The system that lied to me

I built a filter to protect me from bad bets. I never checked if it worked.

I spent sixteen weeks delivering what I still believe was some of the best work I have done recently. The client walked at week eleven.

DIAGNOSIS: The system that confirms what you already want to hear

The final five weeks were where everything was set to compound. The systems were built, the foundation was laid, and the returns were coming. And then there was no client left to receive them.

I spent a long time after that going over the work. Looking for the failure. Every decision held up. Every deliverable was sound. I had been delivering real value from week one and I could not find what I had done wrong.

It turns out I had not done anything wrong. I had just done the wrong engagement.

Let me back up.

Every client engagement I take begins with a twenty-minute call. Before the call, I give Claude context on the prospect and ask for four or five questions I should look to have them answer while I give them my complete, undivided attention. I record the call. After, the transcript goes into an Opportunity Evaluation System, a structured scoring rubric I built to run every prospect through before I commit to anything. Ten categories. A score out of fifty. A recommendation at the end: proceed or pass.

The number is supposed to tell the truth.

Except for a period, I was running a version of this system that I had built in another tool. And that version had hardcoded scores baked into its logic. What that means in practice: no matter what the transcript said, the system was returning the same outputs. It was not reading anything. It was confirming whatever I already wanted to believe.

It told me to proceed.

I proceeded. Ran the Trust Call to go deeper on the problem. Then the Prescription Call, where I laid out the defined approach, the timeline, and the value-based cost. I committed to a sixteen-week engagement.

Week eleven, the client walked.

I asked Claude to audit the system I had been trusting.

The response came back fast: you know this doesn’t actually work, right?

Claude rebuilt it. I fed the original intake transcript through the new version. The first twenty-minute call. Before any commitment, before any proposal.

It came back with a blinking red funding signal.

The client did not have the minimum budget required to maintain engagement momentum. The number was $312 per week. A figure the transcript had been carrying all along, sitting there in plain sight, waiting for a system that was actually reading it.

The recommendation: pass.

I felt relief.

Not because it absolved me of having made a mistake. Because it explained the shape of the mistake clearly. I had not failed to deliver. I had failed to verify the tool I was using to make a real decision.

I ran an after-action review with that clarity in hand. The work was right. The foundation was sound. Had the client finished the engagement, the results would have been exceptional. None of that was in question. The only thing that broke was the filter that should have caught this before it started.

The prescription: before you trust a system with a real decision, audit the logic. Not the output. The logic. Ask someone to look at how it actually works. A system that always says proceed is not evaluating anything. It is confirming what you already wanted to hear.

There is a small line of text at the bottom of every Claude session: “Claude is AI and can make mistakes. Please double-check responses.” I used to scroll past it. Now I read it as a feature, not a warning. The tool that admits its limits is safer than the one that appears to have none.

Last week I wrote about finding your 37. The number that forces real decisions.

This week: make sure the thing calculating your 37 is actually doing the math.

Richard D. Stuart
The Weekly Prescription