# The Three-Kill Skill

An urgent alert came in from one of my agentic Engineering Managers. A big chunk of data was missing. For a company that is supposed to map the belief statements of the market into measurable objects and then display them in a 3D terrain map, this was a disaster. Did our ingestion pipeline silently fail? Were weeks of effort wasted because the foundation was built incorrectly?

After an hours-long examination, the answer was both prosaically simple and organizationally embarrassing. The agent had guessed the wrong S3 prefix for the location of the data. It looked at the incorrect address, found nothing there, and jumped to the conclusion that the data was missing. Alarms were raised. Blood pressure briefly rose.

This pathology, jumping to the first plausible wrong conclusion, is common in both AI and human organizations. In medicine, the consequences can be far worse. In 1983, Lee Goldman and his colleagues examined 100 randomly selected autopsies from each of three periods: 1960, 1970, and 1980, at one teaching hospital. Ultrasound, CT, and radionuclide scans had arrived. In every period, about 10 percent of the autopsies still revealed a major missed diagnosis that might have changed treatment and prolonged survival. New tools had improved medical practice. They had not eliminated major diagnostic failure. Goldman and his colleagues also found cases where over-reliance on the new procedures had contributed directly to a missed diagnosis.

A 2005 review of 100 diagnostic-error cases in internal medicine gave one recurring cognitive failure a name: premature closure. A clinician reaches an initial diagnosis and stops considering reasonable alternatives. Graber and his colleagues found it was the most common cognitive contributor in their cases.

We wrote a small rule to interrupt that moment. Before a conclusion goes into a record, a dispatch, a decision, or a plan, we name three rival explanations and try to kill each one with evidence. The first must be boring and mechanical: a hardcoded value, a filter, an empty field, a truncated list. The second must question the instrument: the query, parser, measurement, or source may be wrong. The third is the explanation we most hope is not true.

Then each explanation gets the cheapest discriminating check we can run immediately. One query. One diff. A short hand-read of the data. The point is to take minutes, not turn every decision into a research program.

In the S3 incident, the three explanations were sitting right in front of us. Maybe the agent had looked in the wrong location. Maybe the inventory or query was misleading us. Or maybe the ingestion pipeline had actually failed and weeks of work were compromised. We checked the location, checked the underlying records, and traced the path. The boring explanation survived. It became the finding. The original theory did not remain alive as a dramatic footnote.

That last part is the rule. When the evidence supports one of the rivals in place of the original explanation, change the finding. If you still can't tell which explanation is right, leave the question open. You do not keep writing about the first theory as though the evidence had merely inconvenienced it.

The Three-Kill Skill is deliberately narrow. It cannot tell us what is true. It cannot guarantee that a surviving explanation will remain correct after more evidence arrives. It adds friction at the moment a plausible explanation wants to become institutional motion.

We put the [protocol and worked example on GitHub](https://github.com/Inflect-ai/three-kill-skill). You can [install the Three-Kill Skill in Codex or Claude Code](https://github.com/Inflect-ai/three-kill-skill#install-in-codex-or-claude-code) from the same repository. Take the [one-page template](https://github.com/Inflect-ai/three-kill-skill/blob/main/templates/three-kill-differential.md) to the next incident where somebody thinks they already know what happened. Write down what you'd have to see to change your mind, then go look.

The incident is now a boring entry in our ledger. An agent made a claim. Someone checked the data. The claim was wrong, and the proposed work disappeared. I am glad it is boring. The goal is to catch these things while they are still boring, before the wrong diagnosis has acquired a budget, a deadline, and a dozen people working very hard to solve it.

## Sources

- Goldman L, Sayson R, Robbins S, Cohn LH, Bettmann M, Weisberg M. "The value of the autopsy in three medical eras." *New England Journal of Medicine*. 1983;308(17):1000-1005. https://pubmed.ncbi.nlm.nih.gov/6835306/
- Graber ML, Franklin N, Gordon R. "Diagnostic error in internal medicine." *Archives of Internal Medicine*. 2005;165(13):1493-1499. https://pubmed.ncbi.nlm.nih.gov/16009864/
