A new academic tutorial argues most prompt engineering is guesswork, and proposes a checklist to fix that.
The paper, posted to arXiv, treats prompt engineering as the work of turning vague human requests into structured specifications an AI system can actually execute. It lays out reusable steps: define the task precisely, include only the context the answer depends on, and assign a role, or a moderated panel of roles, as a lens for the model's attention. Rather than listing what not to do, it pushes writers to state affirmative quality targets and save prohibitions for genuine hard limits. It borrows Occam's razor and Chekhov's gun as editing rules, trimming any instruction that doesn't earn its place, and for riskier tasks adds steelmanning, premortems, and verification, plus an agentic operating loop with built-in escalation points for tasks involving tools or multiple steps.
Most prompt-engineering advice online is a grab-bag of hacks: magic phrases, role-play tricks, all-caps instructions. This tutorial is closer to requirements engineering, a discipline borrowed from software teams and applied to natural-language specs instead of code. That distinction matters because as agentic tools take on more multi-step work, the bottleneck stops being model capability and starts being how precisely a person can specify what good output looks like.
Worth noting: the paper is explicitly not a benchmarking study, so its checklists are reasoned opinion, not measured improvement, until someone actually runs the numbers.