AI in Practice: the September recording.

You can use graded loops in three progressively autonomous ways.

Running a loop by hand works anywhere. That's what the demo was: three prompts in an ordinary chat window. Do the thing, grade it against these criteria without fixing anything, now fix what failed. Any chat will do this, and it's where I'd start anyway.

Turning it into a reusable skill doesn't need Claude Code either. Skills run in the ordinary Claude app on web and desktop, from the free plan upwards, as long as code execution is switched on in your settings. A skill you enable there also loads in Claude Code under the same account, so you can package your loop once and then just ask for it by name.

The question mentioned projects, which are a different thing. A Project holds context and files for a body of work. A Skill is the reusable instruction set. You can use both together, but the loop lives in the skill.

Having it run without you is the level that needs more. In Claude that's Claude Code, which can put work on a schedule. ChatGPT has scheduled tasks, capped by plan at a handful of active ones. Gemini's agent features handle multi-step work but are still limited and marked experimental.

All of this changes fast, so check where things stand before you build around it. What doesn't change is the order: run it by hand first and watch the grader catch a real problem before you let anything run without you.

Only because of the demo. I already have skills built that run graded loops for me, so if I'd left them switched on, Claude would have done the whole thing properly behind the scenes and you'd have seen none of the steps.

I wanted the first pass to come out the way it normally would, mistakes and all, so I asked it to set those aside. That's not a recommendation. When you're doing this for real, keep your skills on.

No list. It knows my key areas of focus, and beyond that I simply asked it to find resources relevant to whatever I'm working on at the time.

I've been really impressed by how relevant they are. Often they come from sectors far removed from mine, which I find helpful in itself.

No. This doesn't need Claude Code and it doesn't need the operating system I've built. Sit down and chat with your AI and tell it you want to find ways to use graded loops to improve the reliability of what it gives you, then go back and forth from there.

I think this is where most people get stuck, assuming it's something only engineers do. If you can have a conversation like the one we had on the call, you can set this up.

Whether AI could score itself against it without guessing what you meant. "Make it sound professional" fails that test. Professional to who? "Every action item you quote has to be linked to a line in the transcript" passes, because it either is or it isn't.

The six I used in the demo are the things a good project manager checks at the end of any meeting. Is it grounded, so every claim traces back to something someone actually said. Is it owned. Does it have a date, and if not does it say so rather than inventing one. Is there a concrete next step.

Writing those down is the part that's yours. You're the expert in what you're trying to achieve, and the loop is only as good as the bar you set.

You tell it when to stop, and there are two ways. Either it stops when it has met the criteria you set, or it stops after a fixed number of passes. I usually say three, because by then it's clear whether it's improving or spinning its wheels and burning tokens.

When it stops without passing, it should tell you what failed and why.

Because you checked it yourself first. When I set one of these up, I validate every single result by hand for the first several runs before I turn it into a skill or put it on a routine.

I have my meeting transcripts running on a routine every morning now, and I do trust what comes out, but only because I spent the time at the start proving the criteria were right. The other framework I use is I do, we do, you do: do it together the first time, check it, and only then let it go on its own.

They're different things. If you connect with someone the default is that you follow them too, but you don't have to. You can stay connected to someone and turn following off, so you keep the connection without seeing everything they post.

You can also follow someone you're not connected to, whether that's a person or a company, and still see their content.

The bell is a third thing, and not many people know it's there. Subscribing to someone puts their posts in your notifications rather than leaving it to your feed to decide. Click the bell, choose "all", and you'll get a notification every time that person or organisation posts.

It depends on what you're there for, but Andrew's answer was to stop watching the vanity metrics. Impressions and interactions on an individual post tell you very little.

What matters is profile views and the messages that follow. He's had plenty of people reach out who never once liked or commented on anything he posted. They saw it, his name kept coming up, and they got in touch later.

Somebody asked whether there's a number to aim for, a hundred profile views a week or a thousand. There isn't one. It depends entirely on what you're there for and who you need to reach.

Less than we all used to think. Andrew's guidance was to post when it's comfortable for you, rather than trying to hit a window.

Being active at all is what makes the difference, so post when you'll actually do it.

Not for the reason we all learned to use them. Loading a post with hashtags used to be how you got found, and LinkedIn's systems now read your post semantically, so a post about global health is understood as being about global health whether or not you tag it.

The one remaining use is search. LinkedIn takes your first three hashtags into the post's URL, so if you specifically want a term in there, that's how. Otherwise Andrew finds them a bit tiresome to read, and I'd agree.

Because it's learning from what you look at, not just from what you click. LinkedIn tracks dwell time, which means how long you spend reading something even if you never like, comment or save it.

So if something political turns up and you stop to read it while disliking it, you've just taught the system you're interested. Use the "not interested" button, or keep scrolling. Don't rubberneck.

One more thing. LinkedIn is testing a second feed toggle, so some accounts now show "For You" next to "Network", which is closer to the way it used to work. Not everyone has it yet.

If you'd like to beta test Orpheus Social, which shows how your LinkedIn presence is likely being read, then hands you a plan to strengthen your signal, simply hop on orpheussocial.com and hit the "express interest" button and Andrew will send instructions, including a code to access the system free of charge (in exchange for feedback, of course). Feel free to connect with him on LinkedIn or send him an email directly, at andrew@ess3.ai.

I was one of the early testers. The version I ran last week picked up specific posts and comments I'd made and tied them back to how my profile reads.