The Whole Pattern

In practice

The pattern has eight parts, one for each unit before this one. First, know what the tool is: a generator of fluent, derived text whose confidence is not evidence, and which confabulates as a matter of design. Second, name the risks, using NIST's twelve and its seven marks of trustworthiness, and scan each use before adopting it. Third, prompt deliberately: instruction, context, input and format, tested against criteria, and changed one thing at a time. Fourth, use safe patterns: use only what is given, invent nothing, ask instead of guessing, mark uncertainty, and never paste what you would not publish. Fifth, verify claim by claim, number by number, against the real source, with experts where the stakes need them, and record what you checked. Sixth, keep humans in the loop by design: a feature gate, a clear line between what AI suggests and what the system decides, review matched to the stakes, and a log of overrides. Seventh, govern the tools: governed services for sensitive work, vendors checked like any other supplier, and short team rules people actually read. Eighth, document the AI and the work it touches, with model cards read critically, data questioned, and reference packs owned, versioned and kept current. Through all eight runs the method's working line: draft, check, decide.