A reusable prompt needs a clear job, a defined input boundary, and evidence that its outputs satisfy the intended task. Review the template with its variable fields, model assumptions, source material, and scoring rules. Preserve revisions so users understand which version was tested and why it changed. This framework treats a prompt as maintained working material: something to document, compare, and revisit when its environment changes. It helps identify what can be reused confidently and which parts depend on untested assumptions.
What to examine
Separate stable instructions from the fields a user supplies for each run. Define required information, acceptable formats, and the response to missing or conflicting inputs. Identify which source material supports the task and keep its contents distinct from instructions that authorize actions or change the template's operating purpose.
Specify the requested structure, required evidence, and conditions for an acceptable answer. Build tests for routine cases and predictable exceptions, including incomplete inputs. Record legitimate output variations so the review evaluates task completion and factual support without rewarding conformity to one preferred sentence or stylistic pattern.
Give the template a version and preserve its text, compatible configuration, test set, and change reason. Identify the person responsible for updates and the conditions requiring a new evaluation. Keep older versions available for comparison so a wording change can be assessed against observed results.
A practical review sequence
Write the task, expected user, required inputs, output structure, and permitted actions. Describe how uncertainty and missing evidence should appear. Make the boundary specific enough that another reviewer can tell when the template has completed its job or requires a question before continuing.
Collect authorized inputs representing ordinary work, difficult cases, and failures the template should handle. Attach expected facts or a scoring checklist to each example. Reserve cases for later evaluation so revisions can be checked beyond the examples that directly guided the wording changes.
Preserve the previous text before making a revision. State the problem the change should resolve, then compare outputs using the same relevant cases and acceptance criteria. Keep new errors visible, and explain whether an apparent improvement depends on additional context or manual preparation.
Provide the prompt version, input instructions, tested configuration, known limitations, and evaluation summary. Describe when users should stop or seek review. Set maintenance triggers for changes in the model, data, task, or tools so the template's evidence remains attached to its intended environment.
Primary-source context: Artificial Intelligence Risk Management Framework (AI RMF 1.0). Use it alongside the asset-specific records relevant to your review.
Questions to resolve
Include its purpose, required inputs, expected output, tested configuration, version, and known limitations. Add representative examples and the rule used to judge them. This gives another user enough context to reproduce the intended procedure and recognize when a request falls outside the tested scope.
Treat each target configuration as a case for evaluation. Preserve the same task requirements and compare outputs on representative inputs, noting any changes needed. Document the tested combinations rather than assuming that an unchanged template carries the same behavior or operating effort into every environment.
Create a traceable revision whenever a change may affect task interpretation, required inputs, output expectations, or permitted actions. Record the reason and relevant comparison results. Version history should help a reviewer connect behavior to the exact text used, including changes introduced to resolve previous failures.
Read the Lab
A persuasive demo is the beginning of an AI review. Define the task, build representative tests, compare outputs without brand labels, track failures and costs, and preserve enough detail to understand what changed between evaluations.
Read the guide: A Practical Evaluation Workflow for AI, LLMs, and PromptsRelated screening lenses
Questions worth asking
Explore a new asset class or suggest a topic for the Lab.