Claude Skills, Flow Agent, and Gemini Omni: Ads Work Without Guesswork

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

Claude Skills, Google Flow Agent, and Gemini Omni finally made ads work feel less like guessing which prompt mood might sell. Flow as a creative studio with Omni underneath, Flow TV showing films with their prompts, an agent that plans projects conversationally, character locks with voices, custom tools like Scene Explorer — the stack can run a sixty-second film from a handful of clips. I still bring a claim sheet and a human taste pass, because consistent panthers and product shots are not the same as truthful ads.

Flow Agent as creative director, humans as editors-in-chief

Conversational planning that removes the need for prompt craft is perfect for speed and terrible for accountability if nobody owns claims. I let Flow Agent propose structure, shot list, and variants. I lock claims, offers, and legal lines before generation. The agent can be the director. It cannot be the advertiser of record.

Flow TV's library of films with prompts is a learning tool I actually use — study, then adapt, never paste blindly into a client brand.

Skills for the boring consistency

Claude Skills hold brand voice, forbidden phrases, and the lighting/backdrop specs that make product photography prompts repeatable. Pro results in the notes came from explicit lighting and backdrop detail, not from mystical adjectives. I encode those details once.

Text-change prompts that fail teach the same lesson: be specific, then verify every character of on-image text before media buy.

Character locks across a travel series

Saved characters with voices keep faces consistent across locations — essential for personal brand libraries and virtual try-on adjacent work. I version the character pack and treat drift as a failed take even when the frame is pretty.

Consent and labelling for synthetic or altered likenesses stay in the runbook. Consistency without consent is still a problem.

Ten use cases, one rights appendix

Blend photos, room redesign, personal-brand library, ads, virtual try-on, restore old photos — the list is long. Each use case inherits the same rights appendix template: source ownership, model releases, music licences, claim sheet. Custom tools like Scene Explorer become one-click only after the appendix is attached to the project.

One-click without rights is how brands buy lawsuits at render speed.

Cartier-panther energy with humility

A sixty-second film from six clips with 2K upscale is a great craft demo. Client ads still need offer truth, landing page alignment, and platform policy checks. I show the craft, then I show the checklist, in that order, every time.

Craft-first meetings that skip the checklist create campaigns that cannot survive legal review.

Omni as multimodal muscle

Gemini Omni underneath Flow is why the studio feels unified. I still arena-test critical shots against other image and video tools when the brand risk is high. Unified UX is not the same as universal quality on every prompt.

Switching costs inside Flow are low; switching costs after a media buy are high. Arena before buy.

What I sell

Skills packs for brand ads, Flow Agent pilots with claim locks, character libraries with consent trails, and human taste passes on a retainer. Guessing is free. Shipping ads that survive scrutiny is the product.

Final gate

Claims locked. Characters consented. Text verified. Rights filed. Human watched the cut. Then publish. Skills and Flow Agent can fill the middle. They do not get the last click without that gate.

Claim sheet in the project root

FLOW or Claude Skills runs refuse to start if CLAIMS.md is missing. Gemini Omni can render the panther; it cannot invent the offer. Editors-in-chief own the file. Guesswork dies at the door.

Human taste pass retainer

A scheduled human watch-through on every cut before media spend. Flow Agent and Gemini Omni can fill the timeline. Taste and truth stay on a retainer line item so they cannot be "optimised away" after the first good panther.

On-image text gate

No media buy if any frame's text was not read aloud by a human. Flow Agent speed is irrelevant when a digit is wrong in the price. Skills can remind; humans still speak the digits.

Character pack versioning

Character locks live in a versioned folder with consent files beside them. Drift fails the take. Flow TV inspiration is allowed; unversioned faces are not. Gemini Omni renders; git remembers.