CHECK 06 15 MIN · HANDS-ON

Explore → Plan → Code → Commit

What you’ll know by the end of this check

  • The four-step workflow that separates power users from one-shot promptscrollers
  • How to set success criteria Claude can actually measure against
  • Why the subagent code reviewer is the single best pre-commit habit

The shortest possible answer

Explore → Plan → Code → Commit. If you take one thing from this whole checklist, take this.

Most people skip straight to “Code.” They type “fix this bug” and then spend an hour re-steering. Walk the four steps once and you’ll never go back.

The four steps

1. Explore

Claude needs context to be right. Ask it to read the relevant files, run the tests, check the existing patterns. The cheapest way to do this is Plan Mode — it reads, it doesn’t write.

2. Plan

In Plan Mode, Claude reads what it needs, asks clarifying questions, and returns a proposed plan. Read it. Push back on it. Revise specific areas by saying “skip step 3, do X instead.” This is the best place to course-correct because no code has been written yet.

3. Code

Approve the plan. Claude executes. You choose whether it auto-accepts file edits or asks each time. Three tips that make this phase dramatically smoother:

  • Define success criteria. “Correct” has to be concrete. A test that passes, a screenshot that matches, an output that parses. Put it in the plan.
  • Add tools. Building web UIs? Install the Claude in Chrome extension so Claude can drive a browser tab. Data work? Give Claude access to the CLI. The less Claude has to guess, the tighter the loop.
  • Include a test suite. Let Claude validate against it continuously. If tests don’t exist, ask Claude to write some first — but verify they’re not false positives before trusting them.

4. Commit

Before you push, run a subagent code reviewer. A subagent gets a fresh context window — it doesn’t carry the main session’s bias. It catches what you miss when you’ve been staring at the diff for an hour.

Then ask Claude to generate a commit message that matches the style of recent commits in the repo. Review it. Push it.

The failure mode this prevents

Skipping Plan Mode and jumping to Code looks faster. It isn’t. You pay the time back in “wait, that’s not what I meant” cycles, and every one of those cycles leaks context. The four-step workflow front-loads the thinking so the coding is boring.

Things to try right now (15 minutes)

Pick one small real feature. A toggle. A validation rule. A migration. Then:

  1. claude → Shift+Tab to Plan Mode.
  2. Prompt with the shape from check 03: context + goal + success criteria.
  3. Review the plan. Push back on at least one line.
  4. Approve. Watch Claude code.
  5. Before you commit, run a subagent code review: /agents → code-reviewer → point it at your diff.
  6. Commit.

Write down one thing the subagent reviewer caught that you would have merged.

The canonical version

Full official lesson is at anthropic.skilljar.com/claude-code-101/469792.

Ready to verify this check?

You’ve done the full loop on a real change. You’ve seen a subagent review catch something. You know the four steps in order. Mark it cleared.