Decomposition & delegation

Do not ask an agent to "build an app". Ask it to "write the database schema."

The idea

AI models have a context window and a limited ability to plan 20 steps ahead. If you give them a monolithic task (e.g. "Migrate our database"), they will hallucinate or get stuck.

You must Decompose the task into verifiable, atomic steps. Instead of one massive prompt, you generate a Build Plan. You insert Human Review Gates between critical steps (e.g., verifying the schema is correct before letting the AI write the migration scripts).

"Migrate Database to Postgres"

(Click to Decompose)

Execution Risk:

HIGH RISK: Handing a massive task to an agent in one prompt guarantees failure.

How it works (The Planner/Executor Pattern)

# The workflow for complex tasks:

# 1. Planning Phase (Agent 1)
prompt = "Analyze the repo and write a markdown checklist of steps to migrate."
plan = agent.generate(prompt)

# 2. Human Review (YOU)
# You review the plan, modify steps, and ensure the schema translation is correct.

# 3. Execution Phase (Agent 2)
for step in plan:
    # We pass ONLY the current step and the overall context
    execute_prompt = f"Implement step: {step}. Context: {repo}"
    agent.generate(execute_prompt)
    # Allows for verifiable checkpoints!