The Explore → Plan → Implement → Commit Workflow¶
For complex, multi-file tasks, always separate phases:
# Phase 1: Explore (plan mode — no edits)
# Press Shift+Tab to cycle permission modes into plan mode, or:
claude --permission-mode plan
> "Read src/auth and understand how sessions work.
Also look at how we handle environment secrets."
# Phase 2: Plan
> "I want to add Google OAuth. What files change?
What's the session flow? Write a detailed plan."
# Press Ctrl+G to open plan in editor before Claude proceeds
# Phase 3: Implement (exit plan mode)
> "Implement the OAuth flow from the plan. Write tests
for the callback handler. Run tests and fix failures."
# Phase 4: Commit
> "Commit with a descriptive message and open a PR"
Skip planning for small, clear tasks. Planning adds overhead. If you can describe the diff in one sentence, go directly to implementation.
What plan mode actually does¶
Plan mode is not just "ask before every edit" — it changes what runs at all. Reads and codebase exploration proceed; edits stay blocked until you approve a plan:
"Plan mode tells Claude to research and propose changes without making them. Claude reads files, runs shell commands to explore, and writes a plan, but does not edit your source. Except in sessions with bypass permissions available, edits stay blocked until you approve the plan."
— Anthropic, Choose a permission mode
What runs without asking in this mode: "Reads, plus classifier-approved commands when auto mode is available" — same source. Everything else prompts.
Entering and leaving¶
Per Anthropic's permission-modes reference:
| Action | How |
|---|---|
| Enter from a running session | Shift+Tab to cycle (default → acceptEdits → plan → default), or prefix one prompt with /plan |
| Start a session already in plan mode | claude --permission-mode plan |
| Edit the proposed plan before Claude proceeds | Ctrl+G opens it in your default text editor |
| Leave without approving | Shift+Tab again |
| Approve | Exits plan mode and switches to the mode the approve option names, so Claude starts editing |
Who does the exploring¶
Two built-in subagent types exist specifically for this phase, both read-only:
| Type | Use when |
|---|---|
Explore |
Fast codebase search; skips CLAUDE.md and git status for speed |
Plan |
Architecture planning in plan mode |
See Subagents for the full table and its footnote on why "Explore is cheap because it's Haiku" stopped being true by default in v2.1.198.
When planning is waste¶
The short version is above: skip planning for small, clear tasks. The Development Workflow has the fuller argument — match the ceremony to the blast radius, not to the task.
Related¶
- Permissions & Auto Mode — the full permission-mode reference
- Subagents —
ExploreandPlan, and every other built-in type - The Development Workflow — where this phase sits in the wider cycle
- Choosing Your Mode — plan mode among the other supervised-autonomy primitives