version: "1.0.2" name: ai-content-generation description: "When the user wants to create AI-generated image or video content: a faceless influencer or persona for TikTok and Instagram, a faceless YouTube channel, an AI clone or talking-head of themselves, or AI ad creative (UGC, product video, static images). Also use when the user mentions 'faceless influencer,' 'faceless channel,' 'AI avatar,' 'AI UGC,' 'talking head,' 'clone myself,' 'AI ad creative,' 'AI video ads,' or tools like Veo, Google Flow, Kling, Higgsfield, Seedance, Runway, ElevenLabs, HeyGen, Arcads, Nano Banana, Midjourney, or Ideogram. Gives step-by-step, copy-paste-prompt playbooks for each use case plus the realism rules that beat AI slop. For paid-ad strategy and targeting, see performance-marketing or paid-ads." metadata: version: 1.5.0

AI Content Generation

Help the user actually make AI image and video content, end to end, with copy-paste prompts at every step. This skill covers four use cases, each a self-contained, do-this-then-that playbook. The hard parts (keeping a character consistent and killing the AI-slop look) are baked into the prompts, so the user can follow the steps without prior experience.

The core idea

The model is the easy part. What makes AI content win is the niche, the script, keeping the character or product consistent, and beating the AI-slop look. Every playbook bakes those in. Roughly 80% of quality is set before the model runs (research, script, consistency, realism); the model is the last 20%.

Pick the use case, then open its reference

Each reference is fully self-contained: a numbered walkthrough split into one-time setup and a repeating loop, with the exact prompts inline.

How to run this skill

1. Identify which of the four use cases the user is after (ask one question if unclear).
2. Open the matching reference file and follow it with the user, step by step.
3. At each tool step, give the user the copy-paste prompt from the reference, filled in with their specifics (niche, topic, product, audience).
4. Keep one character, one voice, and one look consistent across a user's content. Consistency is the most common failure point.
5. Apply the realism rules below to every generation prompt (they are already written into the reference prompts; do not strip them out).
6. Before the user publishes, run the reference's pre-publish checklist.

How the tools are named (app vs model)

The names look chaotic because three different things share the spotlight: apps you chat inside, wrappers that put one interface over many models, and the models that do the generating. Once you know which is which, every new launch slots in. The model generates. The app or wrapper is just the door you walk through to reach it.

Three buckets:

• Assistant / app: where you type. It runs models underneath. ChatGPT, Gemini, Claude.
• Wrapper: one interface over several models, usually for cheaper volume or shot control. Google Flow, Higgsfield.
• Model: the thing that actually generates, split by layer (text, image, video, voice). Veo 3.1, Gemini Omni, Nano Banana, Kling.

The big three, decoded. Each vendor has a company name, an app you chat in, and a family of models underneath. Mixing those three is most of the confusion.

• Google: Gemini is the assistant. Nano Banana is its image model. Veo 3.1 is its cinematic video model (API, 4K, long shots). Gemini Omni is its conversational video generation-and-editing model, now default in the Gemini app, Flow, and Shorts. Veo 3.1 did not go away. It stayed on the API for high-fidelity work while Omni took over the consumer app.
• OpenAI: OpenAI is the company. ChatGPT is the app. GPT (GPT-5 and friends) are the text models inside it. Sora is its video model, DALL-E / GPT Image its image model.
• Anthropic: Anthropic is the company. Claude is the assistant. Opus, Sonnet, and Haiku name the model tiers, from most capable down to fastest (for example Claude Opus 4.x).

To place any new model, ask two questions. Is it an app you chat in, a wrapper over other models, or a model itself? And which layer does it work on, text, image, video, or voice? Drop it in the matching row and move on.

Tool stack at a glance

The user does not need all of these. Each playbook names the few it uses. Start lean (free tiers plus ElevenLabs Starter is enough to ship), scale one tool per layer only when volume demands it.

The rule that beats AI slop

Realism is controlled imperfection. The default settings and generic prompts produce the average, which reads as fake. Add back the texture and flaws real capture has:

• Prompt for a phone-camera look: natural light, slight grain, real-time pace. Drop the words "8K", "cinematic", "perfect", "studio", "hyperreal". Those trigger the waxy plastic look.

• Put real pauses, breaths, and a filler before the key line in voice scripts. Flat delivery is the number one AI-audio tell.

• Keep clips short (3 to 10 seconds) and cut. Faces and physics drift past that.

• Segment the script before generating video. Break the finished script into ~10-second beats and generate one beat per clip, each with its own scene and only that beat's spoken line. Never hand a video model the whole 30 to 45 second script in one prompt. If a clip glitches, regenerate only that clip, then stitch in CapCut.

• No on-screen text in a video prompt. Video models (Veo, Flow, Kling, Gemini Omni) garble any text you ask them to render on screen. Add all captions, hook text, and CTAs in CapCut after generating. Baked-in text stays fine only for the still-image models (GPT Image, Ideogram, Nano Banana), which are chosen precisely because they render legible text.

• Lock the scene and reuse it across clips. Set the environment, outfit, background, camera angle, lighting, and style once, then repeat that same scene description in every beat's prompt. Reusing it verbatim keeps the character and setting from drifting between clips.

• For video, add negative prompts: no morphing, no warping, no melting, no jelly motion, no slow motion.

• Mix in real footage (stock or a real product photo) and apply one consistent color grade so generated and real assets read as one piece.

These are already written into the reference prompts. The reference "Make It Look Real" sections also list the editing and review moves a prompt cannot do.

Run it outside Claude Code (ChatGPT, Claude, Gemini)

A ready-to-use project package lives in assets/chatgpt-project/. It turns this skill into a ChatGPT Project, Claude Project, or Gemini Gem so a non-technical user can run the same steps in a normal chat:

• project-instructions.md — paste into the Project or Gem instructions.
• knowledge/ — upload these files as the project knowledge.

References

• references/faceless-influencer.md
• references/ai-clone-talking-head.md
• references/ai-ad-creative.md
• references/faceless-youtube.md