Make One Image or Narrated Clip, and Label It

T09-L01 · Create media · Level 1 User · 13-minute read, plus 30 minutes of doing

One slide, due tomorrow

You need one illustration for a slide by tomorrow morning, and there is no designer and no budget. You type a request into an image generator. Thirty seconds later you have something that looks professional. Then you look closer.

In a lab it's a postdoc preparing a group-meeting talk. She asked for a schematic of her assay: sample, lysis, ELISA, plate reader. The picture is beautiful. It has five steps instead of four, one label reads "ELISSA", and the arrow between the last two steps points backwards.

In a company it's a marketing assistant who needs a header image for the customer newsletter and a one-minute narrated clip for the website. The image is fine. The voice sounds good too, until she reads the small print of the free plan she used: no commercial use.

A generator is very good at how something looks and unreliable at exactly what it says. And the tool that made the file doesn't decide whether you may use it. This book shows you how to split the work so each part goes to the tool that does it well, and how to label the result. You are at Level 1 User: the asset is for your own talk or your own draft, so a failed attempt costs you only time.

After this you can

• Pick the right kind of tool for an image, a voice-over, a short video or a transcript.
• Write an image prompt with a subject, a style, a composition and exclusions, and fix a weak result by changing one element.
• Keep exact text, labels and structure in an editor instead of asking the generator to draw them.
• Check whether your plan allows your use, and add an AI label, alt text or captions to the finished file.

Before you start

• Books: T01-L01 for the four-quadrant method this book applies to one media asset.
• An image generator. This book uses Gemini with a free Google account. Its images are good, it handles styles well, and every image carries a watermark that tells people AI made it. If your organisation provides ChatGPT or Adobe Firefly, use that instead.
• Clipchamp, Microsoft's free video editor, in the browser at app.clipchamp.com with a personal Microsoft account. It turns a script into a voice-over, adds captions and exports the clip, all in one place.
• Your slide program: PowerPoint, Keynote or Google Slides.
• About 45 minutes: 13 to read, about 30 to do.

Last verified 23 September 2026. The Gemini app makes images with Nano Banana 2 on every plan, including the free one. Paid Google AI plans can also use Nano Banana Pro. Gemini adds a visible watermark and an invisible SynthID watermark to its images. Clipchamp's free plan exports up to 1080p without a watermark. Its text to speech is free for all users (under Record & create → Text to speech), with about 400 voices in over 80 languages. Its autocaptions (the Captions tab, then Transcribe media) can be downloaded as an .srt file. ElevenLabs' free plan gives 10,000 credits a month, but it carries no commercial licence, and anything you publish from it must credit "elevenlabs.io" in the title. Its Starter plan ($6 a month) includes a commercial licence. Adobe says outputs from its own Firefly models are safe for commercial use, and Firefly has a free tier with daily generations. In the EU, the AI Act's transparency rules have applied since 2 August 2026: deepfakes and AI-generated text published to inform the public must be clearly labelled.

The idea in one page

You can't read a picture or a voice-over the way you read a paragraph, so errors hide in plain sight. Four facts decide how you work.

1. Generators are good at the look and bad at the exact. They do style, mood, texture and composition well. They are unreliable at exact words, counts, arrows, numbers and anything that must be scientifically or legally correct. What you do: let the generator make the look, and put every exact thing (labels, steps, figures, your logo) on top in an editor you control.

2. An image prompt is a short brief, not a wish. A good one has four parts: the subject (what's in it), the style (flat illustration, watercolour, photo), the composition (wide, empty space on the left for a title) and the exclusions (no text, no people, no logos). What you do: when a result is wrong, change one of the four parts. Regenerating the same prompt fifty times is a slot machine, not a method.

3. The tool doesn't give you the rights. Three separate questions: may you use what you uploaded (a photo from the web usually no)? Does your plan allow your use (free plans often exclude commercial use)? And do you have a person's consent for their face or voice? Cloning a colleague's voice without written consent is off limits, however easy the button makes it.

