Lesson 3 of 8 · 7 min read

Script for the ear

Write narration from the real process document, in spoken words, with the extras cut, and lock it before any voice or video is generated.

Write from the process, not from memory

The script comes from the real process document, not from what an assistant knows about the topic in general. An assistant asked for a training script with no source writes a plausible one. A plausible script is not the same as your process.

Learners report what happens when nobody checks. One learner judged about 15% of an AI-generated data-integrity script to be nonsense. A trainer asked a manager for bullet points and received a 19-page script two minutes later, for an application the manager did not know. These are single posts, not studies, but they show the failure: a fluent script nobody checked against the process.

So give the assistant the process document and the scenario questions from Lesson 2. Tell it to write only what the document supports. Where the document is silent, the assistant marks a gap instead of filling it. Each gap goes to the process owner, and the answer goes back into the document.

Some creators work the other way round. One creator records the training first, transcribes and edits it in Descript, and uses the cleaned transcript as the script. That method starts from what an expert says aloud. It still needs the same check against the process document.

Talk to the learner

Write the way a colleague would explain the task at the next desk. Use “you” and “we”. Mayer’s review found that a conversational style beat a formal style, with a median effect size of d = 1.11 across 11 lab experiments. One classroom study, with online chemistry tutors, did not replicate it.

In practice, “The return reason must be verified against policy” becomes “Now check the reason against the 30-day rule.” Keep sentences short, because a listener cannot reread. Say what the learner will see and do, in the order they do it.

Cut what the task does not need

The coherence finding is the strongest in Mayer’s review. Learners did better when interesting but unrelated sentences, pictures or clips were removed. The median effect size was d = 1.66, and Mayer calls the evidence strong and consistent.

Cut the history of the old system, the story of why the company changed it, and the jokes. If a sentence does not help a learner answer one of the scenario questions, it is a candidate to cut. A one-line reason for the change can stay. The rest belongs in an email.

Narrate the screen, and keep screen text short

In a screen recording, the narration describes what the screen shows. “Click Returns, then type the order number in the search box” tells the learner where to look. Narration about something else while the screen moves splits the learner’s attention.

Do not print the narration on screen as well. In Mayer’s redundancy studies, narrated animation beat the same animation with the narration also printed on screen. The median effect size was d = 0.72 across five studies.

The effect shrank when the on-screen text was cut to a few key words beside the picture. So put key words on screen, such as “30 days from delivery”, not full sentences. Captions that learners can turn on and off are a separate matter, and Lesson 7 covers them.

Start the pronunciation list now

An AI voice reads what it is given, and it guesses at names and codes. One learner reported an AI voice that said “genAI” three different ways in one course. The captions came from the audio rather than the approved script, so they carried the errors. A buyer who paid about $11,000 for a custom avatar reported that it could not pronounce industry terms.

Start a pronunciation list while you write. List every product code, supplier name, acronym and system term, and write how each one should sound. Lesson 5 uses the list when you choose a voice, and Lesson 7 checks each term in the final render.

Read it aloud with a timer

Read every draft aloud, at the pace you want the voice to use, with a timer running. Reading aloud finds what silent reading misses. It shows sentences too long to say in one breath, words a listener could mishear, and places where the words and the screen do not match. The timer gives you the length, which Lesson 4 needs for the segment plan.

Use one term for each thing. If the screen says “Return reason”, the narration says “return reason” every time, not “cause” or “why they sent it back”. A listener who hears a new word assumes a new thing, and cannot scroll back to check.

Lock the script before anything is generated

EdzLMS, a learning-platform vendor, puts this step first in its production checklist: finish and approve the script before any generation. A change after the voice, the recording and the captions exist means redoing all three. Approval means the process owner has read the script against the process and signed it off.

Keep the process document and your script rules in a Project in Claude or ChatGPT. Every draft then starts from the same source and the same rules, instead of whatever you pasted that day. The steps below show where to set it up.

Worked example · Larkfield Supply

Riley drafts segment one

Riley has the returns process document from the order-system team, plus the objective and the four scenario questions from Lesson 2. Riley sets up a Project that holds the document and the script rules, then drafts segment one.

  1. Riley asks the assistant to draft segment one from the process document only, in “you” and “we”, with gaps marked.
  2. The draft comes back with two gaps marked. The document does not say what staff do when the customer has no order number. It does not say whether a part-used box of paper can be returned.
    What would you do next?What should Riley do with the two gaps?Think it through, then show the answer

    Send them to the order-system team, who own the process. Riley adds their answers to the process document in the Project, so every later draft uses them. A guess would put an unapproved rule into a training video.

