Purpose
Use this lesson to prevent an answer from crossing business, legal, or scope limits. The aim is to produce work that a colleague can review and act on, not merely text that sounds convincing.
Lesson Explanation
Scenario: Write a public reply to an unhappy customer.
Better instruction: Write an empathetic reply; do not promise a refund, admit legal fault, or mention internal investigation details.
Constraints are positive requirements and explicit exclusions. They are especially valuable when the safest answer is narrower than the most helpful-sounding answer. State non-goals plainly rather than hoping the model infers them. The reply is helpful without making an unauthorized promise or disclosure.
Work in a short cycle: provide the relevant input, state the result and limits, inspect the draft against the acceptance criteria, then revise the instruction when a requirement is missing. Keep the final human decision with the person responsible for the work.
Practice Questions
-
A prompt asks an AI to “write a persuasive pitch” for a product with known safety limitations, without stating what claims are off-limits. What does this lesson say is missing?
View Answer
Explicit constraints and non-goals – the prompt should state what the AI must not claim or promise.
-
What does this lesson say constraints actually are?
View Answer
Positive requirements and explicit exclusions.
-
A request for a public statement about a data breach doesn’t specify what details must be excluded. What risk does this create, based on this lesson?
View Answer
The AI might include sensitive, legally risky, or unauthorized details, since exclusions weren’t explicitly stated.
-
Rewrite a bare instruction “respond to this upset customer” into one that includes explicit constraints, inventing reasonable ones.
View Answer
Something like: “Respond to this upset customer with empathy; do not offer a refund, do not admit fault, and do not mention any ongoing internal review.”
-
Why does this lesson say constraints are especially valuable “when the safest answer is narrower than the most helpful-sounding answer”?
View Answer
Because an AI might otherwise produce a broadly helpful-sounding response that inadvertently oversteps into territory (like promises or admissions) that isn’t actually safe or authorized.
-
A prompt states what the AI should do but never states what it must avoid saying. What does this lesson recommend?
View Answer
Stating non-goals plainly, rather than hoping the model infers them.
-
A legal team wants an AI-drafted public statement that avoids admitting fault, promising compensation, or revealing details of an ongoing investigation. How would this lesson describe these three requirements?
View Answer
Constraints – specifically, explicit exclusions (non-goals) alongside whatever the statement is positively required to do.
-
A request produces a response that is empathetic and well-written, but accidentally implies a refund is guaranteed. What earlier step in the prompt likely could have prevented this, per this lesson?
View Answer
Explicitly stating “do not promise a refund” as a constraint/non-goal, rather than hoping the model would infer this boundary.
-
What is the difference between a “positive requirement” and an “explicit exclusion,” as this lesson uses these terms for constraints?
View Answer
A positive requirement states what must be included or done; an explicit exclusion (non-goal) states what must specifically be avoided.
-
A prompt for an internal announcement about layoffs doesn’t specify that the AI should avoid speculating about future rounds. What does this lesson suggest adding?
View Answer
An explicit non-goal/constraint stating that the AI should not speculate about future layoffs.
-
Why does this lesson caution against just “hoping the model infers” non-goals?
View Answer
Because an inferred boundary isn’t reliable – the model may not recognize an unstated boundary the way a human would, so it should be stated plainly.
-
A prompt for a product FAQ states the task, audience, and format, but no constraints at all. Based on this lesson, what risk remains even with those three elements defined?
View Answer
The AI could still include content that oversteps appropriate boundaries (like unverified claims or off-limits topics), since nothing was explicitly excluded.