A useful prompt gives the model a job, evidence and limits. “Write a sustainability report” is not enough. You will get fluent generalities and probably a few claims that need checking.

Try this instead: “You are assisting with a preliminary review of a municipal waste project. Use only the attached feasibility study and the cited public permit conditions. Extract the claimed annual waste throughput, energy output and emissions limits. For each figure give the document name, page and unit. Mark anything you cannot verify as NOT FOUND. Do not invent missing data.”

That prompt defines the task, sources, output and what to do when evidence is absent. Add a table with columns for claim, source, date, unit and verification status. Ask for questions the consultant should resolve before issuing a report.

Now test it. Insert one deliberately inconsistent figure into a synthetic example and see whether the model catches it. If it misses the problem, revise the process rather than pretending a more forceful prompt will guarantee accuracy.

Translation needs the same discipline. An ASEAN project may involve Bahasa Indonesia, Thai and English source material. Ask for the original wording beside the translation and have a competent reader check technical and legal terms.

Planetary.Blue shares practical prompt examples with members. Paid consulting is available where a firm needs a tested workflow.

Test the example on a small, non-confidential task. Planetary.Blue shares member use cases and offers paid consulting support.