Prompt-and-pray

An AI workflow with no safety net: someone writes a prompt, runs it, and hopes the output is correct. No retrieval system, no test, no fallback, no logging.

Prompt-and-pray is an AI workflow with no safety net: someone writes a prompt, runs it, and hopes the output is correct, with no retrieval system, no test, no fallback, and no logging. The fragility shows up in production: an edge case arrives, the model hallucinates, the output format breaks, the source data shifts, and nobody can diagnose what went wrong because nothing was logged or tested. A team builds a customer support bot by writing one long prompt, demoing it on ten friendly questions, and shipping it. The first ambiguous, angry, or out-of-scope message from a real customer reveals that the entire system was a single prompt with no structure underneath.

Builder example

Many AI features fail because the first impressive prompt becomes the permanent architecture. A bare prompt can answer an early feasibility question, but a workflow that affects customers, money, permissions, or stored records needs controls matched to its failure modes. Those controls may include source retrieval, schema validation, evaluations, logging, fallbacks, or approval before an irreversible action. Choose the protections from the consequences rather than attaching every safeguard to every prototype.

Common confusion: A long, detailed prompt can feel robust because of its length. Length adds specificity; it does not add reliability. A 2,000-word prompt with no tests, no retrieval, and no fallback still fails unpredictably on inputs the author did not anticipate.