Creative Director . Agency OS

What it does

Produces the creative direction behind a visual: a structured JSON prompt built from a reference image you upload or from a written brief, a plain-text director's brief for quick exploration, or a video scene brief for a single generated clip. Every prompt runs through the Visual Elevation Protocol, a five-principle gate whose job is to stop you shipping the stock-photo version of your idea, and a kill-switch question that fails any visual fully describable as a person doing an activity at a location. It also carries hard rules on fabricated data, text rendering and platform safe zones, so what comes back is usable rather than just pretty.

Say this to start

This skill has no button. You start it by saying what you want. Any of these will do it:

/create-visuals

> build a JSON prompt from this

> recreate this image

> prompt from reference

> give me a JSON prompt

> ad variants as JSON

When to reach for it

a hero image or ad creative that needs to stop the scroll
reverse-engineering a reference image into a reusable prompt
a set of ad variants from one brief
a five-frame brand campaign from a single product photo
a single-scene video brief
a character or product kept consistent across a series

When NOT to use it

If you actually want	Use this instead
shot-by-shot video sequences with effects timelines and B-roll planning	`video-prompt-builder`
AI-avatar talking-head video production	`heygen`
the headline or body copy that sits on an ad visual	`meta-ads-copywriter`
the written copy around a visual outside of ads	`legendary-copywriters`
designing an actual HTML page or funnel section	`frontend-design`
cutting, captioning or clipping footage that already exists	`video-repurposer`

Before you start

What you need	Why
A brief, or a reference image if you want it reverse-engineered	the decision tree routes on which of the two you have, and the output format differs	Required
The workspace tool-stack file at `protocol/current-tool-stack.md`	it names the image and video tools of record, which decides whether you get a structured JSON prompt or a single flat prompt string	Required
`voice-profile.md` and `positioning.md` in the active workspace	colour preference, energy and aesthetic come from voice; market tier from positioning changes the visual treatment substantially	Required
`epistemology.md` in the active workspace	it holds what the brand rejects, which is what keeps the visuals off the category's default look. Without it you get a one-line warning and visuals built from aesthetic cues alone, which drift towards cliché	Optional
`policy.md` in the active workspace	any never-touch topic is excluded from prompts and references	Optional
A media generation tool connected to your workspace	with one connected you get actual generated images and video. Without one you still get complete briefs and prompts to paste into whatever you use	Optional
The exact real data for anything that must appear on the visual: phone, email, URL, address, price	it will never invent these. Anything you have not supplied comes back as a placeholder plus a production note telling you to add it yourself	Optional
The client's own brand files when you are working in a `client--` workspace	the client's identity applies there, not your agency's	Optional

How it runs

It loads the workspace and the tools of recordVoice profile for personality and colour, positioning for market tier, epistemology for what the brand rejects, policy for excluded topics, past feedback for your visual preferences. It reads the tool-stack file to know which image and video tools it is directing for, and mentions any adaptation it made from your logged preferences.
It answers three questions before writing anythingWhat emotion should the viewer feel, what action should they take, and what makes this memorable. These drive the colour temperature, lighting and composition choices that follow.
It routes to a formatA reference image you want recreated, or a brief for ad variants and structured single-image prompts, goes to the JSON schema. A quick in-conversation exploration goes to the plain-text director's brief. A video request goes to the video scene brief. It picks one and does not output both.
Visual Elevation ProtocolAt least one of five principles is applied to every prompt, two or more for thumbnails, hero images and ad creatives: metaphor over literal, visual tension through juxtaposition, scale disruption, emotional extremes, and one element that breaks the rules. Text-on-solid designs, tutorial screens, literal product demos and simple data graphics are exempt.
The kill switchBefore finalising, it asks whether the visual can be fully described as a person doing an activity at a location with nothing memorable left over. If yes, the prompt is reworked. A face in frame is fine; a face and nothing else is not.
Safe zones and text rulesAny social-platform or vertical output gets safe zones applied so platform UI does not cover the subject or the text. Rendered text is kept to 2 to 5 words, multi-digit numbers are avoided because generators garble them, and anything that must be letter-perfect is marked as a post-production overlay.
Anti-patterns passIt checks the output against the six tells of synthetic imagery: dead-centre symmetry, plastic overlit skin, stock-photo energy, floating or illegible text, uniformly oversaturated colour, and melted background detail.
Output and feedbackJSON mode returns analysis, the JSON prompt, and tweaks. Director's brief mode returns the prompt, the creative rationale, technical specs, two to three variations to test, and production notes. Generated assets are logged to your workspace assets file and it asks how the visuals landed.

What you get

A structured JSON prompt in three sections, analysis, the prompt itself, and tweaks, ready for any tool that accepts structured input
Or a five-part director's brief: the paste-ready prompt, creative rationale, technical specs, 2 to 3 variations to test, and production notes
A video scene brief covering scene, subject, environment, camera, lighting, audio, pacing and mood
Production notes listing exactly which text you must add yourself in post, and where in the frame it should sit
An entry appended to brand/[workspace]/assets.md, plus a logged learning on how the visuals landed

Honest limits

Read this before you rely on it

It directs, it does not paint. Generation happens in whatever tool your workspace has on record, and without one connected you get briefs and prompts rather than finished assets.
It will never invent a phone number, email, URL, street address, price or stat. Those come back as placeholders with a production note, because a hallucinated phone number on a published ad is a real cost.
Image generators are unreliable at rendering text, especially multi-digit numbers. Every prompt containing text carries a verify note, and you have to actually check every character before publishing.
Video clips top out at roughly 8 seconds on current models, so anything longer is chained through the keyframe anchoring workflow rather than generated in one go.
Without epistemology.md in the workspace the visuals default to the category's consensus look. You get a warning, not a fix.
When a product photo is the reference, the product is fixed. Label text, colours, logo, packaging and proportions cannot be altered, so a low-resolution source means asking for a better one rather than a workaround.

Where people go wrong

The mistake	Do this instead
Accepting the first prompt because it reads professionally	Run the kill switch on it yourself. If you can describe it fully as a person doing an activity at a location, it will generate a stock photo no matter how well written it is.
Expecting it to fill in the business phone number, price or URL	Supply the exact data in the session, or accept the placeholder and add the text yourself in post. This is a zero-tolerance rule, not a limitation to argue with.
Loading long rendered stats or numbers into the image itself	Keep rendered text to 2 to 5 words and simple round numbers. Put anything that must be letter-perfect on as a post-production overlay.
Skipping safe zones on vertical or feed output	Load the safe zones for the target platform before writing the prompt. UI overlays will otherwise sit on top of your subject or your text.
Asking for both the JSON prompt and the five-part brief	Pick one. The JSON schema replaces the director's brief format; producing both means two versions of the same direction drifting apart.

Worth knowing

Faces are worth keeping in frame, they raise click-through, but a face alone is exactly what fails the kill switch. Pair the face with one elevated element, an unexpected scale, a juxtaposition, or something that should not physically be possible, and you keep the recognition benefit without the stock-photo penalty.