Guide 05

Work in small phases, not one giant AI prompt

Agent work becomes more reliable when each phase has a clear goal, bounded scope, and concrete verification.

Why small phases work better

Large open-ended tasks make it difficult to know whether the agent is finished, what changed, and where a failure happened. Small phases provide better control and easier recovery if something stops.

Example phase split

  • Environment check and dependencies.
  • Database or configuration.
  • Feature or bug fix.
  • Tests and build.
  • Deployment and live verification.

What a phase should include

  • One clear goal.
  • Which files or areas the agent can change.
  • What the agent must leave alone.
  • Which commands or checks should run.
  • A stop rule: "Stop after this phase and report."

Report after each phase

Ask the agent to summarize changed files, commands run, pass/fail results, blockers, and the recommended next step. Do not continue just because the agent says everything is done.