How to Write a 60-Second Video Script From Facts With AI—With Shots and Sources

You can use this process for an explainer, a product demonstration, a service introduction, or a short news update. The worked example below uses NASA's DART mission. It shows the complete spoken script, the planned visuals, and the changes you need when the material contains dense numbers or conflicting sources.

Start with the brief and checked facts below. You can use the short prompt for a simple first draft, then follow the detailed steps when you need to resolve conflicting sources, plan several shots, or repair an overlong read.

Viewer and question: [who is watching and what they need answered]
Use only these checked facts: [facts, conditions, and source notes].
Write a spoken script for a video no longer than 60 seconds.
Then give the shots, screen text, and source note for each line.
Use these available visuals: [list].
Keep the conditions that make each statement true.
Flag missing information. Do not add statistics or testimonials.
Treat the length as an estimate; I will time the complete read.

For one short, checked source, a note beside each line is enough to track it. Add fact and source IDs when you combine reports or need to distinguish similar measurements. You can do the source checks by hand and use AI for drafting and editing.

Use a text-based AI assistant you can access. Paste the brief and relevant source passages into the chat; this method can work from pasted text. If you use a file attachment, ask the model to repeat the facts and source sections it read before drafting. Check that it has the material you intended to provide.

1. Choose the question your minute will answer

Write the viewer's question in one sentence. For the DART example, it is: "What did the impact change, and how did scientists measure the result?" That question gives the script a target: explain the experiment through a before-and-after measurement.

For your own video, fill in this brief before asking for prose:

Viewer: [who is watching, and what they already know]
Question: [the one question this video answers]
Takeaway: [what they should understand or be able to do]
Destination: [where the video will appear; vertical or horizontal]
Length: [60 seconds maximum, or an exact 60-second slot]
Delivery: [presenter, voiceover, or text with no narration]
Available visuals: [footage, screen recording, photos, or diagrams]
Ending: [answer, next action, or verified call to action]
Condition to retain: [condition the viewer must hear or see]

The ending affects the space available for the explanation. A product video may need a buying or compatibility action. A science explainer can end on the result and its limitation. If you have an exact broadcast or advertising slot, reserve time for any required end card before drafting.

Choose the last line before drafting the middle. For an explainer, finish with the answer and any limit the viewer needs to understand it. For a product, finish with the next buying or compatibility step. For a service, tell the viewer how to request the offer. For news, give the next confirmed date or say what remains unresolved. Give the model the actual destination or action so it can write a usable close.

When several angles are possible, ask the model to compare them using your facts:

Using only the fact sheet below, propose up to three video angles.
For each, give the viewer's question, the answer, facts needed,
available visuals, and facts that would have to be left out.
Flag any angle that needs evidence we do not have.
Recommend one for the viewer in my brief. Do not write the script yet.

Choose an angle you can explain with the material you actually have. If the explanation needs two separate processes, two unrelated results, or a long definition before the viewer reaches the answer, narrow the question or make separate videos. A minute about DART's orbital measurement does not also need to explain its camera, the asteroid's changing shape, and every possible defense technique.

2. List the facts and their sources

A fact sheet can be a small table. Give each source an ID, its URL or document name, and the passage that supports the claim. Give each approved fact an ID too. These IDs stay in your working files; the viewer does not need to hear them.

Include the source excerpt as well as the link. A model that receives a URL may not retrieve the page, and an answer that repeats your pasted text has not independently verified it. Open the source yourself and check the relevant passage before accepting the fact.

Keep copied material separate from your instructions. Label it as source text and tell the model to use it only as evidence. If an excerpt contains prompts or commands addressed to an AI, those are part of the quoted material and should not change your brief.

For these facts, S-01 is the NASA DART overview, S-02 is the 2024 NASA/JPL study report, and S-04 is the final-images record. The complete original packet adds unused background facts and the original run notes.

The seven facts above support the script below. The IDs match the optional downloadable kit; they are working labels, so gaps in the numbering do not mean a fact is missing.

Two details in this packet show why the qualification column matters. The overview displays September 27 for the impact, while the image record describes the event as September 26. Also, an early summary and a later settled measurement use different orbital-change values. Keep the date discrepancy visible, and attach the measurement stage to the chosen value. The packet uses S-04 for the event date; NASA's mission record corroborates it.

