Craft
Writing narration from written prose
Your document reads well. Read aloud, it will not work, and the reasons are specific and fixable.
1. Break the subordinate clauses
A sentence with two subordinate clauses is fine on a page, because a reader can hold the structure visually and re-read the opening if the ending surprises them. A listener cannot. By the time the main verb arrives they have lost the subject.
Split at every clause boundary. Three short sentences beat one correct long one, every time, in narration.
2. Kill the passive voice
'It was determined that the intervention produced no measurable effect' is standard academic register and unlistenable. 'We found no measurable effect' is the same claim, spoken.
Passive constructions hide the actor, and a listener trying to work out who did what has stopped following the argument.
3. Replace every reference to the document
Written prose is full of navigation: 'as noted above', 'see Section 4', 'the following table', 'the aforementioned constraint'. None of these mean anything in a linear medium — there is no above, and the listener cannot see Section 4.
Replace with a restatement. 'As noted above' becomes 'we said earlier that...' with the actual content repeated, because the listener probably does not remember it.
- 'See Figure 3' → 'this chart shows...' with the chart on screen.
- 'As discussed in Section 2' → restate the point in one clause.
- 'The following' → 'here is' or just say the thing.
- 'The aforementioned' → name it again.
4. Unpack the acronyms, once
A reader who forgets an acronym scans back to the definition. A listener cannot, so an acronym introduced at minute one and used at minute six is noise.
Expand on first use in narration, then use the acronym only if it appears frequently after that. If it appears three times across ten minutes, say the full term every time — the redundancy costs four seconds and saves comprehension.
5. Say the numbers the way a person says them
'A 34.7% increase' reads fine. Spoken, 'thirty-four point seven per cent' is precision nobody retains. Say 'about thirty-five per cent' in narration and put 34.7% on screen, where the precision belongs.
The exception is regulated content, where the exact figure has to be spoken. There, slow down and let it sit — a precise number needs a beat either side of it.
6. Add the signposts back
Written documents signpost visually with headings, whitespace, and numbering. Strip those and the listener has no idea where they are in the argument.
Put the signposts into the narration: 'there are three reasons for this — here is the first', 'that is the finding; now what it means for practice'. These feel clumsy on a page and are essential in audio.
The test that catches everything else
Read the script aloud, all of it, at pace. Every place you stumble, run out of breath, or have to re-read a clause is a place the narrator will too, and a place the listener will lose the thread.
This takes as long as the video runtime and catches more than any checklist. It is also the step that most reliably gets skipped.
Frequently Asked Questions
Can AI do this rewrite reliably?
The mechanical parts, yes — clause splitting, passive voice, document references. What it also does is drop qualifiers and firm up hedged claims, so the output needs reading against the source.
How much longer does narration get after rewriting?
Usually ten to twenty per cent, because splitting clauses and restating references adds words. Budget for it when estimating runtime.
Should I write the script or edit an extracted one?
Edit an extracted one. Starting from your document's own sentences keeps the claims accurate; starting from a blank page invites you to restate things slightly wrong.