Talentless AI×Texas AI Alliance

A Studio of One

HOTA · Human Orchestration of Tools and Agents · 3 hours, hands-on

Wednesday, November 18, 2026 · ACC Rio Grande, Bldg 3000

with Steve Mudd, Talentless AI

Hosted by the Texas AI Alliance

Scan to keep this page

AI is a creative multiplier

Steve makes avatar videos, microdramas and client work by himself, with AI doing the jobs a crew used to do. Tonight you build the same setup: a brain, a brand system, a few skills, and the tools that turn one idea into images, audio and video. Bring a topic you care about.

The eight modules

1
Build a brain, a Claude project that knows your topic, voice and brand.
2
Build a brand system, then a one-pager, a newsletter and a deck, back to back.
3
Creative skills, why and when to teach the AI your process.
4
MCP, how your AI connects to the advanced tools.
5
AI images, vibe models vs business models, and a real infographic.
6
Audio, voice, sound and music with ElevenLabs. Good video starts here.
7
Video, four kinds: animated explainers, social cuts, AI avatars and cinematic.
8
Next level: a creative agent, play on a live Luma Labs board with the room.
$
What it costs, free tiers, paid plans and a realistic starter stack.

How the studio fits together

Brainwhat it knowstopic, voice, brandBranddesign systemcolors, type, rulesSkillsyour processsaved and reusableConnected toolsMCP and CLIimages, audio, videoFinished workon brandnot slop

Set up the first three once. After that, every new piece starts most of the way there.

Three ways to think about AI

1Operating system. The layer you work through, connected to your files, apps and tools.
2User experience. You describe what you want in plain words, and the work happens.
3Creative multiplier. One person can now make what used to take a crew.

The three layers

1The LLM. The model you talk to, such as Claude or ChatGPT. It runs everything else.= Operating system
2Agents inside the creative tools. HeyGen, ElevenLabs, Canva and others now have agents you just talk to.= User experience
3Agentic creation platforms. Purpose-built for making, like Luma: agents that run hundreds of generations to get work done at scale.= Creative multiplier

The one-two-three punch

Everything you make runs on three things. Context and references set it up. The prompt is where they meet your point of view.

1Context. What the AI knows before it makes anything.Your brain: topic, voice, brand. Set up once.
The conversation: what you work out together in the chat.
Tells it what
2References. The examples you hand it for this piece.Images: photos, style frames, character sheets.
Audio: a real voice, a sound, a track.
Video: all of it, together.
Shows it how
3The prompt. Where context meets references meets your point of view.Your call: the style, the camera, the feeling.
Your taste: what to keep, what to cut, what to try again.
Makes it yours

Think of it as one console: the LLM is the engine, these three feed it, and MCP is the cables to your tools.

Stay in touch

Connect with Steve on LinkedIn
Text me your work
Made something? Send it. 720-254-7893

The best get featured on this page and shared with the Talentless Hack community.

Subscribe to The Talentless Hack

Weekly AI build notes, and where the next class drops first.

Slop is the first draft, shipped

Slop is what you get when nobody iterates. The AI gave someone an answer on the first try, and they posted it. The first draft costs nothing now. Everything after it is your job.

Slop. Low-effort AI content made fast and in bulk. Merriam-Webster picked "slop" as its 2025 word of the year.
Work slop. The memo, deck or email that looks finished but says nothing, so the reader has to do the real work.
Punk slop. Slop on purpose: weird, trashy, obviously-AI work made as a style or a joke. It works when you meant to do it.

Writing slop

The tells:

"In today's fast-paced world""delve""unlock your potential""game-changer""It's not X, it's Y"everything in threesdashes everywhereno names, numbers or dates"In conclusion"made-up stats"Here's the thing""It's important to note""Let's dive in""Great question!""I hope this helps""Whether you're X or Y""Not X. Y." fragments"ever-evolving landscape""tapestry," "realm," "testament""leverage," "robust," "seamless"emoji bulletsbold on every other phraseeverything is a bulleted listTitle Case Headings: With a Colona summary that repeats the introflattery and hedgingsame length, same rhythm, every sentence
SlopElevate your mornings with our artisanal, handcrafted pastries, lovingly made to delight your senses and fuel your day.
UnsloppedOur croissants come out of the oven at 6:40. By 8, they're usually gone.

