What I noticed when I stopped running AI on my investments for two weeks

I came back to the desk this morning after two weeks away from running AI on my own positions. Seven a.m. UK, before the kettle was on. I’d expected the obvious things — that I’d missed an earnings release or two, that I’d be behind on names I’d been adding to, that the AI investing routine would feel rusty for a few sessions.

Some of that was true. Most of it wasn’t.

What I noticed instead was the shape of the routine itself. When you run something every morning for months, the routine becomes the lens — you stop seeing it. Two weeks without it gave me back the ability to see what running AI on my own positions was doing to my thinking. Where it was earning its place. Where it had quietly turned into ritual.

This post is the field notes from that gap. Three observations about the routines I came back to, then one observation I wasn’t expecting.

What I expected to miss

The honest list, written down before I left: overnight news on the names I hold; the post-earnings prompt for any position reporting while I was away; the thesis pulse check I run on positions I’ve added to recently. I expected the absence of all three to compound — to come back to a backlog of unread context I’d have to work through.

I’d structured the gap to be a proper one. No quick checks. No “just a glance” at the holdings on my phone. I’d asked a friend who watches the market to flag anything that broke materially against any of my positions — that was the only safety net. The rest was off.

Stepping away from a tool you depend on only works as a diagnostic if the gap is real. The reduced version I tried the first morning — a glance at the prices, “just to check” — told me nothing the routine wasn’t already telling me. Only the full pause did the work.

What I noticed about the AI investing routine

1. The routines that earned their place

The overnight news scan is one of them. Twice during the two weeks, my friend flagged something I’d have caught on the 08:00 prompt — once on AAPL (an analyst downgrade that landed during a US session I was sleeping through), once on a sector story that touched MSFT obliquely.

Neither needed acting on. But not seeing them was a real cost — the kind that’s easy to dismiss when you’re running the prompt every morning and the answers usually come back as ALL QUIET. The value of the prompt isn’t the days it produces signal. It’s the days you’d otherwise spend mildly anxious about whether you missed something.

The thesis pulse check is the other one that earned its place — but only on positions I’d added to recently. I’d added to META in late April after the Q1 earnings release. Not running the pulse check on that position weekly during the gap meant I came back to a position I hadn’t reread the thesis on in two weeks. Reading it back, the original case still held, but I’d lost a step of conviction in the interim. That’s a real cost on a position you’ve recently sized up on.

2. The routines that didn’t earn their place

The earnings calendar prompt, weirdly, is the one I missed least. I’d been running it every Monday as a wide scan across all my holdings — a kind of upstream filter to flag earnings releases sitting within the next 30 days. During the gap, my paper diary did the same job. Earnings dates change less than the prompt’s confidence level implied — and when they change, the diary entry is wrong, not the prompt.

The bull/bear prompt was the bigger surprise. I’d been running it weekly across most positions, on the principle that the framing was useful and the prompt was cheap. Two weeks without it didn’t change anything material on any position. The decisions a weekly bull/bear prompt informs aren’t weekly decisions — they’re monthly or quarterly ones. Running the prompt at weekly frequency produced outputs I was filing rather than acting on.

That’s not a small thing. Once you notice you’re filing outputs rather than acting on them, the prompt has stopped being a decision aid and become routine generation. Reducing the cadence from weekly to monthly probably gives me the same signal at a quarter of the noise.

3. The routines I’d quietly stopped doing well

This is the one I should have noticed without needing the gap to flag it. The Tuesday morning post-mortem I’d been running on the previous week’s decisions had drifted into a checkbox — open Claude, paste last week’s notes, ask for the standard review, file the output. The prompt was the same. My attention to the output had slipped.

The gap reset my attention. The first post-mortem I ran after coming back read sharper than any in the previous six weeks. The prompt hadn’t changed. I had.

The thing I wasn’t expecting

The absence improved my reading of the prompt outputs when I came back. That sounds like the kind of thing you’d say at the end of a holiday article. It’s also straightforwardly true, and the mechanism is worth naming.

Two weeks of not having Claude do the first pass on a thesis meant I read this morning’s bear case on NVDA with a fresher set of priors than I’d brought to anything in months. Same prompt I’d run a dozen times before the gap. The output landed harder. The novelty mattered.

Run any analytical tool every morning and the outputs blur. The signal-to-noise ratio drops, not because the model has changed, but because your read of the model has. You start scanning rather than reading. You start treating the output as confirmation of a prior rather than as data. After a gap, the same outputs read sharper.

I think there’s a real frequency past which the routines stop adding signal and start adding noise. The question I came back asking wasn’t “should I keep running this.” It was “should I run it less.”

What I’d change

Three changes I’m making on the routine, from a two-week gap I now think was useful work rather than an absence from it.

1. Bull/bear prompt: weekly → monthly. Run on the first Tuesday of the month, across all positions. I ran this weekly through the spring and the two-week diagnostic showed weekly was producing outputs I was filing without reading.
2. Earnings calendar prompt: weekly → on-demand. Keep the paper diary as the main source. Run the prompt when I add a position, not every Monday.
3. Tuesday post-mortem: rebuild the prompt. The output had drifted into a checkbox. The prompt needs the same treatment I’d give a stale routine on a position — a deliberate review of what it’s for.

The thesis pulse check on recently-added positions stays weekly. The overnight news scan stays daily. Both earned their place during the gap; their absence was a real cost.

The pre-gap version of the morning routine was the five-prompts-in-sequence pattern. The version that goes back into rotation this week is what’s described above — same prompts, lower cadence on three of them.

If you’ve taken a deliberate gap from a tool you depend on — analytical, AI, otherwise — and noticed something similar, tell me. The “what came back into focus” observations are the ones worth swapping notes on.