# How people prompt AI to build a course

> We analysed 37,483 people's AI course prompts to understand what goes with a published course. Fuller prompts are published nearly twice as often.

**Author:** Parshad Desai  
**Published:** 2026-11-09  
**Updated:** 2026-11-15  
**Categories:** L&D Teams, Instructional Designers

---

Coursebox AI's proprietary research data shows that what people write before they press generate goes with whether their course ever gets published.

I joined 3,000 AI course prompts to what the same people did next, and looked at the first prompts of all 37,483 people who used Coursebox's AI to build a course between May and November 2026.

**Prompts that make three or more core design decisions are published nearly twice as often as prompts that make none: 22.7% against 12.0%. Nearly half of all prompts make none.**

> From Coursebox AI's research report: [The State of AI Course Creation 2026](https://www.coursebox.ai/research/state-of-ai-course-creation)

## Half of prompts skip every design decision

**The typical prompt is 21 words, about one sentence.**

I looked for five core decisions in each prompt: who the course is for, their level, how long it should be, what learners should be able to do afterwards, and how learning will be checked. Nearly half of prompts make none of them.

The audience is the decision people make most, in 31% of prompts. A learning outcome appears in 17% and assessment in 16%.

## Fuller prompts get published

**Every core decision in the prompt goes with a higher publish rate.**

12.0% of people whose prompt makes none of the five decisions publish a course. With one or two decisions, it's 14.7%. With three or more, it's 22.7%.

_Fuller prompts are published nearly twice as often_

| Category | Value |
| --- | --- |
| No core decisions | 12% |
| One or two | 14.7% |
| Three or more | 22.7% |

Base: random 3,000 AI-classified learning-related first prompts, each joined to whether the same person published a course, May to Nov 2026 (1,541 with none, 1,208 with one or two, 251 with three or more)

## Structure beats attachments

**A course plan goes further than an attached document.**

Prompts that set the structure, with a number of modules or a list of lessons, are published 23.3% of the time, against 13.0% for prompts that don't. Prompts with a stated learning outcome (19.6%) or a named audience (18.3%) are also published more often than the 14.0% average.

Naming a source document barely moves the number: 14.6% against 13.7%. Across all 37,483 people, attaching a file looks like it helps until device and the number of prompts are equal. Then the gap disappears (11.5% against 11.1%).

Asking for images or video in the first prompt goes with a lower rate, 11.8%. Media can come later; the plan has to come first.

_Structure goes with publishing; media requests do not_

| Category | Value |
| --- | --- |
| Sets the structure | 23.3% |
| States a learning outcome | 19.6% |
| Names the audience | 18.3% |
| Names a source document | 14.6% |
| All prompts | 14% |
| Asks for images or video | 11.8% |

Base: random 3,000 AI-classified learning-related first prompts, each joined to whether the same person published a course, May to Nov 2026

## A paragraph beats a line

**Longer prompts are published more often.**

9.7% of people whose first prompt is under 10 words publish a course. For prompts of 300 words or more, it's 18.7%. The gap holds among desktop users who sent the same number of prompts (8.2% against 15.7%).

The device makes a difference too. 37% begin on a phone, and they publish at about half the desktop rate (7.2% against 15.4%).

Longer prompts don't seem to work by needing less fixing. People whose first prompt runs past 300 words are more likely to edit the course by hand afterwards (34% against 29% for prompts under 10 words) and twice as likely to ask the AI for changes (13% against 6%). They also take longer to publish: a median of 26 minutes after the course is generated, against 13. A long prompt looks like a sign of commitment that carries on after the draft.

_Longer prompts are published more often_

| Category | Value |
| --- | --- |
| Under 10 words | 9.7% |
| 10 to 29 | 12.3% |
| 30 to 99 | 12% |
| 100 to 299 | 16% |
| 300 or more | 18.7% |

Base: 37,483 people who sent a course-creation prompt, 14 May to 10 Nov 2026. Publish rate by length of the first prompt

## The AI asks, and people answer

**Seven in ten follow-up prompts are replies to the AI.**

76% of people send more than one prompt. I read a random 100 second prompts by hand. 71 were answers to the AI's clarifying questions about audience, level, focus and tone, or a choice between options it offered. Only 10 added new detail unprompted.

People who keep going publish far more often: 27.4% of people who sent four or more prompts, against 7.7% of people who sent one. Answering the AI's questions on its own goes with a much smaller lift (11.2% against 9.9%, July to October 2026). The publishing jump comes from people who keep refining.

_People who keep refining publish far more often_

| Category | Value |
| --- | --- |
| One prompt | 7.7% |
| Two or three | 9% |
| Four or more | 27.4% |

Base: 37,483 people who sent a course-creation prompt, 14 May to 10 Nov 2026. Publish rate by number of course-creation prompts sent

## Who writes the fullest prompts

**Course creators state outcomes most; corporate L&D teams name the audience most.**

Independent course creators state a learning outcome in 44% of prompts, training providers in 35%, corporate L&D teams in 27% and teachers in 20%. Corporate teams name the audience in 70% of prompts and attach their own documents in 44%, more than twice as often as creators.

