Skip to content
Tools·5 min read·English

Your Claude Code loop is stuck waiting for a yes. The one line that gets it working again

Since the November 13 update, a loop that runs a skill with allowed tools stops and asks. How to tell if it hits you, the exact settings line and where it goes, and when not to approve.

PART 01

What changed on November 13

Claude Code 2.1.325 brought one change that matters for loops: a loop or a scheduled task now stops and asks before it runs a skill that declares its own allowed tools or hooks.

I tested it on my machine. I started a loop that runs every minute on a skill that writes a morning report. This came up:

Copy
> /loop 1m /morning-report

Use skill "morning-report"?

❯ 1. Yes
  2. Yes, and don't ask again
  3. No

A minute went by, then another. No report. The loop did not fail and showed no error. It just waits for someone to say yes.

PART 02

Does it hit you

It hits you if all three are true:

  • You have a loop (/loop) or a scheduled task that runs a skill.
  • The top of that skill's SKILL.md has a line starting with allowed-tools: or hooks:.
  • Claude Code runs in the normal permission mode, the one that asks before actions.

This is the top of the skill from my test. The third line is the one that triggers the question:

Copy
---
name: morning-report
allowed-tools: Bash(date), Bash(ls), Write
---

In the mode that skips all permissions (bypass permissions) the skill ran with no question. The change shows in the normal mode.

Your Claude Code loop is stuck waiting for a yes. The one line that gets it working again

Contents

  1. 01.What changed on November 13
  2. 02.Does it hit you
  3. 03.The fix, step by step
  4. 04.How to know it worked
  5. 05.When not to approve
  6. 06.A loop nobody is watching
PART 03

The fix, step by step

The fix is one line in the settings file that approves this skill ahead of time. After it, the loop runs it without asking.

Find the skill's name. It is the name in quotes in the question: Use skill "morning-report"?. The same name is on the name: line at the top of SKILL.md.
Open the project's settings file. In the folder where the loop runs there is a folder named .claude, and in it a file named settings.json. If it does not exist, create it. The folder name starts with a dot, so on a Mac Finder hides it: press Cmd+Shift+. to show it.
Paste the block. If the file is empty, paste all of this, and replace morning-report with your skill's name:
Copy
{
  "permissions": {
    "allow": ["Skill(morning-report)"]
  }
}

If the file already has "allow", add only this line inside the square brackets, with a comma before it:

Copy
"Skill(morning-report)"
Save, and close the old loop. In the Claude Code window pick option 3. No on the waiting question, then start the loop again with the same command.

Want the approval in every project, not just this one? The same block goes in ~/.claude/settings.json in your home folder. I prefer the project file: the approval stays next to the skill I checked.

And if you are at the screen when the question comes up? Option 2. Yes, and don't ask again saves an approval. I did not check which file it writes to, so I go with the line in the file, which I can see and control.

PART 04

How to know it worked

On the next run there is no question. On my machine it looks like this, and the report gets written:

Copy
⏺ Skill(morning-report)
  ⎿  Successfully loaded skill · 3 tools allowed

⏺ Write(report.md)
  ⎿  Wrote 6 lines to report.md

Still asking?

  1. Type /permissions in Claude Code. It lists every rule and the file each one comes from. If Skill(morning-report) is not there, the file was not read.
  2. Check that the name inside Skill(...) matches the name in the question, letter for letter.
  3. Check that the file is in .claude/settings.json of the folder the loop runs in, not another folder.
  4. Paste the whole file into Claude Code and ask it to check that it is valid JSON. One missing or extra comma breaks the whole file.
  5. Start the loop again. A loop that is already waiting for an answer does not reread the settings.
PART 05

When not to approve

The new question was not added to annoy you. A skill with allowed-tools gets to run the tools it declares. In a loop, nobody is at the screen when that happens.

  • Approve a skill you wrote, or one whose SKILL.md and requested tools you read.
  • Do not approve a skill you installed from a plugin or the internet and have not read. Read first, approve after.
  • Watch for a skill that asks for Bash with no limit, plain Bash rather than Bash(date). It can run any command on your machine.

The approval is by name. If someone changes what the skill contains, the approval stays. One more reason to give it only to your own skills.

PART 06

A loop nobody is watching

Some loops run at night, when nobody will answer a question. There it is better for the loop to skip the action and keep going than to wait until morning.

Version 2.1.319 added a setting named denyInsteadOfPrompting. Instead of stopping on a question, Claude Code skips the action and keeps going. A skill you did not approve does not run, and the loop does not get stuck.

This setting only goes in your personal settings file, ~/.claude/settings.json in your home folder. A project file cannot turn it on for you. The block:

Copy
{
  "permissions": {
    "denyInsteadOfPrompting": true
  }
}

If the file already has "permissions", add only the line "denyInsteadOfPrompting": true inside it, with a comma before it.

Note: this applies to every question, not only skills. An action that would have waited for your approval just will not happen. The two work together: approve ahead of time the skills you trust, and deny everything else.

One more change from the same update: Bash commands that run inside the sandbox now ask like any other command. To keep the old behavior, add "sandbox": { "autoAllowBashIfSandboxed": true } to the settings file.

code.claude.com/docs/en/changelog (2.1.325, 2.1.319), code.claude.com/docs/en/settings, code.claude.com/docs/en/permissions, code.claude.com/docs/en/skills, checked on November 15, 2026. My test: Claude Code 2.1.325 in the normal permission mode, November 14, 2026.

Daniel Goldman

AI-native products, workshops, and automations. Built from everywhere.

© 2026 Daniel Goldman