Give AI a well-defined assignment
How to ask for useful work without inviting unsupported conclusions. Build a source-bound, constrained assignment.
- Convert a vague request into a structured assignment
- Define source material, scope, output structure, and evidence expectations
- Separate confirmed facts, interpretations, information gaps, and follow-up questions
- Specify how the output will be validated
A language model is eager to please. If you ask it to "review this borrower," it will invent a review. It will confidently guess what you want, pull in external assumptions, and present a polished but potentially unverified narrative.
To get useful work, you must define the assignment exactly as you would for a junior analyst: establish boundaries, require evidence, and state how the work will be checked.
A better prompt defines the assignment
A better prompt defines the assignment. It does not eliminate the need to verify the work.
A good prompt is not a magic spell that forces the model to be accurate. "Do not hallucinate" is an instruction, not a guarantee. The goal of a structured prompt is to constrain the model's output so that its work is easier for you to verify.
The ASSIGN Framework
Use the ASSIGN structure to build a constrained request:
- A — Acting role: What perspective and professional context should guide the draft? (e.g., "Act as a credit-review writing assistant.")
- S — Specific objective: What single task should be completed? (e.g., "Identify material year-over-year financial changes.")
- S — Sources and scope: Which documents, periods, entities, and sections are in bounds? (e.g., "Use only the FY2024 and FY2025 statements and the interim package.")
- I — Instructions and constraints: What must the system do or avoid? (e.g., "Do not invent causes or calculate metrics when inputs are missing.")
- G — Grounding and evidence: How must each material statement trace to source? (e.g., "Cite period, line item, and source page for every figure.")
- N — Needed output and next review: What format should be returned, and how will the reviewer validate it? (e.g., "Return a findings table followed by missing information and reviewer checks.")
Note: The "Acting role" establishes style and task context. It does not transform the tool into an authorized reviewer, lawyer, appraiser, or independent QA function.
Separating Fact from Inference
A common failure mode is the machine blending a verified fact with a plausible but unconfirmed explanation. For example, it might state: "Gross margin declined 180 bps due to supply chain disruptions." The margin decline is a fact in the spread; the cause might just be a plausible guess.
To prevent this, instruct the machine to separate its output:
- Verified facts: Numbers and statements found directly in the text.
- Potential implications: Reasonable inferences that require human judgment to confirm.
- Information gaps: Missing documents or missing data required to complete the analysis.
- Follow-up questions: Items the reviewer should investigate.