Script generation

The prompt

Everything starts from one text box. Click New project in the sidebar, or type in the box on Home. Describe the video in a sentence or two: the topic and the angle you want. The more specific, the better.

Then pick the format (Landscape or Shorts), a length and a tone. For Landscape the lengths are Short, Standard, Long and Deep dive. The tones are Documentary, Investigative, Explainer, Punchy, Cinematic and Casual. The length you pick decides how long the script is. Words in the prompt don't change it.

You can also add a Hook angle: one sentence on how the video should open. Leave it blank to let the AI choose.

Prompts that work

"How spices changed world history, focusing on the trade routes"

Subject + angle. The angle tells the writer what to focus on.

"The collapse of the Roman denarius, and what it did to the army"

A clear subject and a question to answer give the outline a spine.

You can also skip the AI writer and use your own script. Click the + button on Home and pick Use my own script, or open the Paste script tab. Markup gives you control over the visuals: # chapter title, ## headline, > quote, {{source: ...}}, {{speaker: ...}} and {{bullets: ...}}. Everything else works the same way: scenes, voiceover, b-roll and motion graphics, made from your text.

Already recorded your narration? Upload your own voiceover instead (+ > Use my own voiceover, or the Upload voiceover tab). You can add up to 20 audio files. Each one is turned into text and becomes its own scene, so footage, captions and motion graphics line up with your voice.

Research → outline → sections → editor's pass

The writer works like a small documentary team: a researcher, a story editor, a writer and an editor. Each step has a narrow job, which is what lets a small local model write a script you'd want to voice.

1. Research (real sources)

VidNexus reads the live web: Wikipedia plus Ollama web search, up to six sources per topic. It keeps a fact only when the quote it came from is in the source and every number in it appears there too. The outline screen shows a Sources line with links, and "N facts from M sources" while it works.

Web research needs a free Ollama account key. Get it at ollama.com/settings/keys and paste it in Settings > API keys as Ollama web search key. Setup asks for it first, and script writing stays off until it's set. If the web turns up too little, the writer falls back to the model's own knowledge and the outline screen says so in amber, so you know to check the facts.

2. Outline and story plan

The facts become a story plan (the driving question, the promise to the viewer, why it matters, and the answer saved for the end) and a section outline. Every section has a role chip such as Hook, Setup, Turn or Payoff. You can edit, reorder or regenerate the outline before any narration is written.

3. Sections

Each section is written on its own, as plain narration, with the story plan, the facts that belong to it, and a note of what earlier sections covered and what comes later. That keeps long scripts from repeating themselves or giving the ending away. The script review page fills in live as each section lands.

4. Editor's pass

A style check reads every scene for the tells of machine writing: stock phrases, signposting ("in this section"), long sentences, repeated figures, moral endings. Flagged scenes are rewritten without losing their facts, scenes that only repeat earlier lines are dropped, and very long scenes are split so the picture changes often enough.

Then the hook is scored and, if weak, rewritten, and the fact-check compares claims with the source quotes.

Longer videos

Length decides how many sections the outline has: about 6 for Standard, 14 for Long and 24 for Deep dive. Every section works from the same research and story plan, and the writer keeps a running word budget, so if early sections run long the later ones are asked for less. Deep dive scripts take longer on a local model.

Rule of thumb: Ollama running locally is free but slower; a hosted model is fast and costs about a cent per script.

Script review

Before any voice or footage work starts, you land on the script review page. Here you read and fix every scene before anything costs money.

When the script is right, click Continue to voice & visuals. That opens Video settings, where you check the voice, graphics, footage sources and the cost of the build. Nothing is voiced or searched until you click Approve & build there.

What you can do on script review

• Edit any scene's text.

• Click Edit the whole script to change everything in one box.

• Click Back to outline to change the outline and write the sections again.

• Score the opening again and check the facts again after your edits.

Hook scoring

Every generated script gets a hook score out of 30 on the script review page. It is made of three scores from 0 to 10: stakes, specificity and title promise. Next to it, a fact-check panel flags shaky claims and suggests rewrites you can apply in one click.

For a deeper check, open The Lab in Research. It scores your opening from 0 to 100 on how clearly the first minute says what the video is about. It also shows similar videos already on YouTube.

Want a stronger first line? Open Tools > Hook A/B generator, or the Hooks tab in the editor. It writes 8 opening lines, each with a voice preview. Pick one and it becomes the narration of the first scene.

Humanize pass

AI writing has tells: sentences that are all the same length, pet phrases (delve, tapestry, in this video, in the realm of), lots of hedging and a small vocabulary. Humanize in Research rewrites your narration to sound less like a model and more like a person.

What Humanize measures

• Reading ease (60 to 70 suits spoken narration)

• Average sentence length

• Sentence variance (higher sounds more human)

• Banned phrases (common AI phrases)

• Vocab diversity (how many different words you use)

• Repeated bigrams (two-word phrases used again and again)

• Hedging density (actually, essentially, arguably)

You see all seven numbers before and after the rewrite, so you can see exactly what changed. Pair it with voice matching (next section) for the best result.

Voice-matching your channel

Generic rewrites drift toward a generic voice. To match YOUR channel, add its YouTube handle and sync its videos. VidNexus saves your uploads' titles and descriptions on your computer.

Once synced, three tools follow your channel's style:

Voice-matched tools

• Titles. Writes titles that match your channel's rhythm, punctuation and bracket habits, using your 8 most viewed videos as examples.

• Humanize. Rewrites the script using the same 8 videos' titles and description openings, so your tone carries over.

• Channel brief Auto-assemble. Pulls your best past titles into the brief as examples.

To turn on voice matching:

1. Channels > Edit on the channel > paste your @handle in YouTube channel > Save.

2. Click Sync videos on the channel.

3. From then on, Titles and Humanize use it. VidNexus syncs the channel again every day while the app is open.

Channel brief

A brief is a short set of rules the AI follows in every new script on a channel. Once saved, you don't set style rules for each video again. Write it in Channels > Edit on the channel > Channel brief, or in Research > Channel brief.

A good brief contains

• Hook angle. How you like to open (a question, a number, or dropping into the middle of the action).

• Target audience. What your viewers already know.

• Style notes. Rhythm, contractions, banned phrases.

• Competitor patterns. What the top channels in your niche do that works.

• SEO keywords. Search terms every script should try to use.

• Facts to mention. Recurring themes and callbacks.

Click Auto-assemble in Research > Channel brief. It pulls in your saved keywords, your last competitor run and your synced channel history, so you rarely start from a blank form.

What's next

Voiceover: How the narration you approved becomes audio via ElevenLabs.

Motion graphics: How the AI picks graphics for each scene, and the full catalog.