Claude Code has three major modes: edit mode, edit mode with auto-approval, and plan mode. After using the plan mode extensively for a while, I think it is a super underrated feature. It’s actually very very good. I now use it almost exclusively when I start working on a new feature.
The Workflow
- Shift + Tab twice to enter the plan mode
- Brainstorming with Claude, this can be exploring a different solutions, solicit feedback on an implementation, or research on a problem space.
- I use the
@reference to help Claude with additional context so it doesn’t need to spend a lot time exploring the codebase. - For convenience, I sometimes also connect Claude Code to VS Code by using the
/ideslash command (see IDE integration). I open a file in VS, select the lines, and ask Claude Code about the lines. - I iterate with Claude Code until I am happy with the solution. After that, I Shift + Tab twice to enter the auto edit mode. At this point, I simply ask Claude to implement the solution with very little intervention.
The key for this to be effective is to keep the scope small. Plan what you will do in the next 30 minutes or less.
The workflow should be:
- plan > code > debug > commit
- plan > code > debug > commit
- plan > code > debug > commit
… and so on.
If you multi-task well, you can even plan for the next feature in another session while waiting for Claude to finish coding the current one.
Why I Like It
I find that with this approach, I don’t even need to create a plan document (such as PLAN.md) for features anymore. I try to keep the feature iterations small, and commit the changes as soon as the code is working. By doing that I think I accidentally followed the trunk-based git workflow.
What makes this workflow so good is that I get to review what Claude Code is going to do before it does it. I get to see the complete plan, and if I see something that I don’t like, I can ask Claude Code to change it. It’s much easier to change the plan than to change the code during coding. If you vibe code a lot, whether in Cursor, Cline, or anything else, you know exactly what I mean.
Things to Consider
This is not a silver bullet. It has drawbacks. For larger planning sessions, you probably still want to save the plan to a file, such as docs/PLAN.md. See Claude Code Memory Management for more details. For small feature iterations, plan mode + auto-edit mode works beautifully. Claude Code is able to keep the relevant information in the context window.
For longer planning sessions, I prefer ChatGPT’s Canvas feature. This is where I create documents like PRD.md, ARCHITECTURE.md, etc. That’s what I usually do when I start a new project.
If you want to use Canvas on an existing project, you could use Repomix to create a XML dump of your codebase, and upload it to ChatGPT.
That’s it for now, folks!