Spot the slop: a menu and a Facebook ad

This is the slop you see in the wild. Both were made in seconds with a cheap image model and a lazy prompt. Find the tells, then check the key.

AI slop diner menu

The diner menu.

AI slop Facebook ad for burgers

The Facebook ad.

1 The same dish twice. Golden Omelette is on the menu two times, with different copy.
2 Copy-paste errors. The salads come "with Fries or Slaw," and the salad description is the chicken sandwich's.
3 Misspelled words. "Impossisie stack," "lolgning vegetables," "vibranti."
4 The prompt leaked in. "Glowing cheese," "unnaturally perfect," "ethereal greens." Nobody describes their own food that way.
5 Food nobody can eat. A burger taller than your face. Cheese that flows like lava and glows.
6 The plastic family. Same smile, same skin, same face shape. They're all looking at nothing.
7 A hand from nowhere. Whose arm is that? Count the fingers on every hand.
8 Sparkles, god-rays and emoji stickers. When the picture can't sell it, the AI adds glitter.
9 A fake badge. "5-Star Rated" by whom? Same as a made-up stat.
10 Hype instead of facts. "Taste the magic! Best burgers ever!" No price, no address, no hours.

Made for this class with a one-line prompt. The restaurants aren't real.

Work slop

The tells:

headlines like "Overview" and "Key considerations"ten bullets per slideno decision, no owner, no datea recap of the recapa summary longer than the thing it summarizesAI meeting notes nobody read"circle back," "align," "drive impact"an icon on every bulleta deck when an email would doyou can't tell what the sender thinks
Work slopQ3 Update: Key Highlights. We continued to make significant progress across multiple workstreams. Stakeholder alignment remains a priority, and we are well positioned to drive impactful outcomes in Q4.
UnsloppedQ3: onboarding shipped, billing slipped. The new onboarding went live Sept 12, and sign-up completion rose from 41% to 58%. Billing slipped two weeks, so the Q4 launch moves to Nov 3. I need a yes on one QA contractor by Friday.

Example numbers are made up for illustration.

Image slop

The tells:

garbled or invented textplastic, airbrushed skinextra or melted fingersthe same pretty faceteal-and-orange glowshiny everythingperfect symmetryblurry background soupshadows from two sunsobjects melting into handsgibberish signs in the backgroundcolors that aren't yours

Three real generations from an earlier class:

Poster from a vague prompt

Slop. A vague prompt. The model wrote its own copy, dropped the headline and stamped "50% OFF."

Poster from a structured prompt

Unslopped. The exact words, in quotes. "GO BANANAS" and "NOW 50% OFF," exactly as asked.

AI chalkboard menu

Sneaky slop. A full menu that looks right, but it quietly changed a price. Always proofread.

Design slop

Ask Claude, or any AI, for "a landing page" with no direction, and you get the average of every website it has seen. It looks fine and it looks like everyone else. Here is that page, built the way an AI builds it unaided. Count the tells.

Get started 1
Now with AI agents 2
The future of work 3

Supercharge your workflow
with intelligent automation 4

Unlock seamless collaboration and boost productivity with our all-in-one AI-powered platform. 5

Start free trial
Watch demo →
Trusted by 10,000+ teams worldwide 6
ACMEGLOBEXINITECHUMBRELLA
⚡Lightning fast

Built for speed, so your team can move faster than ever.

🔒Secure by default

Enterprise-grade security you can trust. 7

🤖AI-powered

Smart insights that help you work smarter, not harder.

10xfaster workflows
99.9%uptime
24/7support 8

A made-up company, built for this class.

1 Purple-to-blue gradients. On the buttons, the logo and the headline. The number one tell, because it's the most common color in the code AI learned from.
2 The pill badge with a glowing dot. "New," "Beta," "Now with AI," floating above the headline.
3 The eyebrow. A tiny all-caps label above the headline that says nothing.
4 Gradient text in a centered hero. Big headline, middle of the page, two buttons under it.
5 Inter, and copy slop. The default font, plus "supercharge," "seamless," "unlock."
6 Fake social proof. "Trusted by 10,000+ teams" and a row of logos nobody checked.
7 Three rounded cards with emoji icons. Glassy, with soft shadows and a glow.
8 The stat band. 10x, 99.9%, 24/7. Numbers with no source.

