Fable Brain

Explore Plan Code Commit

5 min read

Explore Plan Code Commit

Explore Plan Code Commit is the officially recommended Claude Code loop: research in read-only plan mode, approve a written plan, implement against a runnable check, then commit, and skip the plan stage entirely for one-sentence diffs.

What it is

The canonical agentic coding workflow from Anthropic’s Claude Code best practices. The original engineering post at anthropic.com/engineering/claude-code-best-practices now 308-redirects to code.claude.com/docs/en/best-practices, which is the current official guide. The workflow’s four stages separate context gathering from mutation so the model reads before it writes and verifies before it ships.

How it works

  1. Explore. Enter plan mode (Shift+Tab, a /plan prompt prefix, or claude --permission-mode plan). Plan mode is read-only: “Plan mode tells Claude to research and propose changes without making them.” (https://code.claude.com/docs/en/permission-modes). Delegate scoped investigation to subagents so research does not consume the main context window.
  2. Plan. Claude proposes a plan. Ctrl+G opens it in a text editor for direct edits before proceeding. Approving the plan switches the session into the chosen execution mode.
  3. Code. Implement against a verifiable check. The docs’ top practice: “Give Claude a check it can run: tests, a build, a screenshot to compare.” (https://code.claude.com/docs/en/best-practices). Checks can gate harder via /goal conditions, Stop hooks, or a fresh-context verification subagent.
  4. Commit. Land the change once the check passes, keeping git history descriptive so later sessions can reconstruct intent.

The docs also carve out the fast path: “If you could describe the diff in one sentence, skip the plan.” (https://code.claude.com/docs/en/best-practices).

Key insight

Each stage carries a different permission posture. Explore runs read-only, implementation runs in the chosen execution mode, and the six permission modes (default Manual, acceptEdits, plan, auto, dontAsk, bypassPermissions) are cycled with Shift+Tab or set via —permission-mode and the defaultMode setting. The loop is a permissions choreography as much as a prompting pattern.

Best practice

  • Give Claude a check it can run (tests, build exit code, linter, screenshot comparison) so it iterates until the check passes instead of using you as the verification loop. EVIDENCE-BASED
  • Use plan mode for anything non-trivial; approving the plan switches into the chosen execution mode, so the mode transition is explicit, not accidental. EVIDENCE-BASED
  • Skip planning overhead for small, clearly scoped fixes describable in one sentence. EVIDENCE-BASED
  • Edit the plan directly with Ctrl+G before approving rather than steering with follow-up prompts. EVIDENCE-BASED
  • Run /clear between unrelated tasks so stale exploration does not pollute the next loop. EVIDENCE-BASED
  • After roughly two failed corrections on the same issue, restart: “A clean session with a better prompt almost always outperforms a long session” (https://code.claude.com/docs/en/best-practices). EVIDENCE-BASED
  • Enforce mandatory rules with hooks and permission rules, not CLAUDE.md prose; keep CLAUDE.md short and ruthlessly pruned. PRACTITIONER
  • Gate completion deterministically: Stop hooks or a fresh-context verification subagent catch premature “done” claims. EVIDENCE-BASED
  • Set the project’s defaultMode in settings so every session starts in the intended stage posture instead of relying on manual Shift+Tab discipline. EVIDENCE-BASED
  • In the VS Code extension, review the plan and the inline diffs in the editor; plan review and checkpoints are built into the recommended IDE integration. EVIDENCE-BASED

Pitfalls

  • Relying on the human as the verification loop; without a runnable check Claude cannot close its own loop.
  • Bloated project memory: “Bloated CLAUDE.md files cause Claude to ignore your actual instructions!” (https://code.claude.com/docs/en/best-practices).
  • Applying full planning ceremony to one-line fixes; the docs explicitly say to skip it.
  • Writing a policy hook that exits 1: “Exit 1 does not block anything.” (https://blakecrosley.com/blog/claude-code-hooks-explained). Blocking hooks must exit 2.
  • Grinding through a long correction session instead of restarting with a better prompt.
  • Confusing plan-mode approval with permission approval; the approval also selects the execution mode for the rest of the session.

Sources

Next actions

  • Wire a Stop hook that runs the project test suite so the commit stage cannot fire on red.
  • Compare plan quality with and without a prior subagent exploration pass on one real task.
  • Add a repo checklist linking each stage to its permission mode and hook events.

Graph View

Backlinks