Howdy wizards,
I recently saw a quote that describes how my work feels (and maybe yours too):
“an army of Claudes, with every one asking for my attention”.
That comes from a study Anthropic did on its own employees.
We are the bottleneck, not the AI. Our attention and judgement are the scarce resource now.
They also found that the average Claude session sits around 80%(!) idle, waiting for our input.
That got me wondering: what exactly are all these Claudes waiting for?
So I decided to analyse my own Claude data to figure it out. A little personal research project to pinpoint the friction, so I can remove it and get interrupted less.
If you can't measure it, you can't improve it.
I asked Claude to store all my chat data. Then I had it assign a reason to every single time a chat stopped and pinged me — either the task was finished, Claude legitimately needed me to make a decision, or it was what I consider an avoidable interruption.
Like Claude needing access to something non-sensitive, or stopping just to check in with me.
Here's my personal view of how often and why Claude Code stopped its work in the last 3 months:
I've labelled some of the actual projects I worked heavily on in the bottom part of the chart. I'll get back to one of them in a bit.
I also marked the release of Opus 5.5, to see if it changed anything. The clearest change was that I started asking Claude to shorten or clarify its answers less often. The new model is noticeably more succinct.
Ideally, my agents would never interrupt me for the same thing twice. But I'm not gonna lie, teaching them is a lot of work to do by myself. Luckily, there are really good products that make it easy to build agents that learn from experience:
IN PARTNERSHIP WITH ZEP
Zep Agent Skills turns your agents' traces into step-by-step skills they read at runtime, over MCP or the Zep SDKs.
More tasks done right: agents start from steps that work instead of from zero.
Fewer tokens: agents take fewer steps, and a smaller model can use a skill from a larger model's run.
Works with your agent infra: use any framework, and bring traces from Braintrust, LangSmith or the Zep SDK.
❦
Stops
A little over half the time, Claude stops because it actually finishes what I ask it to do.
About 5% of the time it stops before it's finished and asks me to make some kind of decision. I consider these benign breaks in its work, where human direction is valuable.
A few decision points it gave me lately:
"Now that we've migrated to a new server, should I close the old one's open ports?"
"Should I top up your Gemini credit, or move that job to your Claude subscription?"
"Should the link go to the deep page, or to the overview page?"
"Do you want to have US and Canada on separate lines, or combined?"
Notice that decision points usually don't have a right or wrong answer. They're chances to exercise your direction and taste.
Avoiding decisions altogether in your work with AI is letting AI use you.
Interruptions
About 40% of the time Claude stops for the reason I dread: interruption.
I think of interruptions as stops where, under the right conditions, Claude could have kept going.
Interruptions are basically just friction. And I'm going to war on those.
How I've reduced interruptions from sign-ins and permissions
For sign-ins:
I take the time to create API keys for all my services and add them to Claude's environment, so it can use them fast and securely. Every morning, a check tests my logins and emails me if one is about to run out.
Lots of services don't have an API, or have a crappy one. So my Claude has its own browser—signed into every service—where it can navigate their UI. It keeps a notes file for each service, on how to navigate its UI best. Whenever it wastes time or tokens, it adds a note so it doesn't repeat the mistake.
For permissions:
I give Claude explicit approval in the project's context for specific actions where it would otherwise err on the side of caution and ask me. For important things, I ask it to run a test after to make sure everything works.
Putting my websites live. Working with subscriber data. Emailing reports to sponsors. Creating invoices in Stripe. Making entries in my accounting system. I don't need Claude to stop and ask me for this. I just need to make sure it does it well.
Obviously, think about how sensitive your work, accounts and data are. When I'm doing client work, I'm a lot more careful about which parts I give Claude access to.
How I've reduced interruptions from my unclear instructions
If you look closely at the first chart, you can see I've marked a period at the end of August when I worked on organising my context. I thought deeply about how I work with Claude, and got more deliberate about how I store and improve project-specific context.
What improved my "unclear instructions" was not actually me giving better instructions. It was Claude having access to better context.
Among other things, I created a system for how I work with chats. Each one now has a deliberate name, is colour-coded, reopens after I close Cursor, and is easy to find through a new search interface I built.
I now start most of my chats in existing threads rather than new ones, so my instructions arrive with that chat's context. That makes it easier for Claude to understand what I want.
(I'm planning a future issue explaining my context organising system in more detail — reply if you think that sounds interesting and I'll make it happen)
Why I'm doing this
I wasn't sure what to expect when asking Claude to analyse my chat data.
But it was super worthwhile.
My goal is to let AI carry out the specific tasks I give it, without interrupting me.
My thesis is that if I can minimise these interruptions, it'll be much easier for Claude to take on long tasks. I recently had great success sending Claude on overnight projects, like rebuilding my website and vibe coding my email service provider.
Part of the reason is that I have already done some of the legwork of making sure Claude can work friction-free.
And I now have a way to track if my work on removing the friction is actually working.
If you can measure it, you can improve it.
You can do it too
It's relatively easy to analyse your own AI usage. I'd recommend just pasting in screenshots of my charts and asking your favourite AI to do the same for you.
Your interruption mix is probably gonna look different than mine, and that's the point. Then you know what to improve on. If you do this, I'd love to hear about it.
Towards a smaller army of Claudes asking for our attention
My agent stops ≈2,000/month.
That's 24,000 requests per year for my attention.
I enjoy AI, but that's a commendable amount of distraction.
If I’m able to automate some of it away, my day will be more peaceful, with more time for deep work and slow coffee.
And that's what I enjoy the most.
What's your verdict on today's email?
Subscribe to participate
Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.