## What this means for course builders

- **Make the five decisions first. **Audience, level, length, outcome and how learning will be checked.
- **Write outcomes as actions. **Finish "By the end, learners will be able to..." two or three times.
- **Give the course a shape. **Ask for a number of modules or list the lessons you want.
- **Attach the source, then add the decisions. **A document supplies the facts; the decisions still come from you.
- **Keep refining after the AI's questions. **Check the draft against your outcome before you publish.

For teams using [AI for employee training](https://www.coursebox.ai/ai-for-employee-training), this is the template I use. Replace the parts in square brackets and attach your document:

**Prompt: build a course with AI**

Paste into an AI course builder before generating, with your source document attached.

```
Create a course for [AUDIENCE].

By the end, learners should be able to:
- [OUTCOME_1]
- [OUTCOME_2]
- [OUTCOME_3]

Level: [LEVEL]
Length: [LENGTH], split into [MODULES] modules.
Assessment: [ASSESSMENT]
Build it from the attached [SOURCE] and keep to its procedures and terminology.
Tone: [TONE]
```

Replace these:
- `[AUDIENCE]`: The role, their experience and the setting (for example, new warehouse staff in their first week)
- `[OUTCOME_1]`: Something learners can do afterwards, written as an action (for example, lift and carry loads using safe technique)
- `[OUTCOME_2]`: A second outcome (for example, report a hazard using the site process)
- `[OUTCOME_3]`: A third outcome (for example, respond to a fire alarm)
- `[LEVEL]`: Beginner, intermediate or advanced (for example, beginner)
- `[LENGTH]`: Total time learners have (for example, 30 minutes)
- `[MODULES]`: Number of modules (for example, 5)
- `[ASSESSMENT]`: How learning will be checked (for example, a 3-question scenario quiz after each module)
- `[SOURCE]`: The document the course must follow (for example, warehouse safety policy)
- `[TONE]`: How it should read (for example, plain and practical)

Check before you use it:
- Every outcome has at least one module and one quiz question
- Facts and procedures match your source document
- The number of modules and the length match what you asked for

```json
{
  "_key": "v0109",
  "_type": "calloutBox",
  "body": [
    {
      "_key": "v0111",
      "_type": "block",
      "children": [
        {
          "_key": "v0110",
          "_type": "span",
          "marks": [],
          "text": "You don't have to remember every decision yourself. When you describe a course in Coursebox, the Course Agent reads your prompt and any files you attach, asks you about important details it can't find, such as who the course is for or how learning should be checked, and shows you the suggested details and course structure to review before anything is generated."
        }
      ],
      "markDefs": [],
      "style": "normal"
    },
    {
      "_key": "v0116",
      "_type": "block",
      "children": [
        {
          "_key": "v0112",
          "_type": "span",
          "marks": [],
          "text": "A one-line prompt still works; the gaps get filled before the course is built. "
        },
        {
          "_key": "v0114",
          "_type": "span",
          "marks": [
            "v0113"
          ],
          "text": "See how the Course Agent works"
        },
        {
          "_key": "v0115",
          "_type": "span",
          "marks": [],
          "text": "."
        }
      ],
      "markDefs": [
        {
          "_key": "v0113",
          "_type": "link",
          "href": "https://www.coursebox.ai/2-4"
        }
      ],
      "style": "normal"
    }
  ],
  "markDefs": null,
  "title": "Coursebox asks about what your prompt leaves out",
  "variant": "tip"
}
```

Length is one of the five decisions. The right answer depends on delivery: [corporate training is usually under 30 minutes or over three hours](https://www.coursebox.ai/blog/how-long-should-a-training-course-be). The outcome is also what the quiz should test, and [in 41% of quizzes every learner passes](https://www.coursebox.ai/blog/how-hard-are-online-course-quizzes).

## What the data can't tell us

These are associations. People who already know what they want write fuller prompts, and they may be more likely to finish a course anyway. The data shows prompts and whether a course was published. It doesn't show how good each published course was.

## About the research

The data is anonymised and aggregated: no names, email addresses or company details were used, no prompt text is published, and any group smaller than 30 is left out. I joined a random 3,000 AI-classified learning-related first prompts to whether the same person later generated or published a course, using a one-way hashed ID; prompts whose exact text was shared by several people (usually templates) were left out. Categories come from an AI classification of a random sample of 10,000 first prompts written in people's own words; a manual check of 40 classifications from each of the two AI models found about 92% correct, and 304 prompts labelled by both were checked by hand. Prompt length, file use, device and prompt counts use all 37,483 people who sent a course-creation prompt between 14 May and 10 November 2026, with bots, staff and test accounts removed. I read the 100 second prompts by hand. The template examples are illustrative and aren't taken from real prompts. The full methodology is in [The State of AI Course Creation 2026](https://www.coursebox.ai/research/state-of-ai-course-creation).