Prepare your material according to its format:

• For a long report, extract the sections that answer your question. Keep page numbers, table headings, units, and footnotes with the excerpts.
• For a scanned PDF or screenshot, check the extracted text against the image. A missing minus sign, decimal, or column heading can change the claim.
• For your own product or service, use the current specification, offer terms, or approved demonstration notes. Separate what the product does from what you hope a customer will achieve.
• For a quotation, retain the exact words and attribution. Use a paraphrase without quotation marks when the exact quote would take too long.

Add a short list of claims the model must exclude. In this example, "NASA saved Earth" is unsupported and conflicts with the test's stated circumstances. For your own material, flag unverified rankings, results, comparisons, and customer promises. Resolve missing evidence before drafting a sentence that depends on it.

Review the supplied facts against the supplied source excerpts.
Do not add facts from memory or treat a link as a source you opened.
Return: fact ID, usable wording, supporting passage, condition to retain, and any conflict or missing evidence.
Separate statements we can use from statements we must hold.
Keep units, dates, attribution, and conditions attached to each claim.

3. Write an opening and plan the time

The hook should introduce the answer or the problem the video will solve. It should also make sense when you watch only the first shot. An asteroid aimed at Earth would give the DART example a false premise, even if the narration corrected it later.

Use an opening that fits the evidence you have:

A claim such as "You won't believe what happened" gives the writer no guidance about what comes next. A concrete result lets you build the rest of the minute around an explanation.

Use a word count only to size the first draft. As a practical starting budget, try about 110 to 130 English words for narration with pauses, names, and numbers. This is an editorial starting point, not a universal speaking rate. A quiet visual sequence or a slow presenter may need fewer words. Your measured delivery decides the final length.

If you know your comfortable speaking rate, estimate the space with words = words per minute × speaking seconds ÷ 60. At a measured rate of 130 words per minute, 54 seconds of speech allows about 117 words. The other six seconds might hold a demonstration pause or an end card. Do not assume that a text-to-speech rate setting predicts a human read.

Plan each beat by its job before writing the sentences: opening, explanation, evidence or demonstration, necessary condition, and ending. The example uses seven shots because its explanation needs them. A screen demonstration with one continuous action may need fewer.

4. Draft narration you can read without the table

Ask for the spoken words in one block first, followed by the production table. This lets you hear whether the explanation flows. A table can hide an abrupt jump between facts because each row looks tidy on its own.

Write a factual video script using only the approved facts below.
Follow my brief and use one main takeaway.
Aim for [word budget] words as an estimate; do not claim it is timed.
Give the complete spoken narration first.
Then give a table with planned start/end seconds, voiceover,
on-screen text, shot direction, and fact-to-source mapping.
Keep IDs and URLs out of the spoken copy.
Preserve qualifications that change the meaning of a claim.
Use only visuals I have listed, or label a proposed diagram clearly.
Flag any missing evidence. Do not invent quotations, results,
customer experiences, urgency, or a call-to-action destination.

This DART example is about 110 words on the page. Read the names and measurements aloud as you would record them; the word count only helps size the draft.

NASA deliberately crashed a spacecraft into an asteroid moon. After the impact, the moon took less time to complete an orbit.

The spacecraft was DART, a test of asteroid deflection by impact.

In 2022 it hit Dimorphos, which orbits the larger asteroid Didymos.

Neither asteroid threatened Earth. Scientists used the pair to test the technique.

Before impact, one orbit took about eleven hours and fifty-five minutes.

A 2024 study found that one orbit took thirty-three minutes and fifteen seconds less, once the orbit had settled.

The test demonstrated one possible way to deflect an asteroid. More research is needed to understand how it would work on other targets.

The fact sheet supplies the sources for each paragraph. The revised ending leaves the viewer with the meaning of the experiment. Source links belong in the accompanying description or another usable location you arrange before publishing; reading internal IDs aloud would give the viewer no way to find the evidence.

The time allocations below are estimates. This example has not been measured with a complete read; adjust the words and shot boundaries to your delivery.

