AI capability films · music-cut · landscape 16:9
How to make a video like the Grok 4.5 ad
The Grok 4.5 ad demonstrates an AI product by typing at it. Each capability beat opens with a prompt being typed into a bar, and the scene ends when the prompt finishes — so the film is content-led rather than locked to a beat grid. Between beats, plates collapse to a line and reopen, and the device stage grows continuously across the film before a brand wordmark wipes over a photographic sky. baz ships this as the Grok 4.5 Ad skill: about 56 seconds across 17 scenes.
The shortcut: run a skill instead of starting from scratch
A skill is a context file that teaches an AI agent one specific video style — its structure, pacing, typography, motion grammar and sound design. Hand an agent the skill plus your product, and it rebuilds that video for you. The steps below are what the Grok 4.5 Ad skill already encodes, so you can follow them by hand or let the skill carry them.
baz is a library of video skills: context files that teach an AI agent a specific motion-graphics style, so it can rebuild that video for your product, with your brand, from your inputs. Run a skill inside your own coding agent — Claude Code, Codex — or directly in the baz editor.
What you need before you start
Required
- Product screenshots
- Your feature list — one capability per beat
Makes it materially better
- The demo prompts themselves — the actual sentences you want typed on screen
- Your product URL and brand assets
- A tagline for the closing wordmark
How to make a video like the Grok 4.5 ad: step by step
Write the prompts first — they are the script
In this format the typed prompts carry the argument. Write the four or five you want on screen before anything else, and make them things a real user would ask. A vague prompt produces a vague beat.
Let content set the timing, not a metronome
This film is content-led: a scene ends when its prompt finishes typing. That is why it feels responsive rather than mechanical. Do not force the beats onto a fixed grid — let a longer prompt take longer.
Add capabilities by adding beats, not by stretching
If you need a longer film, add another capability beat. If you need a shorter one, remove a beat. Stretching the existing scenes to fill time is the one adjustment that visibly breaks the format.
Collapse plates to a line between beats
The signature transition squashes the outgoing plate toward a line and reopens the next from it. Note that not every collapse is a full hairline — demo-to-demo swaps use a partial squash, and treating them all as hairlines makes those cuts far harder than they need to be.
Grow the device stage across the whole film
The stage the product sits on gets larger as the film progresses. It is a slow, cumulative move that gives the piece a sense of building without any single scene doing anything dramatic.
Close with a wipe over something real
The wordmark wipes over a photographic sky rather than a flat colour. That contrast — synthetic UI against a real photograph — is what stops the ending feeling like a slide.
What the finished video actually contains
Approximate timings for a ~56 seconds cut. Treat them as a shape to hit, not a spec — the exact durations shift with your script and your music.
| Time | Beat | What is on screen |
|---|---|---|
| Open | Hook and prompt transition | The opening claim, then the first prompt bar arriving. |
| Establish | Device rise, first output | The device stage rises and the first generated output resolves. |
| Per capability | Capability demo card | A prompt types, the output appears, the plate collapses to a line into the next. |
| Midpoint | Act break and overview | A capability overview and hero-card focus reset the film before the second half. |
| Build | Prompt pills and grid pan | Prompt pills are born and a capability grid pans, compressing many uses into a short beat. |
| Close | Final prompt, brand close | One last typed prompt, then the wordmark wipes over a photographic sky. |
Mistakes that cost you the view
Easing in and out across a seam
Do this instead: Accelerate out, decelerate in. Symmetric easing creates a velocity dip exactly at the cut, which reads as a stutter.
Resizing the demo card box
Do this instead: Scale it instead. Resizing the box itself makes small type shimmer on every frame.
Treating every plate collapse as a hairline
Do this instead: Use a partial squash for demo-to-demo swaps. Modelling them all as full hairlines makes those cuts roughly three times harder to land than intended.
Stretching scenes to hit a target runtime
Do this instead: Add or remove capability beats. Stretching breaks the content-led timing that makes the film feel responsive.
Baking in image URLs you have not verified
Do this instead: Confirm every asset resolves before publishing. A broken plate in the middle of a capability beat is unrecoverable at render time.
Run it from your own agent
Skills are designed to run inside a terminal agent like Claude Code or Codex, against a project you control. The same skill produces the same style whether it runs in your agent or in the baz editor.
npm install -g bazaar.it@latest
# Install this skill into your agent (Claude Code, Codex or Cursor)
baz skill add grok-4-5-ad
baz project create --name "Grok 4.5 Ad for my product"
baz prompt "make a video like the Grok 4.5 ad" --stream-json
baz review --summary --json
baz export start --wait --format mp4 --jsonOperations draw from a prepaid balance. See the CLI reference for the full command set.
Common questions
What if my product does not have a prompt bar?
Then this is probably the wrong skill — the typed prompt is the structural spine, not a decoration. Look at the Supahub or Outbidd grammars, which are built around UI interaction rather than a text input.
How many capabilities should I show?
Four or five. The source runs 17 scenes across roughly 56 seconds, and most of those are transition and framing beats. Each additional capability adds real runtime, so pick the ones that differentiate you.
Do I need video editing experience?
No. The skill carries the editing decisions — shot order, pacing, type treatment, transitions, sound design. You supply the raw material and the facts about your product; the agent assembles the cut. You review it and ask for changes in plain language.
Can I run this in my own coding agent instead of the baz editor?
Yes. Skills are designed to be run from a terminal agent like Claude Code, Codex or Cursor. Install the CLI with `npm install -g bazaar.it@latest`, then `baz skill add <slug>` to pull the skill into your agent. The same skill produces the same style whether it runs in your agent or in the baz editor.
What does it cost?
baz bills against a prepaid dollar balance rather than a seat subscription — you top the balance up and each operation (generation, voiceover, render, export) draws a listed amount from it. There is no per-seat fee to make a video, and the price of each operation is published.
Can I change the video after it renders?
Yes, and this is the point of doing it as code rather than as a flat render. Every scene is a real component you can edit — ask the agent to re-time a beat, swap a screenshot, rewrite a line, or change the accent colour, and only that part rebuilds. You are not re-rendering from scratch each time.
Will every video from this skill look the same?
The grammar is the same — that is what makes it a skill and not a template. The content is not: your brand colours, typography, screenshots, copy and data all come from your inputs. Reusing one skill across a series is a feature, because episode three lands in the same visual language as episode one instead of being rebuilt from scratch.
Related guides
Pick a skill. Get the video.
Run the Grok 4.5 Ad skill in your own coding agent, or in the baz editor. Usage comes out of a balance you top up, with a listed price per operation — no seat subscription to make a video.