  3. Riley reads the draft aloud with a timer. It runs about 5 minutes, and the opening explains why Larkfield changed systems.
  4. Riley cuts the opening to one sentence and removes a paragraph about the old system. The segment drops to about 4 minutes. Each remaining paragraph supports at least one scenario question.That answers the question you guessed at: Riley’s first draft runs about 5 minutes and opens with why Larkfield changed systems. Guess: what should happen to that opening?
    What would you do next?The draft says, “The 30-day window is calculated from the date of delivery as recorded in the dispatch module.” How should Riley rewrite it for the ear?Think it through, then show the answer

    In spoken words that point at the screen: “Look at the delivery date, here. The 30 days start from this date, not the order date.” The on-screen key words read “30 days from delivery”.

  5. Riley starts the pronunciation list. LKS is read as three letters, “L-K-S”. Aubriot is read “oh-bree-OH”. The list goes into the Project with the process document.
  6. Morgan and the order-system lead read the script against the process and approve it. Riley marks it locked, with the date.

Segment one is a locked script that runs about 4 minutes aloud, in short spoken sentences, with key words marked for the screen. Nothing has been generated yet. Lesson 4 turns the script into a storyboard.

Starter prompts

  • Draft a script from the process document

    Write the narration for one segment of a training video. Use only the process document in this project, or this one: [paste the process document]. The learning objective is: [paste the objective]. The segment must prepare learners to answer these scenario questions: [paste the questions]. Write [segment name], about [length] minutes when read aloud. Speak to the learner as "you", and use "we" for the team. Use short spoken sentences. Describe what the screen shows at each step. Leave out history, background and anything the questions do not need. Mark on-screen key words in [brackets], a few words each, never full sentences. Do not invent any rule, number or step. Where the document does not say, write "GAP:" and the question I need to ask.

  • Read it as a listener

    Here is a draft narration script: [paste the script]. Review it as someone who will hear it once and cannot reread. List: sentences over 20 words; formal or passive phrasing to rewrite in spoken words; any sentence that does not help with these questions: [paste the questions]; any place where the narration does not say what is on screen; and every product code, name, acronym or term that a synthetic voice could mispronounce. Do not rewrite the script. List each problem with one suggested fix.

  • Project instructions for script drafts

    You draft narration scripts for training videos at [organization]. Use only the documents in this project. Never invent a rule, step, number or name. Mark anything the documents do not cover as "GAP:" with the question to ask. Write in a conversational style: "you" for the learner, "we" for the team, short spoken sentences. The narration describes what the screen shows. On-screen text is key words only, in [brackets]. Cut anything the learning objective does not need. End every draft with a pronunciation list of each product code, name and acronym, and how it should sound.

Steps for your assistant

In Claude

Checked on October 15, 2026

  1. Open Projects in the left sidebar, click + New Project, and give it a name. Claude does not see the name or the description.
  2. On the right side of the project page, click + and upload the process document and your pronunciation list. Claude uses project knowledge in every chat in the project.
  3. Click Set project instructions, paste the “Project instructions for script drafts” prompt, and click Save instructions.
  4. Draft each segment in a new chat inside the project. Chats do not share context unless it is in project knowledge, so add the locked script there when it is approved.

Anthropic’s help page lists Projects on every plan, including Free. Free users can create up to five projects.

In ChatGPT

Checked on October 15, 2026

  1. Select New project in the sidebar and enter a name.
  2. Upload the process document and your pronunciation list, or paste them as text.
  3. Open the more options menu (•••), choose Project settings, and paste the “Project instructions for script drafts” prompt as the project instructions. They apply only in this project and override your global custom instructions.
  4. In Project settings, under Memory, choose Project-only memory to keep this project’s context out of other chats.

OpenAI says projects are available to signed-in users, subject to plan and workspace settings. Free plans allow 5 files per project, Plus and Go allow 25, and Pro allows 40.

When not to use AI

  • Filling holes in the process. If the process document is missing or out of date, an assistant will supply plausible steps. Get the document fixed first.
  • Approving the script. The process owner reads it against the process and signs it off.
  • Deciding how a name is said. Ask the supplier or the person, and write the answer in the pronunciation list.
Recall from memoryWhich finding is the strongest in Mayer’s review, and what test decides whether a sentence stays in your script?Answer it without looking back, then show the answer

Coherence: learners did better when interesting but unrelated material was removed. If a sentence does not help a learner answer one of the scenario questions, it is a candidate to cut.

Exercise · your project

Draft, cut and lock the script for your first segment.

  1. Set up a Project with your process document and the “Project instructions for script drafts” prompt.
  2. Draft one segment with the first prompt, and send every gap to the process owner.
  3. Read the draft aloud with a timer, and cut anything your scenario questions do not need.
  4. Start a pronunciation list with every code, name and acronym in the script.
  5. Get the process owner’s approval, and add the locked script to your video plan.

Sources