Writing a better prompt

Start with the story, not the scenes

Your first prompt should describe what the video is about -- not how each scene should look. Think of it like briefing a creative partner: give them the big picture and let them bring it to life.

This works well:

"Create a 15-second sales video for Zoho CRM. Audience: sales managers frustrated with scattered spreadsheets. Show how Zoho gives teams one place to track every deal. Tone: confident and human, like a colleague -- not an ad. End with a CTA to book a demo."

This leads to frustration:

"Scene 1: show a blue background with our logo centered. Scene 2: show a person at a desk looking stressed. Scene 3: zoom into a laptop screen..."

When you try to direct every scene upfront, the system has no room to do what it does best. You'll end up in a loop of "that's not what I meant" corrections. Instead, let it generate a first draft, then fine-tune the specific scenes that need work.

Tell ngram who this video is for

One sentence about your audience changes everything. "For investors" produces a very different video than "for existing customers" or "for teens on TikTok."

Examples:

• "Audience: retail buyers evaluating new cosmetic brands to stock"

• "This is for nonprofit fundraisers in Belgium"

• "Aimed at parents managing youth sports schedules"

Paste your script (even a rough one)

This is the single most effective thing you can do. Users who include a script or outline in their first message get to a finished video much faster than those who don't.

It doesn't need to be polished. Even bullet points work:

"Key points:

• Hook: The threat landscape evolved. Your systems didn't.

• Problem: Teams are running disconnected tools that don't talk to each other.

• Solution: TalosConnect gives you a single view across all your systems.

• CTA: Book a demo today."

If you have a full script with timing, even better -- but a rough outline is already a huge improvement over "make a video about my product."

When you upload images, say what they're for

Uploading screenshots, logos, or reference images is great. But the system needs to know what to do with them.

Do this:

"I've attached 3 screenshots -- use the dashboard view for the product demo section, the settings page for the feature walkthrough, and our logo should appear in the intro and closing."

Not this:

[uploads 5 images with no explanation]

If you just drop files in without context, there's a good chance they won't be used the way you intended -- and you'll spend the next 10 messages asking "why didn't you use my images?"

Set the tone

A short note on tone gives the system a direction for the voiceover, music, and pacing. One line is enough:

• "Tone: calm and authoritative"

• "Make it upbeat and punchy"

• "Professional but warm -- like you're explaining it to a smart friend"

• "Confident, not salesy"

After the first draft: edit surgically

Once you have a first draft, you'll probably want to change a few things. That's normal. Here's how to do it effectively:

Be specific about which scene:

"In scene 3, replace the generic chart with the screenshot I uploaded"

Not vague:

"The visuals are wrong, fix them"

If something isn't right, tell the system exactly what's wrong and what you want instead. "Try again" without changing your instruction will usually get you the same result.

The quick checklist

Before you hit send on your first message, check if you've included:

• What the video is about (one sentence)

• Who it's for (your audience)

• What to say (script, outline, or key talking points)

• What you've uploaded and where each thing should go

• How it should feel (tone, energy, style)

• Any hard requirements (duration, language, aspect ratio)

You don't need all of these every time. But the more you include, the faster you'll get to a video you're happy with.

Good prompts: 5 examples

1. SaaS product explainer

"Create a 60-second explainer video for our project management tool, TaskFlow. Audience: team leads at mid-size companies who are tired of juggling Slack, email, and spreadsheets to track work. Key message: TaskFlow replaces the chaos with one place where your team plans, tracks, and ships -- without learning a new system. Tone: clear and direct, not corporate. End with a CTA to start a free trial."

Why it works: clear product, specific audience with a specific pain point, message in one sentence, tone direction, and a defined CTA.

2. Startup investor pitch

"We need a 90-second pitch video for Greenhouse Robotics. We build autonomous robots that handle weeding and soil monitoring for mid-size vegetable farms. Our audience is seed-stage investors who've seen a hundred agtech pitches -- so skip the 'farming is broken' cliche and lead with our traction: 12 pilot farms, 40% labor cost reduction, $800K ARR. Tone: confident and grounded. Close with 'Join our $3M round -- let's scale.'"

Why it works: tells ngram what NOT to do (skip the cliche), leads with specifics instead of generalities, and gives real numbers to work with.

3. Course launch promo with outline

"Make a 30-second promo video for an online photography course called 'Light & Story.' Audience: hobbyist photographers who have a good camera but feel stuck shooting the same things.

Outline:

• Hook: You have the gear. Now learn how to see.

• Problem: Most photography courses teach settings. This one teaches you how to find the story in any scene.

• Proof: 2,000+ students, 4.9 star rating

• CTA: Enroll now, first module is free.

Tone: warm and inspiring, not lecture-y. Think campfire, not classroom."

Why it works: provides a full outline with each section defined, gives a metaphor for tone that's easy to follow.

4. Feature announcement with screenshots

"Create a 15-second feature announcement for our design tool's new 'Auto Layout' feature. Audience: existing users who still arrange elements manually. I've attached 2 screenshots -- the first shows the old manual workflow (use this as the 'before'), the second shows Auto Layout in action (use this as the 'after'). Keep it simple: show the pain, show the fix, done. No voiceover -- just text overlays and upbeat background music."

Why it works: explains exactly what each attachment is for, specifies format preferences (text overlays, no voiceover), and keeps the ask focused.

5. Local business promo

"I run a family-owned bakery called Mila's in Portland. I need a 20-second Instagram Reel to promote our weekend brunch menu. Audience: local foodies aged 25-40. Key things to mention: house-made sourdough, seasonal pastries, coffee roasted in-house, and the cozy atmosphere. I've attached our logo and 4 food photos -- use them throughout. Tone: warm, inviting, a little playful. End with our address and 'Walk in this Saturday.'"

Why it works: gives the business context, a specific platform (Instagram Reel implies vertical format and short length), clear items to feature, attachment instructions, and a local CTA.

How NOT to write a prompt: 3 examples

1. The empty ask

"Make me a video."

What's it about? Who's it for? What should it say? ngram will try its best, but it's working with nothing. You'll spend 10 rounds correcting a first draft that was always going to miss, because there was nothing to hit.

2. The scene-by-scene screenplay

"Scene 1: black background, logo fades in center, hold for 2 seconds. Scene 2: split screen, left side shows a stressed office worker in blue lighting, right side shows a green forest. Text appears letter by letter: 'There is a better way.' Scene 3: camera zooms into a laptop showing our dashboard. The cursor clicks on the 'Reports' tab. Scene 4: ..."

This feels thorough, but it fights the system. You're describing exact camera moves and compositions before ngram has even understood what your video is about. The result: constant corrections because every scene is slightly different from what you pictured. Instead, describe the message and let ngram handle the visual storytelling -- then adjust specific scenes after.

3. The file dump

[uploads 8 screenshots, a logo, and a PDF] "Use these."

Use them for what? Which screenshot goes where? Is the PDF the script? Are some of these reference images or actual assets to include? Without context, ngram has to guess -- and it will guess wrong. Always tell it what each file is and where you want it used.