5. Make the script sound natural aloud

Read the narration aloud without the shot directions or source columns. Mark the places where you run out of breath, stumble over a term, or have to reread a sentence to understand its subject. Give those specific problems to the model. "Humanize this" on its own leaves too much room for invented personality and extra claims.

Contractions and ordinary verbs can help, but you do not need to force jokes, slang, or personal reactions into a factual explainer. Avoid having the model add "I tried this" or "our customers love it" when your material contains no such evidence.

Be especially careful when shortening qualifications. Keep "about" with an approximate measurement, "up to" with a conditional maximum, and the study population with a research result. If a qualification is too long for the minute, simplify or remove the claim that needs it.

Edit this narration for [viewer] and [delivery style].
Keep its facts, qualifications, names, numbers, and attribution.
Use ordinary verbs and clear subjects. Split sentences that make
me carry several unfamiliar names or measurements at once.
Remove repeated explanations, hype, staged openings, and filler.
Do not add personal experience, humor, quotations, or new claims.
Keep the factual hook. Return the revised narration, then identify
any edit that changes which fact or qualification a line contains.

After the edit, compare the new lines with your fact sheet. A smoother sentence still needs to say the same thing. For a presenter, leave room for their natural phrasing after they understand the fact; then check that final version too.

6. Assign shots that explain each line

Each shot should give the viewer something useful to look at while the line is spoken. A diagram can show a relationship. A close-up can show a product feature. A screen recording can show where an action happens. A decorative montage gives you little help when the sentence needs a precise explanation.

Here is a proposed 60-second layout for the DART script. The intervals are planning allocations. Adjust them after recording the narration; each spoken beat must fit its own window as well as the whole minute.

Use the same object names and units in the narration, screen text, captions, and description. Label an orbit diagram as simplified. A diagram showing a circular path explains the relationship; it should not suggest that every detail of the post-impact orbit is drawn to scale.

Keep "about" on the rounded baseline. Use the study's reported change instead of calculating a new exact difference from that rounded value.

On-screen text can shorten what the viewer must read, but it cannot quietly strengthen the claim. "33 min 15 s shorter" needs enough context to identify the orbital period and measurement stage. A giant "NASA SAVED EARTH" headline would conflict with the narration and the sources.

Keep captions and emphasis text separate in your plan. Captions follow the spoken words; a measurement label stays near the diagram it explains. Preview them together on a phone-sized display so you can catch collisions, text hidden by interface elements, or a value that disappears before it can be read.

For every visual, record the exact file, its credit, and whether it is a photograph, illustration, screen recording, or original diagram. If you replace a diagram with NASA imagery, check that asset's page and NASA's media guidance. The guidance addresses third-party material, credits, endorsement, and AI applications. A source link alone does not establish permission for every use.

7. Adjust the workflow to your video type

A product demonstration

Start with the task the buyer wants to complete. Choose the feature that helps with that task, then show it in use. A specification can support a feature claim. A customer outcome, comparison, or performance claim needs evidence that supports that separate statement.

Use this sequence to organize your facts: [task] → [feature in action] → [relevant condition] → [available next action]. If the specification gives an "up to" result under particular conditions, retain those conditions or choose a simpler feature to demonstrate. A quick demo of one task cannot support a promise about every user's results.

Give the model the exact shots you have: the control being pressed, the attachment fitting, or the output appearing. Ask it to write around those shots. If the draft requires footage you cannot obtain, change the script before recording the voiceover. Keep price, availability, and compatibility claims tied to current information. For producing the footage itself, use the product-video guide.

A service or small-business introduction

The viewer needs to know whether the service fits their situation and what happens next. Put the intended customer and deliverable near the start. Then explain the scope, any condition that affects the offer, and the action the customer can actually take.

A useful script skeleton is: We help [customer] with [specific task]. You receive [deliverable]. The service includes [scope] and requires [condition]. To start, [action you can offer]. Replace every bracket with confirmed information. If you have no evidence for a promised turnaround or financial result, leave it out.

Show the deliverable or a clearly identified sample of the process. Do not have the AI turn a blank template into a completed customer result, or invent a testimonial to give the script an ending. Make sure the named booking page, form, or contact method exists before inviting people to use it.

