ChatGPT Prompts for Commercial Real Estate That Produce Reviewable Work

A prompt library meets the first two controls of the production standard the moment a team stops prompting from a blank chat. The work order described above is a named job in miniature: the role line defines the seat, the objective defines the finished output, and the source line draws the boundary. When the instruction says to use only the uploaded files and to cite a page, tab, or cell for every material number, the source boundary is enforced inside the prompt itself, not in a policy nobody opens. The test of a named job is whether a second analyst could run the same work order against the same packet and grade the result. A stored prompt passes that test; a clever one-line request never will.

The five standardized patterns each carry a layer of the output discipline. Document inventory and structured extraction produce the sourced facts, every field tied to a reference a reviewer can open. Reconciliation produces the deterministic layer: differences between two sources quantified, not asserted, with the likely resolving document named. Red-team review surfaces the working assumptions, listing unsupported claims and ranking unanswered questions by decision impact. The executive draft carries the recommendation, written from approved facts and model outputs with fact, calculation, assumption, and judgment kept visibly apart. A team that runs the patterns in that sequence has produced a fully layered output without asking anyone to memorize a taxonomy.

Stop conditions are the library's exception path, and they belong in the prompt before any pattern runs at volume. Conflicting unit counts, missing lease dates, unclear actual-versus-budget periods, a source the model cannot open, a calculation that cannot be reproduced from the supplied data: each should end the run with a specific question rather than a polished guess. The escalation is cheap because it is small. The operator resolves the one item that unlocks the work and reruns the pattern, which is far safer than finding the guess three reviews later.

The approval record and the measurement loop live in the same place as the prompt. Store each pattern beside its source packet, output template, and review checklist, and track the input packet, version, output, reviewer corrections, and failure reason on every run. When the same correction appears more than once, the instruction changes, and the record of that change stays with the pattern. That is decision-quality measurement at the scale of a single instruction, and it is the difference between a folder of clever sentences and a system the next analyst can inherit. The full statement of this standard is published separately on the PSV newsroom as “The PSV Production Standard: How AI Work Ships in Commercial Real Estate Operations.”

None of this authorizes the model to act. The boundary holds where this guide drew it: no external communication and no financial decision moves without a person reviewing the final output. The patterns prepare, reconcile, and challenge; the operator signs. Every improvement sharpens the preparation without moving the judgment, which is why the library survives repeated use instead of decaying into private habit.