The Elephant and the Goldfish

In September 1999, NASA lost the Mars Climate Orbiter. One file gave its figures in pound-force seconds; the specification said newton-seconds. The software did exactly what it was given. Nobody checked what it was given against the page.

Most people now work with AI in the same way. The tools have become very good at doing what they are told, so the quality of the answer depends on what they are told. This issue is for the majority of users who are not AI specialists but want better results. It sets out why clever prompts matter less than they used to, why the person writing the request is the worst judge of what is missing, and a simple method, adapted from an approach published by a Google engineer, for testing a brief before anyone relies on it.

• Why you are the worst judge of your own brief: people tapping out a tune expected listeners to name half of them; listeners got 2.5% (Stanford, 1990).
• Think with the elephant, test with the goldfish: one long chat to work the job through, then a fresh chat that sees only the written brief.
• Six habits that need no technical team, and a one-page card your team can use this month.