10 things my AI does that nobody sees

Aloha PLB Ohana,

My default AI is Opus 5.5, and lately I've been noticing it do things I didn't ask it to do.

The first time I noticed this was a couple of weeks ago, before a call that a local business owner booked with me.

A few days before the call, my AI noticed the call coming up, built her AI Visibility Report on its own, and notified me in Slack.

The AI Visibility Report is the document I prepare before every sales call, and it reports on how the AI engines see a business, what a visitor runs into, and what I'd fix first.

But here’s what I want you to pay attention to:

Even though I have a skill created for the report generation itself, I do not have a skill or any AI workflow set up that creates the report on a trigger (i.e., when a call is booked).

My AI noticed a pattern in my business (sales call booked > Cindy runs the report), and it proactively took action.

Even though I know that this is where AI is heading, I was still surprised when this happened.

I haven’t heard any of the YouTube creators I follow talk about this development, haven’t read any updates from Claude about this, and quite frankly didn’t think we’d arrived at this moment yet.

And that moment is that AI is becoming autonomous and agentic, which means it notices what needs doing and does it, instead of waiting for you to type a prompt.

So how did this happen?

I can only assume that my AI could do this because of what sits underneath it.

My AIOS (AI Operating System).

It knows what that report is for and when it's due, because my system holds my learnings, my judgment, my decisions and my boundaries.

The judgment part is what I call fidelity, meaning how closely the AI makes the call I would have made.

Without those four, an AI that acts on its own is mostly guessing faster.

In my last post, I wrote about the Disclosure Test. If your clients knew exactly how AI shows up in your work, would your offer still be worth what they pay?

It's only fair that I take the test myself.

So this is my full disclosure, in two parts. This post covers the system underneath, the part nobody sees. Part two covers what it builds.

But first, the part I don't hand over.

What I don't hand to AI#

My voice. AI drafts a lot of what I write. (Including the first draft of this post. Disclosure, remember?)

The rules it drafts come from me, though.

Everything that I publish that AI assists me with goes through a voice calibration, and it looks like this:

• AI drafts V1
• I edit (usually butcher) and hand back to AI
• AI takes note and adds notes to the voice calibration file. It notices patterns in my edits, the way I write, and the goal is that each V1 thereafter is a little closer to how Cindy actually writes.

My convictions. AI doesn't get to have opinions on my behalf. When I want to write about something I believe, it interviews me first and works only from my answers. If the interview doesn't go deep enough, we do it again. (I sent one back for a second round just last month.)

My judgment. AI does the research and the drafts, and then the decision comes back to me. It isn't allowed to email a client, change a price, or post anything under my name.

10 things AI does inside my business#

Below is a list of 10 things that AI does inside my business. Some things I’m sure you’ve heard other business owners talk about having AI do for them, and I hope some of the things I’m having it do are new ideas for you.

1. It briefs me every morning.

Every morning a business briefing lands in Slack. Mine is called a “Daily Pulse.” Before writing it, the AI reads my email, my calendar, my CRM, my meeting notes and Slack, then checks all of that against what my project notes say. Project notes are summaries of each Claude Code session, so anytime I work on something in Claude Code (for myself or for a client), that work is documented and progress is tracked.

When the two don't match, it flags the difference.

My Daily Pulse has saved me more times than I can count over the last 4 months or so, and I believe it is one of the key factors that has allowed me to essentially double the number of agency clients I can work with at any one time.

2. It keeps my client tracker honest.

Every morning, a separate AI reads client emails and meeting notes and updates where each client stands. It only writes down what it can point to, like a sent email or a reply.

Anything it's guessing at, it holds for me to confirm.

3. It preps me for the week ahead.

Every Sunday morning my AI looks at my calendar, my inbox, Slack, payments and every open project, then asks me three to six questions about the week ahead. Each question comes with its own recommendation, so "yes" is a complete answer.

Then on Sunday evening it turns my answers into a one-page plan, with the emails I need to send already drafted and an honest look at the money coming in.

4. It remembers every correction.

When I correct it, it writes the lesson into a memory it loads every time I start working, in any project. There are more than 180 of those lessons now, plus nearly 500 more tied to specific projects.

Here’s an example: Not too long ago I was noticing that my AI kept reminding me to do things I had already handled. The notifications that AI was sending me were getting annoying, so we troubleshot and did a system check of its process. Turns out, it wasn’t checking my sent email folder, so I corrected it and told it to check that folder before assigning me a task.

So now that is what it does, and I’m getting a lot fewer of these notifications from AI.

The AI does occasionally make mistakes, but the goal is that I only ever have to say something once.

5. It checks its own work before I see it.

This one came from frustration. This summer my AI kept handing me confident answers that turned out to be guesses, and I lost hours following them. So I added a rule to its instructions, in my own words:

“I do not mind things taking longer to do right. I DO MIND being led down the wrong path because AI assumed rather than checked.”

Now it has to point to the file, email or page behind a claim before it tells me something is true.

