Claude Code approval fatigue: what the numbers say and what to change

You click Allow on the twentieth npm install prompt without reading it. That is fine. Then the twenty-first prompt is a git push --force to the wrong branch, and your finger clicks Allow before your eyes get there. That is approval fatigue, and Anthropic has now published numbers on it.

Anthropic says users accept 93% of Claude Code permission prompts (March 2026), and later 97% (August 2026). Neither figure comes with a published method. In a controlled test with 1,053 paid testers, people blocked about 17% of dangerous commands early in a session and about 5% after 50 or more prompts. The prompt was still on screen. The attention was gone.

What the numbers say

Every figure below comes from Anthropic. Three of them have no published method, so read those as Anthropic's own statements.

Anthropic names the problem in the auto mode post: "Over time that leads to approval fatigue, where people stop paying close attention to what they're approving." The sandboxing post says the same thing from the other side: constantly clicking approve "slows down development cycles."

A yes to everything is a rubber stamp

Look at the controlled test again. A person was shown a prompt, and the prompt was dangerous. Most of the time they let it through. The longer the session, the more often they did.

So a prompt you see 50 times a day does two things. It slows you down, and it trains you to stop reading. By the time something risky shows up, the habit has already decided. Developers say this in plainer words. One Hacker News comment from Jan 20, 2026: "Once approval fatigue and ongoing permission management kicks in, the temptation is strong to run --dangerously-skip-permissions." Another, from Jun 23, 2026: "per-call approval you stop reading by day two."

The usual ways out each trade something away:

• Skip permissions. --dangerously-skip-permissions ends the prompts by ending the check. Nothing asks, so nothing gets caught. The Claude Code skip permissions guide covers what it turns off and a safer setup.
• Auto mode. A classifier decides instead of you. Anthropic reports it caught 89% of the dangerous commands in its test, and it hands the session back to a person after 3 denials in a row or 20 in total. You still need a way to hear about that when you are not at the screen.
• Babysit. Stay at the terminal and click. This is the 97% world, and the numbers above say how well it works.

Ask fewer, better questions

The goal is a small number of questions you actually read, each one worth your attention.

1. Let rules decide the routine ones

Most prompts are the same commands again and again: git status, npm test, ls, a lint run. If you have never said no to a command, it should not ask you. Write a rule that approves it.

The risky list is short. Anthropic's autonomy research found that only 0.8% of actions on its public API appear irreversible. Force pushes, deletes outside the project, deploys, production migrations, payments and messages to other people belong on it. Everything that is not on either list can ask.

A permission policy holds both lists in one place, for every agent you run. See one permission policy for every agent and the four levels of AI agent oversight, where policy is the level that decides without you.

2. Make each question answerable at a glance

A prompt that says "Claude is waiting for your input" is a beep. A question that shows the command, the repo and the agent is a decision you can make from a lock screen. When you run three sessions at once, it also tells you which one needs you.

3. Send the ones that matter to the person, where they are

You are not always at the terminal. Put the important prompts on your phone, with yes and no on the notification. Then you can walk away from a long run, and the first risky step comes to you instead of waiting for you to come back.

4. Escalate only when the question is missed

A single push can sit unseen under a group chat. Sending more pushes only adds noise. What helps is one louder try, and only when the first one is missed. Pushary sends the question to your phone first and, if it is still unanswered after your Call after time, calls you so you can answer on the call. For Claude Code the default is 2 minutes. It rings only inside your calling hours (08:00 to 22:00 by default), and a call that falls outside them waits for your hours to start if the question is still open. At most 5 calls in any 24 hours by default, and never once the question is answered. Calls are a beta: numbers in Germany, and the calls are in English. The Claude Code phone call page has the setup.

5. Keep a record

If rules decide most prompts, you need to see what they decided. Keep every question, every answer and who gave it, so a bad rule shows up in the record before it shows up in production.

Put together, that is ask, wait, escalate, record: ask only when it matters, wait for a real answer and let silence count as a yes only where you wrote a rule that auto-approves on timeout, call only when the question is missed, and keep every answer.

Check your own numbers

You do not need Anthropic's data to find your own fatigue. For one day, count the permission prompts you see and the ones you said no to. Then look at the commands you approved every single time. Each of those is a rule waiting to be written. What is left is the list of questions you should actually read.