Claude Code Plan Mode: Get a Plan Before It Touches Your Code
Engineering
Claude Code Plan Mode: Get a Plan Before It Touches Your Code
The cheapest bug to fix is the one still in a sentence. Plan mode makes Claude Code research your codebase and lay out its approach before it edits anything — so you catch a wrong direction while it's a paragraph you can veto, not a tangle of broken files you have to unwind. Here's how to use it.
What plan mode does
In plan mode, Claude Code can read, search, and think — but it can't write, edit, or run commands that change things. It investigates the task, then presents a plan: what it intends to do, in what order, touching which files. You read it, approve it (or send it back), and only then does it start executing. Research first, act second.
How to turn it on
Cycle your input mode with Shift+Tab — Claude Code rotates through its modes, including plan mode (watch the indicator by the prompt). Ask your question in plan mode and, instead of diving in, it comes back with the plan. Approve it to switch into execution.
When it earns its keep
Plan mode pays off most when:
- The task is non-trivial or spans several files — sanity-check the approach before it fans out.
- You're in unfamiliar code — the plan doubles as a map of how the agent understands the codebase, and surfaces where it's wrong.
- The cost of a wrong direction is high — a migration, a refactor, anything hard to unwind.
- You want to steer, not supervise — approve a good plan and let it run instead of watching every edit.
For a one-line fix, skip it. For "refactor the auth flow," always.
Plan, execute — then watch
Plan mode fixes the front of the loop (a wrong approach). The back of the loop is execution drifting anyway — the plan was fine, but halfway through it got stuck. That's where you want eyes.
Mwah is those eyes: once an approved plan is running, every live Claude Code session floats on your desktop as a robot showing its live activity, ALIVE or STUCK — so you see the drift the moment it happens, feed the ones that nailed it, and retire the one that lost the thread. Plan up front; watch the run.
Related: Claude Code stuck in a loop? · Run multiple Claude Code sessions
Keep reading
View all dispatches →Shopify AI Agent Tools: What Your Agents Can Actually Do in Your Store
Clearly now ships Shopify store tools your AI agents can call — browse the catalog, read orders, roll up revenue, edit products, and attach generated images — over MCP and the beehaven CLI, with a dry-run preview on every write. Here are the tools, with example commands.
How to Manage Your Shopify Store with AI Agents
AI agents can now do the work of running a Shopify store — writing descriptions, generating product images, answering customers, updating the catalog. Here's how agent-run commerce works, and how to set it up.
Claude Code for Shopify: Run Your Store from Your Terminal
If you already run your codebase with Claude Code, the same agents can run your Shopify store — edit the catalog, generate product images, answer customers — with a connector and the right guardrails. Here's the setup.