And every new article for my site's Playbook gets reviewed by a second AI that has never seen the draft. If the draft fails that review, it never reaches me.

6. It has to justify giving me homework.

My AI knows that I am actively building a system that I want to improve and get better over time. So whenever I start and end a session in Claude - it usually has a few “Suggested Improvements” for me.

But it’s kind of like a self-fulfilling prophecy…

Ask it to improve, and it’ll always find something to improve. And this was resulting in it giving me more and more tasks to do. Mine got so good at suggesting improvements that I told it this was getting out of hand.

So now, before it can put anything on my plate, it has to answer five questions I wrote:

• Is this necessary?
• Will it create more issues?
• Does it make the system better?
• Can you just do it yourself?
• Can we live without it?

If it can do the work itself, it does.

7. It keeps a record of every decision, including what I said no to.

Every meaningful decision goes into a log: what we chose, what we chose against, and why. It's about a hundred decisions long now.

When an old question comes back around (they do), the record is there, so we don't argue it all over again.

How this shows up daily: I recently rebranded the website (and business) with new fonts, new colors, and a new brand style guide. So now when I want it to design a PDF, create a graphic, or do something visual, it knows that we retired the old brand style guide and are using the new one.

Regardless of what we’re designing, it knows how to build it.

8. It takes action from my calls.

Probably the one integration that I couldn’t live without in my AIOS is its connection to my notetaker, Granola.

This is probably my ‘secret weapon’ if I had one.

Before a client call or a sales call, I usually do some sort of prep with the AI. And as soon as the call is done, I simply go into that session, tell the AI to process the most recent call, and tell it to take action. It may help me strategize next steps, prep the next thing, generate a proposal, etc. While it’s at it, it’ll log the call notes, update the client projects, generate content ideas, suggest improvements, etc.

Note takers have been around for a while now, but having mine connected to my AIOS has changed what happens after every single call.

9. It keeps my clients' files private from each other.

We build for several clients at once, and each client's files belong to them. So the AI is blocked from mentioning one client anywhere in another client's work. It's a hard block, which means it holds even on a day the AI forgets the rule.

I know that this is an obvious thing, but it’s one of those things that you probably will never think about when designing your AIOS until you notice it's an issue… so consider this a bit of a heads-up if you do client work.

Nuance like this is emotional intelligence… which AI has none of.

10. It runs a hygiene sweep every month.

There is an inherent amount of bloat that comes with building a self-improving AIOS. Decisions, judgments, learnings, issues, loops, etc. all stack and stack and stack.

CLAUDE.md, AGENTS.md and other files get increasingly bigger and bigger and bigger.

So every month I have my AI do a hygiene sweep - and its job in that session is to look for what to remove, simplify, fix or automate.

Instead of changing anything on its own, it hands me decisions I can approve with one word.

Where these 10 things came from#

One of the ironic aspects of AI models like Claude from Anthropic or Codex from OpenAI is that two of the most powerful technology tools in human history came with absolutely no instructions.

No onboarding guide.

No best practices.

No knowledge base.

I asked AI about this one, and its response was interesting…

It said that they were intended to be “primitive” out of the box. Nimble.

My best interpretation is that the engineers of the tools had no idea themselves what best practices would be.

It was our job as early adopters to figure this stuff out.

Like cave men and women discovering fire.

Most, if not all, of the list of ten above came from me experimenting with AI.

Making mistakes, rubbing two sticks together, and simply talking to the system and telling it what I wanted it to do.

Quick Win: Just tell it what you want#

When I work with clients and help them install their own AIOS, there’s always a bit of carried-over rigidity from how they’re used to learning new systems.

This is totally understandable and expected. Working with AI and with an AIOS is a fundamentally new way to work.

Here’s usually what happens with my clients…

Incoming Slack message: “Hey Cindy, my AIOS is doing this thing I don’t want it to be doing. I’d rather it do X. Should we update its instructions somewhere?”

My reply: “Sure. In your next session, tell it, ‘You’re doing this thing I don’t want you to be doing. I’d rather you do X.’”

I then usually follow that message up with a laughing face emoji 🤣 and then say something like this:

This is super cliché, but treat your AIOS as if it's a very eager new hire that wants to do everything right by you. You have to set boundaries, give it direction, tell it when it fails, and to a certain extent tell it when it does something great so it can do more of it.

Work with us#

So which of those ten would give you back the most hours this month?

That's the work we do with clients. We set up the version of this that fits your business, and we teach your team to run it.

If you'd like to see what that could look like for you, book a call. Before we talk, I'll put together an AI Visibility Report on your business, so the call starts with something useful whether or not we work together. (These days, my AI might beat me to it.)

Part two covers what AI builds: the client portal, the apps, the websites that won't publish copy above an eighth-grade reading level, the proposals that are their own web pages, the support desk I built in Manus, and the always-on agent on my Mac Studio that I can text from my phone.

Read part two: 10 things my AI builds that everyone sees

I'd love to know which of the ten surprised you most. Leave a comment below; every one comes to me first.

Mahalo for being here.

Cindy