4. Label it by default. Say what AI made, in one line, where the audience will see it: "Illustration: generated with Gemini; labels and layout by the author." It costs ten seconds, and in the EU it's the law for deepfakes and for AI text published to inform the public. Add the other things people need to use the asset: alt text for an image, captions for a clip.

Which tool for which job:

Worked example: one slide, one clip, four quadrants

The method comes from T01-L01. Here it is on a lab talk, then on a company newsletter.

Lab: a schematic and a 60-second explainer

Structure it. "For Thursday's group meeting I need one slide showing my assay in four steps: sample, lysis, ELISA, plate reader. I'd also like a 60-second narrated version for new students joining the lab. A useful result: correct, readable in two seconds, and nobody mistakes it for real data."

Understand it. She asks Gemini for the whole schematic with labels. The first image is lovely and wrong: five stations, "ELISSA", one arrow reversed. Two rewordings later the spelling is fixed, the arrow is still reversed and a sixth vessel has appeared. That's fact 1: the generator draws what a lab schematic looks like, not her assay. Back to Structure it: "Four identical-style icons from the generator. Labels, arrows and order are mine, in PowerPoint."

Choose tools. Going down the ladder: the group's slide template already has the colours and fonts. PowerPoint handles the boxes, arrows and labels. Gemini makes the four icons in one style. Clipchamp turns the slide and a short script into the narrated clip, with captions. She needs no video generator: nothing in her story moves.

Build it. She asks Gemini for one icon at a time with the same prompt frame:

Subject: a single microcentrifuge tube, closed.
Style: flat vector icon, two colours (dark teal and light grey), thick clean outlines.
Composition: centred on a plain white background, square, lots of empty space.
Exclude: text, letters, numbers, labels, hands, people, logos, shadows.

Only the subject line changes for the other three icons: a tube with a lysis buffer drop, a 96-well plate, a plate reader. The third icon has a faint number printed on the plate. She doesn't regenerate everything. She changes one element, adding "no markings on the plate" to the exclusions, and the next one is clean. She drops the four icons onto her slide, types the labels and draws three arrows. The structure is right because she typed it.

For the explainer she writes a 120-word script, one sentence per step, and reads it aloud once to catch awkward phrasing. In Clipchamp she imports the exported slide, adds the script under Record & create → Text to speech, picks a clear, neutral voice and slows the pace a little. She runs Transcribe media on the Captions tab and corrects one error ("lice is" for "lysis"). Along the bottom of the slide she adds the label: "Icons generated with Gemini. Steps, labels and narration script by the author." She exports at 1080p and watches the clip once with sound and once muted.

What it taught her: the generator saved her an hour of icon hunting. It would have cost her credibility if she'd used its schematic. Next time she'll reuse the same prompt frame so all her slides match (T09-L02).

Company: a newsletter header and a product clip

Structure it. "The monthly customer newsletter needs a header image for the autumn issue, and the website needs a 60-second narrated clip introducing our new sample-storage box. A useful result: on brand, correct, and ours to use commercially."

Understand it. She makes a voice-over in ElevenLabs' free plan. It sounds great. Then she checks the licence page: the free plan carries no commercial licence, and anything published must credit ElevenLabs in the title. The website clip is commercial. She also sees that the header needs the newsletter name in it, and the generated lettering is slightly off. Back to Structure it: "A header image without text, the name added in our template, a voice we may use commercially, and captions, because many people watch muted."

Choose tools. Going down the ladder: the newsletter template already has the logo and font, so the name goes there. Gemini makes the header background. For the voice she has two options. She can use Clipchamp's free voice, checking its terms for commercial use first, or she can pay for ElevenLabs Starter, which includes a commercial licence. She picks Clipchamp because the clip lives there anyway, and asks her manager whether the company wants a paid voice later. The product shots come from the company's own photos, not from a generator: a customer should see the real box.

