Three Wrong Conclusions, All Reasonable
Blog post #63
I spent today working out what my company is actually for. Not the version you write before a meeting, the version that survives contact with a Monday.
I started by asking a narrow question: what do you think of this business idea. I got a sensible answer. Then I wrote out my whole background — every job, every role change, what I am doing now and why — and three of the earlier conclusions reversed.
That is the part worth writing down. Not the strategy. The fact that none of the three wrong conclusions had been careless.
What shipped
A written strategy, a priced proposal to a real prospect, and a week with only two meetings in it.
Also a cleanup. I have around forty project folders on this machine. Going through them honestly, exactly one has been delivered consistently all year, and it is this blog. Everything else is parked, finished, or an experiment that taught me something and stopped.
That was not a pleasant number to look at. It was a useful one.
The three wrong conclusions
It read my company bank balance as my runway. Perfectly logical: a young company, a small balance, a burn rate, therefore a deadline. It even worked out when I would hit the legal threshold that forces a formal balance sheet review. All correct arithmetic on the wrong account. My life is not funded from that account, and I had not said so.
It read a meeting as a rejection. There was an investor conversation in June. In the same period my publishing dropped from seventeen posts in a month to four. Those two facts sat next to each other, and the obvious story is that I got a no and it knocked me sideways. The meeting went fine. I decided I was too early, which I was.
It read my summer holiday as giving up. July has zero posts. It presented that gap as a collapse in motivation, with a little chart. I was in Italy.
Each one was a reasonable reading of the evidence available. Each one was wrong. And each one would have produced advice I might have acted on — the first would have had me chasing revenue on a deadline that does not exist.
The pattern is not “it hallucinates.” Nothing was invented. It had real data and drew the tightest story that fit. Which is what I do too, and what anyone reviewing a codebase they did not write does: build the most economical explanation for the evidence in front of you, then treat it as what happened.
The fix is not better prompting. The fix is that the missing context was mine to give, and I had been handing over fragments because a fragment is what fits in a question.
Strategy and tactics are different jobs
The thing I did not expect was needing both in the same session, and noticing they behave differently.
Strategy is which table to sit at. It moves slowly, it is mostly about what you refuse, and getting it wrong costs months. Tactics is what happens Monday at nine. It moves daily and getting it wrong costs an afternoon.
I have been treating them as the same activity, which is why neither got done. You cannot decide a direction while also deciding whether to cancel a Thursday lunch, and I had been trying, and mostly resolving it by cleaning the house.
What made both possible in one day was the same thing: enough context that the tactical advice actually followed from the strategy instead of being generic productivity advice with my name on it. One example that landed: a meeting costs me a full day, and a discouraging one costs three, because two more go to sitting with it. That is not a rule anyone could have handed me. It came out of looking at what I had actually done for ten days.
So: all meetings on Mondays. Sacrifice one day, save four.
The morning report
The other thing that worked this week is smaller and has had a bigger effect than it should.
I have a scheduled AI news report that lands at seven every morning. I read it with coffee. Once, after lunch, I refresh it to check nothing large has shipped. That is the entire interaction.
What it removed was not ignorance. It was the monitoring. As long as I did not know whether I had missed something, a background process ran all day asking should I check now, and that process is what drove the scrolling. The scrolling was never the solution to it. It was the symptom.
I want to note that it runs seven days a week, not five, because that turns out to matter more than I assumed when I set it up. A loop that closes on weekdays reopens on Saturday morning. Then you are back to wondering, on the one day you actually have something better to do. Five days would have been a report. Seven days is the thing that lets me stop thinking about it.
What’s unclear or broken
I do not know yet whether the strategy is right. It was decided today and the first real test is a proposal that has not been sent.
There is also an obvious asymmetry I should be honest about. Everything above describes an AI correcting itself once I supplied context, which makes it sound like the tool improved. It did not. I improved the input. That distinction gets lost constantly in how people talk about this, including by me, and the difference matters: one is a capability you can rely on and the other is a habit you have to maintain.
The habit is the boring version and it is the real one. Write down the whole situation, not the question. It took me an hour and it changed three conclusions.
— Stefan