The second wave. Tell an AI to make it "look less AI" and it swaps the purple for warm cream with a terracotta accent, or near-black with one acid-green highlight. Same defaults, new paint. And yes, this page has eyebrows too. Now you'll see them everywhere.

Design brief before you build
Before you design anything, ask me five questions about my audience, my brand and the feeling I want. Then propose three different visual directions with a named font pairing, a palette of four to six colors with roles, and a layout idea for each. Don't use purple gradients, Inter, emoji icons, pill badges, eyebrow labels, glow effects or a centered hero with three cards unless I pick them on purpose.

Video and audio slop

mouths out of synchands and faces that morpha character who changes between shotsfloaty camera with no reasonflat AI narrationgeneric stock musicno story, just shotsevery shot a slow push-inevery clip the same 5 to 8 secondsphysics that floatseyes that never blink or lookwrong number of fingers mid-gesturetext that crawls and changeslip-sync on a mouth that isn't moving

Make some slop on purpose

Slop Paint is a free paint program from Talentless AI. Draw a doodle, pick a style and press Slopify. About 20 seconds later you get an AI version of your drawing. No sign-up.

Try Slop Paint

The unslop checklist

1Read it out loud. If you wouldn't say it, cut it.
2Check every fact. Every number, name, date and quote against your sources.
3Cut the tells. Swap generic phrases for specifics.
4Add what only you know. One detail, opinion or story the AI couldn't have.
5Put it next to your brand. Colors, voice, logo. Fix what drifts.
6Change one thing at a time. Then look again.
7Run the tough reviewer. Let the AI play your hardest critic.
8Would you sign it? If not, it isn't done.
Slop detector
Read this draft and list every phrase that sounds like AI wrote it, every claim without a source, and every place I could use a specific name, number or example instead. Don't rewrite it yet.
Unslop pass
Read this like an editor. Cut generic phrases, check every fact and number against my sources, replace anything vague with something specific, and keep my voice. Show me what you changed.
Go deeper

1 · Build a brain

Claude

Projects, deep research, your voice and brand.

Open / install

A brain is what the AI knows about your topic, your voice and your brand before you type anything. Set it up once and stop pasting the same background into every chat.

The layers of a brain

InstructionsWho it serves, what to always and never doVoice layerHow you sound: tone, words you use and avoidBrand layerHow you look: colors, type, logo, photo styleYour contentPast posts, decks, bios, notes, linksResearchSourced findings on your topic, saved as documentsEvery chat,skill and toolstarts herePermanent brain: the same layers in a shared database, like Supabase

Build from the bottom up: research and your own content first, then the brand and voice layers, then the instructions that tell the AI how to use it all.

References

What goes into a brain

Your writing: posts, decks, bios, emailsTranscripts of you talkingResearch with sourcesYour brand guideExamples you loveExamples you hate, and whyClient briefs and past feedback

Your brain is a stack of references. Feed it only what you'd sign: slop in, slop out.

Two types of brains

Project brain (tonight). A Claude project for one topic or client. Fast to set up, and it lives inside Claude.
Permanent brain (show and tell). A database that many tools and agents read and write, like the Supabase setups behind a newsroom or an app. More setup, and it lasts.

Exercise 1: build your project brain

1Set up a project in Claude. Name it after your topic, and add short instructions about who it serves.
2Run deep research on your topic. Turn on Research and ask for a sourced report.
3Write up the findings. Save the report as a document in the project.
4Upload your content. Past posts, decks, bios, notes, links. The more real material, the better.
5Add the big thing: a voice layer and a brand layer. How you sound, and how you look.
Deep research prompt
Research [topic] for [audience]. Give me: the five most important facts with a source for each, the three biggest open questions, who the key people and organizations are, and what has changed in the last 12 months. Cite every claim.
Voice layer prompt
Read the writing samples in this project. Write a voice guide for me: tone, sentence length, words I use, words I never use, and three before-and-after examples. Then ask me five questions to sharpen it.
Project instructions (paste into the project)
You are my content partner for [topic].
Audience: [who], who care about [what].
Always: write in my voice (see the voice guide), cite sources, flag anything you are not sure of.
Never: make up numbers, quotes or names; use jargon I would not use.
Before any big piece, ask me up to five questions first.
Brand layer prompt
Look at the brand examples I uploaded. Write a one-page brand guide: colors with hex codes, fonts, logo rules, photo style, three words for the look, and five do and don't rules. Keep it short enough to paste into any prompt.
Prompt tip: ask before it acts. End big requests with "Ask me five questions before you start." The questions show you what it doesn't know yet.
Prompt tip: keep it in the brain. When a chat gives you something good, save it to the project as a file. That's how the brain grows.
Unslop it

Read the research before you trust it. Open the sources, cut anything you can't verify, and add what only you know.