Build it. Her header prompt: subject "autumn leaves and a few laboratory sample boxes on a wooden shelf"; style "soft watercolour, warm orange and the company's blue"; composition "very wide banner, empty calm area on the left third for the title"; exclude "text, letters, logos, people, brand names". She places the image in the template and types the newsletter name over the empty area. For the clip she combines three real product photos with a Clipchamp voice-over of her approved script, corrects the autocaptions and adds the line "Narration: AI voice (Clipchamp). Product photos: our own." The header caption says "Illustration generated with AI."

What it taught her: the licence question took five minutes and prevented a real problem. Her next trip round the quadrants is a short style sheet (colours, prompt frame, label wording) so next month's issue matches this one.

In both cases the generator did one job: the look. Everything exact stayed under human control.

What goes wrong

Asking the generator for exact text. Symptom: labels that are almost right ("ELISSA", "Recieve"), or text that turns into gibberish at small sizes. Fix: add "no text, no letters" to the exclusions and type every word yourself in the editor.

Regenerating instead of deciding. Symptom: your twentieth attempt at the same prompt, each one differently wrong. Fix: name what's wrong, change the one part of the brief (subject, style, composition or exclusions) that controls it, and try again. After three failed tries, do that part by hand.

Using an input you have no rights to. Symptom: you uploaded a photo from a supplier's website as a "reference" and the result looks just like it. Fix: upload only your own material or material licensed for this use. When unsure, describe the style in words instead.

Assuming the plan covers your use. Symptom: a free-plan voice-over on the company website. Fix: before anything commercial leaves your desk, open the tool's pricing or licence page and look for the words "commercial licence". Write down what you found.

Cloning a voice or face without consent. Symptom: "Let's have it sound like the boss." Fix: use a stock voice that belongs to no real person, or record the real person with their written consent for this exact use. A different name on a recognisable voice doesn't solve anything.

No label. Symptom: a colleague asks, weeks later, "Is that a real photo of our lab?" Fix: put a one-line AI label on the asset itself, not only in the file name. It should say what was generated and what you did.

Trusting the preview. Symptom: captions that cover the labels, or a clip that cuts the last word. Fix: check the exported file, not the editor: once with sound, once muted, at the size people will see it.

Do it yourself (30 minutes)

Use a real need from your own week: a slide, a header, a short explainer.

1. Structure it (5 min). Write one sentence: what the asset is for, who sees it, and what a useful result looks like. List the exact things it must get right (words, steps, numbers).
2. Understand it (8 min). Try the generator once on the whole thing. Note what it got right (the look) and what it got wrong (the exact parts). Check whether your plan allows your use. Rewrite your sentence if you learned something.
3. Choose tools (2 min). Decide which part the generator makes and which part you make in your editor. Pick an image, or a narrated clip, not both.
4. Build it (15 min). Write the four-part prompt, fix one weak result by changing one element, assemble the asset in your editor, add the AI label and the alt text or captions, and export. Check the exported file once as your audience will see it.

Exit check

Delivered artifact: one finished image or clip, with:

• its AI label, visible on the asset;
• its alt text (for an image) or corrected captions (for a clip);
• the prompt that produced the generated part;
• one line on rights: where each input came from, and whether your plan allows your use.

It passes when a colleague looks at the exported file and finds no wrong word, no missing label and no input you lack the rights to use.

Rule to remember

Make it, then say you made it.

Further reading and tools

Next books

• T09-L02 · Consistent assets for your company or lab: one style across many slides and images.
• T01-L02 · Trust but verify: checking output before it leaves your desk.
• T12-L02 · Privacy and labelling in everyday work: what to label, and what must never go into a generator.
• T16-L01 · Choose a tool for your task: comparing two tools on the same job before you commit.

Tools

• Gemini: Taught. The free plan is enough for this book.
• Adobe Firefly: Catalogued. Pick it when commercial use has to be clear.
• ElevenLabs: Catalogued. More natural voices; the commercial licence starts with the Starter plan.
• Veo, Sora: Catalogued. Generated video, for when nothing real can be filmed.
• Whisper: Catalogued. Transcription on your own machine.
• C2PA: Catalogued. The standard behind content credentials.