A news or announcement update

Anchor the video to the confirmed development and its event date. Separate an announced plan from something already available, and distinguish the date of the announcement from the date the change takes effect. An article's publication date does not always establish when the event occurred.

Ask for a fact sheet with columns for "confirmed," "attributed statement," and "unresolved." Write the narration from the confirmed facts. Attribute a claim when the source only establishes that someone said it. If two sources disagree, check whether they describe different stages or dates before calling the difference a contradiction.

Choose what the viewer needs to know now: what changed, who it affects, and what remains unknown. Avoid making an old announcement sound like a new event through words such as "today" or "just launched." Recheck time-sensitive details when the final video is ready to go out.

A research result or number-heavy explainer

Identify what the number measures before presenting the change. Preserve its unit, baseline, population, time period, and any qualification the conclusion needs. With percentages, check whether the source reports a relative change or a percentage-point difference. Do not let the model convert an association into a causal claim.

Use one comparison the viewer can follow. In the DART script, the central comparison is the time needed for an orbit. Extra measurements about size or another orbit would need their own explanation and screen labels, so they stay outside this minute.

If the exact value is essential, give it a dedicated visual and enough time in the read. If a rounded value is appropriate for your question, record the approved rounding in the fact sheet and use it consistently. You can put the exact table or source in the accompanying material. Never remove a qualifier solely because it makes the number look cleaner.

A script translated into another language

Translate the checked script after choosing its angle and facts. Ask the model to preserve units, qualifiers, names, attribution, and how certain the wording is. Keep the original fact IDs beside the translated lines so you can check the meaning.

Then record a complete read in the target language. An English word budget and timing record cannot establish the translated duration. Expand numerals into the form the narrator will actually say, and check technical terms with someone who understands both the topic and language.

If the translation runs long, remove a lower-priority fact or simplify the explanation in both the narration and visuals. Recheck the translated captions separately; a short display label may lose a qualification that remained in the spoken line.

A video with screen text and no narration

Build the timing around reading and seeing the action. Keep each screen focused on one statement, and let a diagram or demonstration carry the part it can explain. Test the sequence without pausing it. If you cannot read the text before the next card, shorten it or lengthen the hold.

Audit text cards as claims in their own right. A caption-only version cannot rely on a voiceover to explain an ambiguous number or supply a missing condition. Keep evidence in the production table and provide usable source links with the finished piece.

8. Time the whole read and each difficult beat

Record a scratch read at the pace you would use for the audience. Include pauses, pronunciations, and any moment when you want the viewer to inspect the screen. Time from the intended first sound through the final word, and account separately for opening silence and the end card in the total video.

Mark the end of each beat as well as the total duration. A 58-second voiceover can still fail a plan if the measurement paragraph takes longer than its assigned shot. Let the recording determine the final shot boundaries, or revise the paragraph that cannot fit.

In the DART example, the comparison paragraph has the most demanding pronunciation. Its planned window is 16 seconds. If your read takes longer, give that shot more time or simplify the wording while keeping the measured change and its stage.

A number-heavy draft could say: "A 2024 study put the settled orbit at eleven hours, twenty-two minutes and three seconds. That was thirty-three minutes and fifteen seconds shorter. Earlier summaries reported about thirty-two minutes."

The example's spoken version says: "A 2024 study found that one orbit took thirty-three minutes and fifteen seconds less, once the orbit had settled." It gives the listener one comparison to follow. The exact period stays in the on-screen label, and the source notes explain the earlier measurement. Keep the voiceover and label attached to the same study result.

When a read is too long, make cuts in this order:

1. Remove repeated framing and sentences that restate the previous line.
2. Remove a supporting fact that the takeaway can survive without.
3. Replace a long noun phrase with an ordinary verb, while preserving the claim.
4. Move a secondary detail to accompanying material if the viewer can understand the video without it.
5. Narrow the question when the necessary explanation still cannot fit.

Keep the condition that makes a claim true. Increasing the delivery speed may help a presenter who naturally speaks faster, but it will not repair an explanation with too many unfamiliar ideas.

