Using AI to Slow Down

One of the recurring themes Every morning when I plan my day is "slow down." Over the last few weeks, Claude has turned out to be a good orchestration layer for that. The usual pitch on AI is acceleration: the model can do almost anything you describe in plain English, so the only skill left is knowing what to ask. That may be true, but the bottleneck isn't only what to ask, it's when. The default move is to use AI the second I want to ship. The orchestration move, the one my skills are built around, is to use it before that moment: to record decisions, get a panel of experts to weigh in, and help with implementation.

I have around 40 custom skills built into Claude. Here's how they break down.

Recording decisions

The memory layer isn't actually Claude, it's Obsidian. I've written about my second brain setup before. Every meeting, decision, person, project, and idea lives in cross-linked markdown files I've built up over years. Claude reads from it and writes to it, but the durable record is mine.

The piece I've been trying to add since I read Annie Duke's Thinking in Bets is a real method for recording decisions. Her argument: the quality of a decision is in the inputs you considered and the process you used, not the outcome. A good decision can produce a bad result and a bad decision can produce a good one. If you only look at outcomes you'll update away from sound processes that got unlucky and toward sloppy ones that got lucky.

The implication is that you have to record the decision at the time you make it. What you knew, what you considered, what alternatives you weighed, what you chose, why. Then you can come back later and review the process separately from the result.

I've been bad at this for years. Obsidian solved the storage problem but not the discipline problem. The skill stack is starting to solve the discipline problem too: the want-elicitor and idea-validator force me to write down the inputs to a decision before I let myself make it, and the daily and weekly reviews give me a structured place to audit the process.

It's the institutional memory of a one-person operation. It's also slowly becoming an honest log of how I actually make calls.

Panel of experts

My most used skills exist to pressure-test decisions before I make them.

The idea-validator walks me through customer discovery on any new business concept I get excited about. The want-elicitor interviews me before I start a project, asking one sharp question at a time until it's confident about the real goal. Both force me to name the customer or the actual goal before I let myself sketch the build.

The pattern underneath: any time I catch myself about to score a decision from feel, I build a skill that requires me to answer specific questions instead. They're stop-losses against my own loss aversion and my own impulse to fall in love with a plan before I've pressure-tested it.

These aren't thinking tools, they're scaffolding around the parts of my judgment that may have blindsides.

The same pattern shows up for emotional state. My check-in skill runs a parts-work session when I notice I'm about to give up. My pick-me-up skill runs cognitive reframing when I'm spiraling. Both blend clinical modalities (Cognitive Behavioral Therapy, Internal Family Systems, Matt Mochary's emotional clearing framework) and treat them as design patterns for decision-quality tools.

Most people use AI to accelerate. I've put a row of pauses in front of my own action.

Implementation

The skills that actually produce output sound like a small team: chief of staff for the morning brief, researcher for my publication, publisher for the daily roundup, critic for security and code review.

Some publish to public surfaces without me in the loop: the daily roundup ships posts to my CMS, the SEO agent fixes indexing issues and resubmits pages, the Instagram kit generates the image and caption.

The through-line

Across all three, the purpose is the same. I'm using AI to externalize the parts of myself I've found to have blind spots and bias. The frame I find most useful comes from Daniel Kahneman's System 1 thinking and the mental models Charlie Munger laid out in his Psychology of Human Misjudgment.

The biases I keep tripping on: loss aversion. Doubt-avoidance, the rush to remove discomfort by making a call. Authority-misinfluence, the Milgram-style deference to people I should question. And the lollapalooza, several biases compounding in the same direction, which Munger considered his most important model.

What this looks like in practice for me: deferring when I should push back is authority-misinfluence. Jumping into a build before I know what I actually want is doubt-avoidance reaching for the comfort of action.

Each skill is a prosthetic for a specific failure mode. The system is a way of running myself that doesn't depend on me being in good shape on any given day.

That, more than acceleration, is the point. The bottleneck isn't in my head, it's that my head is variable. AI doesn't fix that, but it can sit at the chokepoints and route me when I'm not.

What's next

The biggest thing I want to add is a local-only stack. The skills I run now mostly go through cloud models, which works for everything I'm willing to send to a cloud model. But I have data I'm not willing to send: bloodwork, genome decode, Apple Watch exports, anything related to my health. The next build is running an open-source model locally on my laptop and pointing the existing skills at it for those workloads. A personal health dashboard is the first real test.

That gives me the same orchestration layer for things I won't put in the cloud, which closes the privacy gap I've been working around.