A good ChatGPT prompt for a social story names the scenario, the student's reading level, and the 2:1 descriptive-to-directive sentence ratio that Carol Gray methodology requires, because ChatGPT will not apply that ratio unless you ask for it. In Emoquest's 2024 community survey of 16 parents, school SLPs, OTs, and SPED teachers, 94% reported spending 30 or more minutes on a single social story, and a vague ChatGPT prompt often adds editing time rather than saving it.
Why does a vague ChatGPT prompt produce a bad social story?
ChatGPT defaults to directive, instructional language because most prompts implicitly ask for advice, not a Carol Gray methodology narrative. A prompt like "write a social story about the fire drill" tends to return a list of "I will" statements: I will line up, I will walk quietly, I will not run. That output reads as a behavior plan, not a social story, because it skips the descriptive and perspective sentences that explain why the fire drill happens and how others feel about it.
What should the prompt actually contain?
| Prompt element | Why it matters | Example line |
|---|---|---|
| Scenario in one sentence | Anchors the whole story to a specific event | "[Student] has a school fire drill this Friday." |
| Reading level | Keeps sentence complexity appropriate for K-5 | "Write at a kindergarten reading level, 5 to 8 words per sentence." |
| Sentence-type ratio | Forces descriptive and perspective sentences to outnumber directive ones | "Use at least 2 descriptive or perspective sentences for every 1 directive sentence." |
| Student trait, not diagnosis | Personalizes without putting a clinical label in the chat log | "[Student] is sensitive to loud, sudden sounds." |
| Length cap | Keeps the story short enough for a K-5 student to sit through | "4 to 6 pages, 8 to 12 sentences total." |
| Ending instruction | Avoids a flat or punitive close | "End with a cooperative sentence about who will help, not a warning." |
90% of the work quote from the same survey: "Getting suitable pictures is 90 percent of the work." A better ChatGPT prompt fixes the text half of that ratio. It does not fix the picture half. Plan on sourcing or shooting photos separately no matter how good the prompt is.
What does a full prompt look like?
A prompt school SLPs report using successfully combines all six elements above into one instruction, for example: "Write a social story for [Student], a kindergarten student who is sensitive to loud, sudden sounds. The scenario is a school fire drill this Friday. Use a kindergarten reading level, 5 to 8 words per sentence. Include at least 2 descriptive or perspective sentences for every 1 directive sentence. Keep it to 4 to 6 pages, 8 to 12 sentences total. End with a cooperative sentence about who will help, not a warning." This is close to the prompt structure used by Emoquest's separate review of ChatGPT's Carol Gray compliance, which found the text output usable but never methodology-audited by default.
What will the prompt still not fix?
Three things a prompt cannot solve on its own. First, illustrations: ChatGPT's image generation does not hold a consistent character across multiple pages, so most SLPs still source their own photos or flat illustrations. Second, the ratio audit: ChatGPT does not check its own output against the 2:1 rule, so you still have to count sentence types yourself before using the story. Third, FERPA boundaries: putting a specific student's real name and identifying details into a general consumer AI product has not been formally cleared by most districts, so a placeholder name in the prompt, swapped for the real name afterward in your district-managed document, is the safer pattern.
Is AI-drafted text still considered evidence-based?
Social narratives as a category are recognized as an evidence-based practice by AFIRM's Social Narratives brief packet and by the National Clearinghouse on Autism Evidence and Practice (NCAEP), and neither body specifies who or what drafts the first version of the text. What the evidence review emphasizes instead is that the story is accurate, individualized to the student, and re-read on a consistent schedule. A ChatGPT draft that meets those bars starting from a well-built prompt is not disqualified by being AI-assisted, provided a human reviews it for the sentence ratio and factual accuracy before it reaches the student.
Frequently Asked Questions
What is the single most important thing to add to a ChatGPT social story prompt?
Naming the descriptive-to-directive sentence ratio (at least 2:1) explicitly. Without that instruction, ChatGPT defaults to a list of "I will" statements, closer to a behavior plan than a Carol Gray methodology social story.
Can ChatGPT generate the illustrations too?
ChatGPT with image generation can produce illustrations, but character consistency across pages is unreliable. Most school SLPs still find or shoot their own photos, since real photos are reported as the most effective visual for K-5 students.
Should the prompt include the student's name?
Use a placeholder like [Student] in the prompt itself and swap in the real name after ChatGPT returns the draft. Putting a specific student's name into a general consumer AI product is a FERPA gray area most districts have not formally cleared.
Does the prompt need to mention the student's diagnosis?
No. Describe the relevant trait (sensory sensitivity to loud noise, needs extra transition time) instead of naming a diagnosis. This keeps the story accurate and avoids putting a clinical label into a consumer AI chat log.
How long should the ChatGPT output be before you edit it?
Ask for 4 to 6 pages and 8 to 12 sentences total. Longer drafts usually mean ChatGPT padded with generic reassurance sentences that a K-5 student will not sit through.
What will ChatGPT still get wrong even with a good prompt?
The sentence-type ratio, since ChatGPT does not self-audit against the Carol Gray methodology criteria. Always re-read the draft once specifically counting directive sentences before you use it with a student.
Is a ChatGPT-written story still an evidence-based practice?
Social narratives as a category are recognized as evidence-based by AFIRM and NCAEP regardless of who or what drafts the text. What makes an individual story effective is that it is accurate, individualized, and re-read on a schedule.
One approach for school SLPs short on time is to keep a 5-tool stack: a methodology checklist (the ratio table above), a saved prompt you reuse and edit per scenario, a folder of stock photos sorted by scenario, an AI text drafter (ChatGPT, Claude, MagicSchool, or Emoquest for one-sentence-in story output), and a delivery format your district already uses (Google Slides or PDF). The prompt gets you a draft. The audit step still has to be yours.