Homework

Ask your project to write one paragraph about your topic in your voice. If it sounds generic, your brain needs more of your real material.

Go deeper

Claude

Your brain, design system, skills and connectors.

Web + desktop + iOS + Android. Paid plan for Cowork.

claude.ai

2 · Build a brand system

Claude

Design systems, docs, decks and pages.

Open / install

A brand system is your colors, type, spacing and rules, written down where the AI can use them. Set it up once and every one-pager, newsletter and deck looks like you made it.

Exercise 2: brand system, then three assets

1Build a brand system in Claude's design system. Give it your logo, colors, fonts and a few examples you like.
2Make three assets from it, back to back: a one-pager, a newsletter and a presentation.
Brand system prompt
Build a design system for [brand]. Use these colors: [hex codes]. Fonts: [headline font] and [body font]. Voice: [three words]. Include a type scale, spacing, buttons and cards, plus do and don't rules. Base it on the examples I attached.
Asset prompt
Using my design system and the brain in this project, make a one-page overview of [topic] for [audience]. Then make a short newsletter edition on the same topic, then a 6-slide presentation. Keep all three on brand.

Borrow a designer's eye

Your brand system says what your stuff looks like. A designer prompt says how to think about it. Name a designer or a movement as a starting direction, then make it yours. Use it for the style and the thinking, never to copy someone's actual logo, poster or work.

Paul Rand. Simple shapes, wit, a single idea per piece.
Swiss style. Grids, sans serif type, lots of white space.
Saul Bass. Bold cut-paper shapes and strong color.
Bauhaus. Primary colors, geometry, function first.
Designer prompt
Redesign this one-pager the way Paul Rand would think about it: one clear idea, simple geometric shapes, a touch of wit, lots of white space. Keep my brand colors and fonts. Don't copy any of his actual work. Explain the three changes you made.

Save it as a skill. Once a direction works, put it in a skill with your brand system, so every new piece starts with that eye.

Refine, don't regenerate

Never accept the first draft. Point at one piece and fix it.

Section refinement
Rewrite just this headline in my voice. Make the value clear in under eight words. Give me three options, then tell me which one you'd pick and why.
Make it a newsletter
Turn the one-pager into a newsletter edition: a subject line (three options), a two-sentence opener, three short sections with subheads, and one clear call to action. Keep my voice and my brand.

Using ChatGPT instead?

ChatGPT has no design system of its own. The closest option is to connect Canva and use your Canva Brand Kit, so the brand lives in Canva and ChatGPT borrows it.

Unslop it

Put the three assets side by side. Fix anything off brand, cut filler words, and check every fact against your brain.

Go deeper

Claude

Your brain, design system, skills and connectors.

Web + desktop + iOS + Android. Paid plan for Cowork.

claude.ai

Canva

Brand Kit; the way ChatGPT gets your brand.

Web + iOS + Android.

canva.com

3 · Creative skills

A skill is a saved set of instructions for a job you do over and over. Write it once, and the AI does that job your way.

The rule: if you explain the same thing twice, make it a skill.

Why build a skill

Consistency. The same steps, rules and quality bar every time.
Memory of lessons. Every round of feedback becomes a rule, so mistakes don't come back.
Speed. One short request runs a whole workflow.
Handoff. Anyone on your team can run your process.

What skills can do

Write in a voice. House style, banned phrases, dialogue rules.
Build an asset. A newsletter, a one-pager, a buyer package for a show.
Run a pipeline. Script, voice, visuals and render, step by step.
Prompt a tool. Video prompts written the way a specific model likes them.
Check work. Quality checks against a brand or a client's rules.
Sort and file. Name, dedupe and organize generated media.

Start with two skills that work with your brain

Your brain only helps if it stays current, and it only pays off when you make something from it. These two skills do both.