My complete read took [seconds]. The target is [seconds], including
[seconds] reserved for pauses or the end card.
The difficult beats were: [line, observed duration, and problem].
Suggest the smallest cuts that preserve the takeaway and facts.
Keep these qualifications: [list].
Return the revised narration and a list of removed claims.
Treat the new timing as unmeasured. I will record it again.

If the read is shorter than the slot, check whether the viewer has enough time to inspect the demonstration or comparison. You can use that time for a useful hold. Add narration only when a missing explanation prevents understanding. A maximum of 60 seconds does not require filler to reach exactly 60.

For synthetic narration, measure the audio produced from the exact final text. Check how it says names, units, abbreviations, and numerals. Changing the text, voice, or rate creates a new version that needs timing again.

9. Check what the words and images tell the viewer

Give a fresh chat the fact sheet, source excerpts, final script, and shot table. Ask it to identify problems without rewriting them first. A separate conversation can help expose an assumption repeated during drafting, but it is another AI review and may repeat the same mistake. Open the sources yourself for claims that determine the result.

Audit this script and shot plan against the supplied fact sheet
and source excerpts. Do not repair it silently.
For every claim in narration, screen text, and visual implication,
return the line or shot, fact ID, exact support, and verdict:
supported, needs qualification, unsupported, or conflicting.
Check dates, units, baseline comparisons, actor/target identities,
causal wording, promised results, and the call to action.
Separate factual problems from unclear delivery or missing assets.
Suggest the smallest fix for each problem.
Do not treat this review as proof of timing or media permission.

Review a combined row claim by claim. In the DART table, S-04 supports the impact fact, while S-01 and S-02 support the asteroid relationship. Listing three sources beside a sentence does not establish that each source supports every part of it.

Check the hook especially closely because it establishes the viewer's interpretation. Then check whether the visual suggests more than the sentence. A dramatic collision path, a before-and-after image, or a customer reaction can add a claim that never appears in the narration.

Finally, check the proposed source invitation and call to action. In the original run, the reviewer rejected "sources listed with this video" because no accompanying source list had been established. The practical repair is to arrange a usable source list and verify it before using that invitation. Keep internal file labels in the working table.

10. Fix common script problems

11. Hand over a script someone can record

Save the final narration as plain text and the shot table as a spreadsheet or CSV. Include start and end times, the exact spoken words, screen text, shot directions, notes showing the source for each fact, asset files, credits, and unresolved notes. Link to the source passages from the evidence sheet so another person can check a line without repeating your research.

Before handing it to the editor or presenter, confirm:

• The hook, explanation, and ending answer the same viewer question.
• The recorded read fits the intended slot, including planned pauses and the end card.
• Every beat has enough time for its narration and visual task.
• Each factual line and screen label has support, including its conditions.
• The proposed shots can be produced from the available assets.
• The final source invitation or other call to action leads somewhere usable.

After assembly, watch the exported video once without pausing and once with the fact sheet beside it. Check captions, names, measurements, credits, and the actual images. If the editor changes a factual line or adds a visual comparison, review that change and measure the affected timing again. For a more detailed claim review, use the source-fact audit workflow.

The original DART run and downloadable evidence

The August 24, 2026 example used a signed-in ChatGPT drafting chat and a separate review chat. The model identifier was not retained. The public packet was pasted; no successful upload was claimed. The original seven-row plan covered 0 to 60 seconds, and its exact final synthetic narration measured 56.43 seconds. A human performance and final rendered video were not tested.

The original evidence files below retain that run. The rewritten narration and proposed timings earlier in this article are separate teaching material. The original checker and timing record do not validate the rewrite.

• Complete original review kit: the fact packet, prompts, table, narration, audio, screenshots, and checks.
• Fact packet and source ledger: approved facts, qualifications, and source conflicts.
• Original prompt pack and run record: the drafting and review interactions.
• Original production table and original narration: the retained words and shots.
• Original synthetic read and timing record: the voice setting and measured duration.
• Optional Python checker and retained test results: the original bundle's structure checks and nine rejected changes.

View the original production board and ChatGPT screenshots

These retained images document the original run. Their wording, counts, statuses, and timings apply to the original files.

The optional checker checks whether the original files still match their recorded content and timing, and whether required table fields are present. It cannot decide whether a source is true or a visual is misleading. After editing the files, update its expected content and timing before using it.