23d (edited) • Posts
I spent MONTHS hand-tuning Claude Code permissions. Then the guy who built it told me I didn't have to.
I'm going to be honest about something I'm a little embarrassed by.
I spent weeks — months, really — hand-tuning my Claude Code permissions.
You know the pop-up. Every time Claude wants to do something real — run a command, hit an MCP, query my Data Hub — it stops and asks: "Can I run this? Yes / No?"
It drove me nuts. So I went to war with it. Allow-lists. Deny rules. Patterns I tested, refined, broke, and fixed again. I was basically writing a rulebook to teach Claude what it was allowed to do without bugging me. And it kind of worked — I was proud of it.
Then I was listening to Boris Cherny — the guy who actually built Claude Code at Anthropic — do an interview. And he described the exact problem I'd spent months solving. Except he had a much better answer than mine.
THE PART THAT STOPPED ME
Boris said the permission pop-up had a problem nobody expected: "Over time you get kind of lazy... at least for me, I just kept saying yes. I wasn't really reading the commands."
That was ME. After enough prompts, you stop reading. You just hammer "yes." He said Anthropic's own security team flagged it — the human-in-the-loop was supposed to make things safer, but prompt fatigue turned everyone into a reflexive yes-clicker, which actually made it LESS safe.
So they built a different answer. He called it auto mode: "Every permission prompt, we route to a model, and the model decides yes or no based on what you've said so far in that conversation."
Instead of YOU rubber-stamping every action (or me writing a 200-line rulebook), a model checks each action against what you actually asked for and approves the routine stuff automatically.
Then the two lines that got me:
"This is what we use internally at Anthropic, and the vast majority of our users use it also."
"Because you don't have to sit there saying yes or no, I can now run Claude for hours, or for days."
That last one is the whole game. The pop-up wasn't just annoying — it was the leash. As long as you have to babysit yes/no, you can't ever walk away. Kill the babysitting, and you can set Claude loose on something big and actually go do something else.
THE KICKER FOR ME
I got that result the HARD way. Months of permission tuning to stop the pop-ups.
Auto mode gets you the same benefit — without the months. No allow-lists to maintain. No rulebook. You flip one setting and the model handles the judgment call I was trying to hard-code by hand.
HOW TO TURN IT ON (30 seconds)
Open your user settings file at ~/.claude/settings.json (on Windows: C:\Users\<you>\.claude\settings.json) and add:
{ "permissions": { "defaultMode": "auto" } }
Save, restart Claude Code. Done. (Or just tell Claude "set my default permission mode to auto" and let it do it for you.)
A few things I learned setting it up:
- It has to go in your USER settings file, not a project's. Claude Code deliberately ignores "auto" if a project tries to set it — so a random repo can't grant itself free rein.
- Your hard deny rules still apply. rm -rf, force pushes, DROP TABLE — if you've blocked them, they stay blocked. Auto handles the routine stuff; the dangerous stuff still stops.
- Needs a recent Claude Code version and a current model. On Team/Enterprise your admin may need to enable it.
WHY I'M ASKING YOU TO TEST IT
I'm trying this on myself right now, swapping out the elaborate permission setup I spent months building.
Boris and the whole Anthropic team swear by it. But I want to know if it holds up for US — people running real Amazon workflows, real MCP queries, real money on the line.
So here's the ask. Turn on auto mode. Run it for a few days. Then come back to this thread and tell me:
- Did the pop-up fatigue actually go away?
- Did you feel like you LOST any control, or did it feel fine?
- Did it let you kick off something long-running and walk away?
Comment "I'm in" below if you're going to try it, and report back once you've lived with it. I'll compile what everyone finds.
Interview source: "Fireside Chat with Boris Cherny, Head of Claude Code" — https://www.youtube.com/watch?v=Z47vatpsGPI
0
0 comments
I spent MONTHS hand-tuning Claude Code permissions. Then the guy who built it told me I didn't have to.
skool.com/sellerlabs
AI for Amazon sellers. Automate PPC, inventory, listings, and reports with AI agents and MCP. Built for serious Amazon operators.
Leaderboard (30-day)
1
+1
Powered by