3 Claude Code prompts I actually send: before code, in circles, overnight

These are the three prompts I type into Claude Code most often, word for word. The first I send before it writes any code, so we agree on a plan first. The second is for when it keeps trying the same fix. The third is for when I leave it working overnight. Each one is short on purpose. Change the task in the first line, keep the rest, and copy it with the button.

By Dmitry Mishunin, who builds Meowdoro · November 8, 2026

Before any code
When it goes in circles
Before I go to bed
Make them yours
Questions

1. The prompt I send before Claude Code writes any code

I want to add dark mode to the settings page. Don't write any code yet.

First, read the files this touches. Tell me in a few lines how it works now.

Then write a plan: what you'll change, in which files, in what order. And what could break.

Write the tests first. Tell me which ones should fail until it's done.

If anything is unclear, ask me. Don't guess.

We'll change the plan until it's right. When I say go, follow it. If it goes wrong halfway, stop and tell me.

It’s tempting to let Claude start coding right away. In my experience it’s slower. Why each line is there:

“Don’t write any code yet.” Without it, Claude starts editing on the first turn, and you review code for a plan you never agreed to.
“Read the files… tell me how it works now.” If its summary is wrong, the plan will be wrong too. This is the cheapest place to catch that.
“What could break.” It makes Claude look past the happy path before it starts.
“Write the tests first.” The failing tests are the definition of done, and you can read them in a minute.
“Ask me. Don’t guess.” A guess turns into code you have to undo.
“When I say go… stop and tell me.” You decide when the work starts, and a run that goes wrong stops instead of improvising.

I change the plan up to ten times before I say go. That’s where most of my time goes, and it pays off. Claude Code has a mode for exactly this: type /plan, or /plan followed by the task, and it switches to plan mode. Claude reads the files and writes a plan, but doesn’t edit your code until you approve it. Shift+Tab switches to plan mode too.

2. When Claude goes in circles

Stop. You've tried this three times and it's not working.

Don't write any more code.

Tell me what you think the problem is, in two lines.

List what you've tried, and why each one failed.

Give me three different ideas, not variations of the last one.

I'll pick one. Then go.

It stops the loop. No new code until you’ve both looked at the problem again.
Two lines for the problem. If Claude can’t say it in two lines, it doesn’t understand it yet, and now you know.
Why each attempt failed. It turns three failed fixes into information instead of noise.
Three different ideas. “Not variations of the last one” is the important part: it pushes Claude off the path it’s stuck on.

If the session has grown long while it went in circles, the failed attempts are still in the context and keep pulling it back. Then /rewind to before the first attempt, or /compact after you pick an idea. When to use which.

3. The prompt I send before I go to bed

I'm done for today. You're not.

Keep working through the night. Follow the plan step by step, and run the tests after each one.

Stop only if it's critical: something you can't undo, or a change to the plan itself. Then write the question at the top of your notes and wait for me.

Everything else, decide yourself and write down why.

Commit what works. Leave what doesn't on a branch, with a note.

In the morning, show me what you finished, what broke, and where to start, small enough to start with a coffee.

Thanks. See you tomorrow.

It needs a plan. Send it after the first prompt, once the plan is agreed. A night without a plan is a night of guessing.
Only two reasons to stop. Something it can’t undo, or a change to the plan. Everything else gets decided and written down, so you can check the decisions in the morning.
Working code stays separate from experiments. What works is committed. What doesn’t is left on a branch with a note.
A morning summary you can act on. Finished, broken, and the first small step.

Two things to check before you leave. Claude Code still asks before actions your permission settings don’t allow, so a run can sit at the first prompt all night: allow what the plan needs, and nothing that can’t be undone. And a long run uses your plan’s limits, so it may stop when the five-hour window runs out (how the limits work).

Make them yours

Change the first line, keep the rest. The task changes every time. The structure is the part that works.
Save the ones you use daily. A prompt you paste every day can become a skill you run with a slash command (how skills work).
They work in Codex too. They’re plain instructions, not Claude Code features.

Where Meowdoro fits

We build Meowdoro, a focus timer around the MacBook notch, and these prompts come from how we work with it. Once the plan is agreed and Claude is working, I don’t watch it. The notch shows one quiet line when a session finishes or asks me something, and during a break nothing reaches me at all. In the morning, after the overnight prompt, the island shows which sessions finished and which are waiting for an answer. Try it and tell us what’s missing.

Questions

What’s a good prompt to start a task in Claude Code?

One that asks for a plan before any code: read the relevant files, explain how it works now, list the changes and what could break, write failing tests first, and ask instead of guessing. The full prompt is above; /plan keeps Claude in plan mode while you work on it.

How do I stop Claude Code from going in circles?

Stop it from writing code, ask it to state the problem in two lines and why each attempt failed, then ask for three genuinely different ideas and pick one. If the context is full of failed attempts, /rewind or /compact first.

Can I leave Claude Code running overnight?

Yes, with an agreed plan, permissions that allow what the plan needs, and clear rules for when to stop. It also uses your plan’s usage limits, so a long run may pause when a window runs out.

Do these prompts work in Codex?

Yes. They’re plain text and don’t depend on Claude Code features.

The prompts are the ones I send, as typed in our “Typing” Reels on Instagram. The task in the first prompt is an example; replace it with yours. Plan mode and permissions: the Claude Code documentation on commands and permission modes, read on November 8, 2026.