Define 'Explain the Essence, Don't Read Aloud' in 4 Parts for AI── A complaint that narration just reads the slides becomes a definition an AI can act on once it is split into tone (speaker and listener) and role (for each point, why it matters and what it means concretely)

01The operator sent back narration that only read the slides aloud

In April 2026 the operator was working with an AI on a slide video with spoken commentary. The narration the AI drafted turned the text on each slide into speech, almost word for word.

The operator did not accept it. If the narration only reads the slides, there is no reason to have narration at all. The voice should explain the substance of what each slide says. That was the reply sent back to the AI.

The same day, the operator went a step further. A specific manuscript was named as the reference version, and the narration was to match its way of speaking. The operator also spelled out what that way of speaking was. A presenter talks directly to the audience. The presenter unpacks each key point on the slide. And the presenter explains why the point matters and what it means in concrete terms.

The first reply was a complaint: this is pointless. The second request was a definition of how to speak. This episode looks at what happened between the two, and turns it into a method: how to convert a vague sense that something is wrong into a definition an AI can act on.

02The complaint about reading aloud is really a complaint that the narration has no job

Saying "it just reads the slides" is not a remark about voice quality or pace. It says the narration adds nothing the slide does not already show.

Slide and narration reach the same person at the same moment. If the voice repeats the slide text, the listener receives the same information twice. Nothing new is learned.

So the fix is not better wording. It is a defined role. What is the narration responsible for? Until that is settled, the AI falls back on the only material it has, which is the slide text.

03Putting the same words before the eyes and the ears lowers understanding

If reading aloud were merely useless, it would do no harm. Research in educational psychology says it does harm.

The educational psychologist John Sweller summarised a body of findings showing that presenting the same information in more than one form interferes with learning rather than helping it. This is known as the redundancy principle. Listeners spend effort matching the words they read against the words they hear, and have less capacity left for the content itself.

A 2023 literature review added a condition. Repeated content is most likely to hurt learners who already have background knowledge in the field. Beginners facing complex material can sometimes benefit from the repetition.

In pharmaceutical work, the audience for slide material is often physicians, pharmacists or specialist staff. The more the audience already knows, the more likely reading aloud is to get in the way.

Speaking directly to the listener has support too. Richard Mayer reported that when the words of a lesson move from formal written style to conversational style, learners get better at applying what they learned to new problems. The effect appeared in all eleven comparisons he reviewed. It weakens, though, when the tone becomes too casual.

04The definition has two parts: the speaker's tone and the two things the narration adds

Break the operator's second request apart and the definition of narration has four elements. Each becomes one line of the instruction given to the AI.

Elements 1 and 2 set the tone. Elements 3 and 4 set the role. Specify tone alone and the voice becomes friendlier, but what it says can still be the slide text. Specify role alone and the content grows, but it may come out as written commentary rather than speech. Only with both does it become a definition.

The definition has a boundary as well. Explaining the substance does not mean ignoring the slide and talking freely. The order follows the slide's key points, taken one at a time. What gets added is reasons and examples, not new claims.

05In material review, unpack the point but keep efficacy wording and numbers unchanged

Pharmaceutical work has plenty of slides with a voice track: internal training videos, presentation decks for product briefings, explainer videos for healthcare professionals. The first draft of that narration can be requested from an AI.

Here the definition needs one more fence. When unpacking a point, do not rephrase indications, dosage and administration, adverse event frequencies or trial figures.

The reason is plain. Ask an AI to say why a point matters and it will try to state the point strongly. Along the way, phrasing beyond the approved scope or rounded numbers can slip in. Spoken material is subject to review just as the slides are.

So for promotional or medical narration, the definition has two layers. What may change is the conversational tone and the added reasons and examples. What may not change is any wording or number tied to the approved label. Writing both layers into the instruction also tells the reviewer where to look hardest.

The reviewer's own questions follow from the same layers. Has the narration added a claim the slide does not make? Are the added examples inside the approved scope? Do the numbers match the slide exactly?

06Without a stated role, an AI restates the slide text

Why does an AI return narration close to reading aloud? The mechanism is simple.

Ask only "add narration to this slide" and the AI's sole material is the slide text. Nothing says who is speaking, to whom, or for what purpose. When the text is all there is, restating it is the answer least likely to stray from the request.

Anthropic's published guidance on prompting Claude addresses exactly this. Setting a role, even in a single sentence, focuses Claude's behaviour and tone on the use case. Explaining the context or the reason behind an instruction helps Claude understand the goal and give a more targeted response.

The operator's second request met both conditions. It assigned the role of presenter and specified a tone of speaking to the audience. It said what to add, in two parts: why the point matters and what it means concretely. And it supplied a reference manuscript. Putting the complaint into words supplied the material the AI had been missing.

07From tomorrow, hand over speaker, listener, the additions and the fixed wording

When you ask an AI for narration, give it these five things along with the slides.

1. Name the speaker. One sentence on whose voice this is, such as "speak as the presenter".
2. Name the listener. Who is being addressed and how much they already know. Physicians and first-year staff need different depths of explanation.
3. State what to add. For each key point, add why it matters and what it means concretely, and do not read the slide text as written.
4. State what not to change. Indications, dosage, figures and warnings are to be spoken exactly as on the slide.
5. Supply a model. If there is a manuscript or an earlier episode that sets the tone, name it.

Then have the AI draft one slide first. Check whether it has slipped back into reading aloud, and whether the added examples stay in scope. If it is right, extend it to the whole deck. If not, find the missing element and add it to the instruction.

Written once, these five lines can be reused. For the next video, only the listener and the model need to change.