ChatGPT Prompts for Commercial Real Estate That Produce Reviewable Work

Most weak CRE output starts with a one-line request and no deal context. The model does not know the firm's definitions, the approved sources, the asset type, the relevant time period, or what the reviewer considers acceptable. It fills the gaps with generic language.

A production prompt should read like an analyst work order. Define the seat the model is supporting, the job to complete, the source files it may use, the facts it must not assume, the format it must return, and the conditions that require a question or escalation.

• Role: 'Support a U.S. multifamily acquisitions associate preparing a first-pass screen.'
• Objective: 'Normalize the supplied T-12 and reconcile it to the rent roll.'
• Source boundary: 'Use only the uploaded files; cite a page, tab, or cell for every material number.'
• Rules: 'Preserve reported values, separate adjustments, and flag ambiguity instead of guessing.'
• Output: 'Return the normalized table, reconciliation, open questions, and confidence by field.'

The exact language matters less than the operating pattern. A good team stores prompts alongside the source packet, output template, and review checklist so the work can be repeated and improved.

Each pattern below should be adapted to the firm's data policy and approval requirements. None should authorize external communication or a financial decision without a person reviewing the final output.

• Document inventory: list each file, date, period, and the decision-relevant information it contains.
• Structured extraction: map unstructured fields into a defined schema and attach a source reference to each field.
• Reconciliation: compare two sources, quantify differences, and identify the likely document needed to resolve them.
• Red-team review: challenge the draft thesis, list unsupported claims, and rank unanswered questions by decision impact.
• Executive draft: write from approved facts and model outputs, clearly separating fact, calculation, assumption, and judgment.

A prompt should say when the model must stop. Examples include conflicting unit counts, missing lease dates, unclear actual-versus-budget periods, a source it cannot open, or a calculation that cannot be reproduced from the supplied data.

Stop conditions improve both safety and speed. They prevent polished guesses and direct the operator to the smallest question that unlocks the work. The model becomes more useful when it is allowed to say that the record is incomplete.

A prompt becomes valuable when it survives repeated use. Track the input packet, version, output, reviewer corrections, and failure reason. Update the instructions when the same correction appears more than once.

That feedback loop is the beginning of an AI operating system: a controlled workflow that improves from actual review, rather than a folder of clever sentences disconnected from the work.