If you’re constantly going back and forth correcting Claude, I hate to break it to you, but you’re the problem. These are the 4 moves I use to stop correcting and actually delegate.
Most people use Claude like a slot machine. Type a quick request, pull the lever, hope. It builds the wrong thing, they correct it, it builds another slightly wrong thing. Twenty messages later the result is mediocre and they’re exhausted.
The people who save hours spend a few minutes getting aligned up front, then let Claude run. These moves are built for Claude Code. Move 2 works in the regular Claude app too.
01
Make it show you the plan
In plan mode, Claude reads what it needs, thinks through the approach, asks you questions and hands you a written plan without editing your files. You approve it, then it builds.
Press Shift + Tab until the status bar says plan mode on, or start a single message with /plan. Get the plan right and the mistakes get caught on paper instead of in your files.
02
The 95% rule
Half the time Claude builds the wrong thing, it filled a gap you left with a guess. Add this line to almost any request: "Ask me questions until you are 95% sure exactly what I need. Do not assume anything."
Now Claude surfaces the decisions you should be making, like which file to touch or what happens in the edge case. A few quick answers up front delete whole rounds of "no, not like that."
03
Turn on auto mode
Once you’ve approved the plan, you don’t need to approve every step. Auto mode runs the task start to finish while a separate safety check reviews each action and blocks anything that goes beyond what you asked. It’s the starting mode in current versions of Claude Code, and Shift + Tab gets you back to it.
This works because you did Moves 1 and 2 first. You’re not letting it run wild on a vague request. You’re letting it carry out a plan you already approved.
Do This Safely
Skip bypass permissions on your real computer. Anthropic says to use it only in isolated setups like containers or VMs, because it has no protection against unintended actions. Auto mode also isn’t a guarantee, so set up an allow list and a deny list first. Deny rules still apply in every mode. My exact lists are in The Claude Code Safety Hack.
04
Keep a living instruction file
CLAUDE.md is the file Claude Code loads at the start of every session to learn your project, your preferences and how you work. The better it is, the less you explain. This is the move that compounds.
Any time you correct a recurring mistake or agree on a new way of doing things, say "add that to CLAUDE.md." Keep it under about 200 lines, which is Anthropic’s own target, because longer files eat context and get followed less. When you add something new, delete something old. Running /doctor prompt-audit flags outdated or conflicting lines for you.
New To These Files?
A CLAUDE.md is a plain text file with simple formatting. If the whole "md file" thing is fuzzy, here’s the 5-minute crash course plus a template: What’s an .md File?
This is how I open a real task. Paste it as your first message, fill in the bracket, and it triggers Moves 1 and 2 automatically.
Answer its questions, because that step is where the time gets saved. Approve the plan, then let it run. If you do one kind of task all the time, save this as a slash command so you never retype it.
Delegating isn’t losing control. You make the decisions up front, where they’re cheap and fast, then let Claude carry them out. A 20-message correction marathon turns into a few good questions and a finished result.