). Build Agents need competent coding (mid-tier model). Test Agents only need to follow steps and report results (cheap model). By separating the instructions, you can use the right model tier for each role and keep costs manageable.

Below are the three documents that make up the planning system. Each tab shows the full markdown content that should be saved as a file. Click each tab to view and copy the content.

ARCHITECT.md

High-capability (Opus, GPT-4)

Sr. Architect: Receives the user goal and creates the entire plan: master-plan.md, task files, and test files. Thinks deeply, writes sparingly.

BUILD-AGENT.md

Mid-tier (Sonnet, GPT-4o)

Build Agent: Receives one task file and executes it precisely. Claims the task, builds the code, logs decisions. Does not plan or redesign.

TEST-AGENT.md

Lightweight (Haiku, GPT-4o-mini)

Test Agent: Receives one test file and verifies the build. Follows test steps literally, documents pass/fail results. Does not fix issues.

How to Set This Up (AI Guide)

  1. Create a planning/ folder in the project root.
  2. Save the three documents from the tabs above as planning/ARCHITECT.md, planning/BUILD-AGENT.md, and planning/TEST-AGENT.md.
  3. Add planning/ to .gitignore if you don't want project plans in version control (they contain temporary working state). Alternatively, keep them in git as a record of how the project was built.
  4. When the user has a new feature or project to build: paste the contents of ARCHITECT.md into a capable AI model along with the goal. The AI will create the planning folder structure.
  5. Hand individual tasks/task-NNN.md files to build agent sessions. Each session gets one task. Multiple agents can work in parallel on independent tasks.
  6. After builds complete, hand tests/test-NNN.md files to test agent sessions to verify the work.

Key Principles

  • Self-contained tasks: Each task file has everything the build agent needs. No cross-referencing other tasks.
  • Under 300 words per task: Shorter tasks have higher execution quality. If a task is getting long, split it.
  • Explicit outputs: “Create /src/auth/login.ts” not “implement authentication.”
  • File-based coordination: No shared memory. Status is tracked in markdown files. This works across any AI tool.
  • Right model for the job: Expensive models plan. Mid-tier models build. Cheap models test.
← Back to all resources