Underspecified vs Constraint-Rich Prompts: The Ambiguity Test

Beacon Pharmacy fires off about 9,000 medication reminder texts a week across 41 stores, and a single bad batch can trigger refill spikes that floor staff cannot absorb. So when operations manager Nadia Okonkwo asked ChatGPT to “write a medication reminder SMS”, she expected one tight line of copy. What came back was a 118-word brochure that told patients to download the app, praised the pharmacy’s pharmacists, and finished with three exclamation marks.

Nothing was wrong with the model. The request was simply underspecified: no length, no audience, no tone, no legal limits, no example. This article puts the prompt Nadia typed next to the prompt that finally worked, explains why the second one holds its shape, and hands you a skeleton you can reuse for any request that keeps returning the wrong shape.

Why One Request Returns Two Different Answers

When a prompt has no boundaries, the model fills the empty space with whatever is most common in its training data. For “write a medication reminder SMS”, the most common thing is marketing copy. So you get feature lists, benefits, and calls to action, because that is what most SMS writing samples look like.

The failure is predictable once you name it: an underspecified prompt asks the model to guess your intent, and the guess will always be the average of everything the model has seen. If your need is specific, the average is wrong. Beacon did not need average SMS copy. It needed a 160-character reminder that named the drug, the dose window, and the store phone number, with no promotional language and no medical advice.

  • The symptom: the output is longer, warmer, and more salesy than anything you would ever send.
  • The cause: no stated length, audience, purpose, or forbidden content.
  • The fix: convert each silent assumption into one explicit constraint.

The Prompt That Produced the Brochure

Here is, almost verbatim, what Nadia typed. Note how short it is and how many decisions it leaves open.

Write a medication reminder SMS for our pharmacy patients.

What came back: a paragraph that opened with “Your health is our priority!”, listed the benefits of staying on schedule, invited the patient to book a consultation, and signed off with the brand name. It was 118 words. It read like an advertisement for the pharmacy rather than a nudge to take a pill.

Every editable field was left to chance: who the text is for, how long it should be, what it must contain, what it must never contain, and what a good one even looks like. A model cannot respect rules it was never given.

Reading the Constraint List Against the Output

Nadia rebuilt the request by naming the decisions she had been making silently in her head. The table below is the exact list she used. The middle column is what the vague prompt actually delivered; the right column is what she would have accepted.

FieldVague prompt resultRequired constraint
Length118 wordsUnder 160 characters including the brand name
AudienceGeneral publicExisting patient who opted in to refill reminders
Must containNothing specificDrug name, dose window, store phone
Must avoidNo ruleNo medical advice, no promotions, no exclamation marks
VoiceCheerful marketingCalm and factual, second person
ExampleNoneOne reference line to match

That is the whole test. For each row where the vague prompt left a blank, the model invented something. Fill the blanks and the inventing stops.

If you find yourself juggling length against completeness, the trade-offs are the same ones we break down in our prompt length comparison.

Rewriting Beacon’s Prompt With Hard Limits

Same tool, same task, six lines longer. This version adds the constraints as labeled sections so nothing is buried inside a sentence.

You are writing a medication reminder text message for a regional pharmacy
chain. The recipient is an existing patient who opted in to refill reminders.

Write exactly one SMS, under 160 characters including the pharmacy name.
It must include: the medication name, the dose time window, and the store
phone number as plain digits.

Do not include: medical advice, dosage changes, promotional language, emoji,
links, or exclamation marks. Do not greet the patient by first name.

Voice: calm, factual, second person. No adjectives about the pharmacy.

Reference example of the target shape:
"Beacon Pharmacy: your lisinopril refill is ready by 6pm today. Questions?
Call 5550142. Reply STOP to opt out."

Return only the message text. Then, on a new line, state the character count.

What came back: one line, 152 characters, naming the drug window and the store number, followed by the character count. It passed compliance review on the first pass. Nadia then generated variants by swapping only the medication and the time window, which the constraint list kept stable.

Why Constraints Beat Adjectives

People instinctively fix bad output by adding adjectives: make it shorter, make it less salesy, make it more professional. Adjectives are graded and relative. “Shorter” could mean 100 words or 40. “Less salesy” has no unit of measurement, so the model guesses again, and you are back in the same loop.

Constraints with units are checkable. Under 160 characters is checkable. No exclamation marks is checkable. Must include the store phone number is checkable. When every requirement can be verified with a glance, the model has no room to substitute the average for your intent, and you can reject bad output in one second instead of negotiating over six rounds.

A useful habit: after any disappointing reply, list the three things you judged it against. Those three things were your real constraints; you simply had not written them down before the model replied.

Watch for one more habit that keeps prompts vague: describing the deliverable by its name instead of its parts. “Write a summary” assumes the model shares your definition of a summary. “Write 120 words of plain-language summary with no jargon, aimed at a new patient” does not. The moment you spell out the parts, the ambiguity test passes itself.

The same test runs in reverse when output is too short. A two-line reply where you wanted a report usually means the format row was missing or contradicted another constraint. Re-read the prompt and ask which single line the model obeyed; that line is the one to sharpen, and the others are the ones to move up.

The Ambiguity Test Skeleton You Can Reuse

Run any new prompt through this template before you send it. Fill each bracket, or delete the line on purpose rather than by accident.

Role: You are [specific job].
Task: Produce [one named deliverable].
Audience: The reader is [exact person], who already knows [context].
Must include: [list of required elements].
Must avoid: [list of forbidden elements and tones].
Format: [structure, length or character count, and unit].
Voice: [three words, no vague adjectives].
Reference: Match this example of the right shape: [one example].
Output: Return only [the deliverable], then [any verification line].

Copy the skeleton, then fill the two rows people skip most: Must avoid and Reference. The avoid list removes the average junk, and the reference example pins the shape so you stop re-explaining it every time. For prompts that drive a formatted system output, the same discipline is covered in our output format guide.

Finally, keep the winning prompt, not the winning answer. Beacon saves the constraint block as a reusable template, so the next campaign only changes the two variables that should change. That is the real payoff of the ambiguity test: it turns a lucky one-off reply into a repeatable one.

Sources & Notes

The constraint checklist and the ambiguity-test skeleton are our own framing, reviewed on 2026-10-12. Beacon Pharmacy, Nadia Okonkwo, the 41 stores, and the 9,000 weekly texts are illustrative examples. The character limit reflects a common SMS segment of 160 characters and is a design choice, not a legal standard. Compliance rules for patient messaging vary by region, so confirm any reminder template with your own counsel before sending. Language models change between versions, so re-check any prompt that depends on an exact character count or formatting rule.

声明:1.本站所有文章,如无特殊说明或标注,均为本站原创发布。本网站中【文本生成/文生图】使用了DeeSseek人工智能生成内容技术, 算法来源说明:本网站未自行开发、训练或部署深度合成算法,所使用的人工智能算法来自:DeepSeek、Chatgpt、Google Gemini。 2.为了网站加载速度优化,本站预览图片进行压缩处理,所以有些不是太清晰,并不影响AI提示词生图、AI设计等内容。3.任何个人或组织,在未征得本站同意时,禁止复制、盗用、采集、发布本站内容到任何网站、书籍等各类媒体平台。如若本站内容侵犯了原著者的合法权益,可联系我们进行处理。--GUUNN.COM