Planning, Iteration, and CI Automation

Claude Code can either execute tasks directly or produce a plan first and wait for approval before making any changes. The choice between these modes is an architectural decision with real consequences: direct execution is fast, but irreversible changes happen immediately. Plan-first gives the developer a checkpoint to catch wrong assumptions before they become wrong files. This lesson covers both modes, when each is appropriate, and how to run Claude Code in automated pipelines without interactive sessions. By the end of this lesson, we will be able to:

• Explain the difference between direct execution and plan-first mode in Claude Code

• Identify the task types where plan-first mode prevents expensive mistakes

• Use claude --print to run Claude Code non-interactively and capture structured output

• Sketch a CI automation pipeline that uses a project command to produce reviewable output

Direct execution vs. plan-first

In the default interactive mode, Claude Code reads a task, decides what to do, and starts executing: reading files, writing code, and running shell commands. The developer sees each tool call as it happens and can deny individual steps. Plan-first mode changes this sequence. Claude Code produces a written plan describing what it intends to do and then stops. No files are changed. No commands are run. The developer reads the plan, requests changes if needed, and then approves execution. Only after approval does Claude Code begin making changes.