Plan-Driven Agentic Engineering

Traditional AI coding is often "prompt-to-patch"—asking for a single fix. This skill shifts the workflow to a "teammate" model, where the human and agent collaborate on a structured plan before execution to ensure the agent can handle long-running, complex tasks without drifting or losing context.

The Workflow

1. Collaborative Planning (The plan.md Phase)

Before asking the agent to write a single line of production code, align on the logic.

• Create a plan.md file: Write out the intended changes in markdown within your IDE.
• Context Loading: Explicitly provide the agent with relevant files, API documentation, or existing patterns (e.g., "Look at the iOS implementation to plan the Android port").
• Verifiable Steps: Break the plan into a checklist. Each step should have a clear "Definition of Done" that the agent can eventually verify itself.
• Iterate: Ask the agent, "Given this plan, what am I missing?" or "What are the biggest risks in this architecture?"

2. Execution Loop

Once the plan is finalized, move into execution.

• Delegation: Point the agent to the plan.md and instruct it to execute a specific subset of tasks.
• Sandbox Execution: Allow the agent to run code, tests, and shell commands within a sandbox to validate its own work as it goes.
• Compaction: If the task is long-running (exceeding the context window), instruct the agent to "summarize the current state and remaining tasks into a fresh context" to maintain focus.