A generated clip does not make a finished video. What makes AI video useful is everything around it: a script you can cut, narration you control, visuals that support the claim, and captions people can read on a phone.

This workflow is for one person or a small team: a freelancer, a content creator, a shop owner, a two-person agency. Four tools appear as examples, and you do not need all four.

1. Write the script

Start with the script, not the visuals. A general assistant such as ChatGPT can turn a description of your business into a structured short script, as long as you give it the structure rather than a topic.

Write a 45-second promotional video script for a small bakery that sells sourdough bread and runs weekend baking classes.

Structure: one hook line, three short value points, one call to action.

Constraints: plain spoken English, no statistics, no claims about health benefits, maximum 110 words of narration. Add a column for the visual that supports each line.

Ask for two versions at different lengths, then choose one and edit it: you want a script you can cut, not one you have to repair.

Two things need a human eye. Facts, because an assistant will happily invent a delivery radius, an opening time or a customer quote. And tone, because the default voice is often more promotional than a one-person business sounds. Read it aloud before you record.

2. Create the narration

Text-to-speech is good enough now for explainers, tutorials, social clips and local advertising. ElevenLabs is one service for it, with a free tier that has character limits and paid plans with more.

Generate slightly slower than feels natural, since pacing is easier to tighten than to stretch. Keep one voice across a series. Check pronunciation of names and places. Generate the whole script in one pass where possible, so the tone stays even.

Settle two things before publishing. Voice consent: cloning a voice, your own or a client's, means confirming you have the right to do it, and ElevenLabs' cloning flow asks for that confirmation. Commercial use: check what your plan permits, especially for client work.

3. Create or select visuals

There are four sources, and they are not interchangeable: AI-generated clips for abstract or conceptual sequences where no footage exists; your own footage for anything that must show the real product, premises or people; stock media for establishing shots and B-roll, under the licence you hold; and templates when the deliverable has to match a brand system quickly.

Runway is an example of an AI video platform: it generates and edits video, images and audio in one workspace, with a free trial and credit-based paid plans. It fits the shot you cannot film.

Two limits to plan around. Generated clips usually need trimming and colour matching before they sit together, and continuity drifts, so the same character or room can change between generations. Plan short cuts rather than long takes, and keep shots of the real product real.

4. Edit, caption and publish

Editing is where the video becomes watchable, and transcript-based editing is the biggest time saver in the workflow. Descript transcribes your recording and lets you cut the video by deleting words, with filler words and long pauses removable in one pass.

Captions are not optional, because most social video is watched with the sound off, and auto-captions are a starting point rather than a finished file. Read them: product names, numbers and accents are exactly where they fail. Check line length, keep captions inside the safe area on vertical formats, and use your own font if you have a brand. Then watch the whole thing once with sound and once without, on a phone.

5. A worked example

A freelancer is making a 45-second promo for a two-person landscaping business.

  1. Script. They ask ChatGPT for a 110-word script with a hook, three services and a call to action, from the owner's description of recent jobs. They rewrite one line that promises a response time the owner cannot guarantee, then read it aloud.
  2. Narration. They generate the voiceover in ElevenLabs, slow it slightly, and regenerate one sentence where the town name is mispronounced.
  3. Visuals. They use the owner's phone footage for the finished garden and the team at work, plus one short generated clip as an opening transition. No stock: the business is local.
  4. Edit and captions. They cut to the transcript in Descript, remove filler words, correct the captions, add the logo at the end, and export vertical and landscape versions.

What they had to review: the response-time claim, the services list, the pronunciation, the licence on the generated clip, and every caption. No tool checked that the video tells the truth about the business.

6. Do you need all four tools?

No. General assistants script and some now generate images or audio; editors with generative features, like Descript, cover narration and clips; video platforms, like Runway, include editing and audio tools; and your phone may already handle captions.

Start with what you have, and add a specialist tool when one step costs you hours, not because a workflow diagram said so. Buying four subscriptions before making one video is the most common way this workflow stalls.

7. Limitations to plan for

  • Accuracy. Nothing in this chain checks facts. Every number, date, price and quote is your responsibility.
  • Copyright and licensing. Generated and stock assets carry licence terms. Read them, and keep records for clients.
  • Voice consent. Cloning someone's voice without documented permission is a legal and ethical problem, not a technical one.
  • Visual consistency. Generated scenes drift, so plan to hide cuts.
  • Privacy. Do not upload footage of clients, customers or children to a service you have not checked.
  • Human review. Every stage produces something plausible, and plausible is not correct.

Your first-video checklist

  • One audience, one message, one call to action
  • Script read aloud, facts checked, timing confirmed
  • Narrator voice chosen and pronunciation fixed
  • Rights checked for every asset and voice
  • Captions read and corrected
  • Watched on a phone, with sound and without

This guide is editorial. It does not report firsthand product testing, and no vendor reviewed it.