1A research skill that feeds the brain. It researches your topic, keeps what's new and sourced, and adds it to the brain in the same format every time.
2A newsletter skill that uses the brain. It picks the freshest material from the brain and writes an issue in your voice and your brand.
Create a brain-building skill
Create a skill called brain-builder. When I give it a topic or a link, it researches the topic, keeps only facts with a source, checks them against what's already in my brain, and writes a short update: what's new, why it matters, and the sources. It adds the update to my project brain in the same format every time and tells me what it skipped and why.
Create a newsletter skill
Create a skill called newsletter-from-brain. It reads my project brain, picks the three freshest and most useful items, and writes one newsletter issue in my voice: three subject lines to choose from, a two-sentence opener, three short sections, and one call to action. It uses my design system for the layout, cites where each fact came from, and runs my unslop checklist before it shows me the draft.

Run the research skill once a week, then the newsletter skill. That's a publishing schedule you didn't have to staff.

What goes in a good skill

1A clear name and trigger. When should the AI use it? "Use when I ask for a newsletter."
2The steps. What you do, in order.
3The rules. Voice, brand, banned words, what never to do.
4A checklist. What to check before showing you the result.
Interview me for a skill
I want to turn how I make [asset] into a skill. Interview me one question at a time about my steps, my rules, and the mistakes you should avoid. When you have enough, write the skill.
Make a skill
Turn what we just did into a skill called [name]. It should make a [asset] from my brain and design system. Include the steps, the rules I gave you, and a checklist to run before you show me the result.
Homework

Make a skill for one of your three assets, then run it on a new topic.

Go deeper

4 · MCP: how Claude runs your other apps

MCP (Model Context Protocol) is a plug that lets Claude use other apps. If a tool has an MCP connector, you can ask Claude to run it for you, so you don't have to learn every video tool yourself.

Youask in plainwordsClaudeyour agent+ brain+ brand+ skillsMCP connectorsVoice and soundElevenLabsImages and designCanva, Luma, IdeogramVideoHeyGen, Runway, DescriptYour workGmail, Drive, Calendar, NotionCLI: Claude runs commands on your computer. Render video, convert files, run scripts.HeyGen and Luma have their own agents. Claude can direct them.

You talk to Claude. Claude runs the other tools.

What it means for you

One place to work. You ask Claude, and Claude calls ElevenLabs, HeyGen, Canva or Luma.
A test for any tool. Before you learn a new tool, check whether it has an MCP connector.
What it connects to. Voice, video, image and design tools, plus your own files, email, calendar and docs.
Stay safe. Use official connectors, read what each one can do, and keep approvals on for sending, posting and paying.

Which creative tools have MCPs

Two kinds. Hosted connectors run in the cloud: you paste a URL or pick them from Claude's directory and sign in. Local ones run on your computer next to the app, and need the app open plus a one-time setup.

ToolWhat it doesHow to connect
HeyGen HostedAI avatars, video agent, translationAdd custom connector: https://mcp.heygen.com/mcp/v1/
HeyGen MCP docs
HyperFrames HostedHeyGen's motion graphics: Claude writes a web page, HyperFrames renders it as videoAdd custom connector: https://mcp.heygen.com/mcp/hyperframes
HyperFrames MCP guide
Descript HostedEdit video by editing text: transcribe, cut filler, caption, make clipsIn Claude's connector directory
Connect Descript to Claude
Canva HostedSearch, create, autofill, resize and export designsIn Claude's directory, or https://mcp.canva.com/mcp
Canva AI Connector setup
DaVinci Resolve LocalPro editing and color. Claude can build timelines, import media, add markers and renderCommunity server. Needs Resolve open with external scripting set to Local
davinci-resolve-mcp on GitHub
Blender Local3D. Claude builds scenes, objects and materials from a promptBlender add-on plus a small server
blender-mcp on GitHub
ArtCraft suite LocalFree, open-source apps modeled on Adobe's: PhotoCraft (Photoshop), VectorCraft (Illustrator), FilmCraft (Premiere), EffectCraft (After Effects), LightCraft (Lightroom), SoundCraft and more. Built with AI, mostly alphaDownload or run in the browser: artcraftsuite.org
Community MCP for PhotoCraft (Windows): photocraft-mcp

Local servers are community projects, not made by the app company. Read what they can do before you install, and expect some setup. Hosted connectors need a paid Claude plan.

The CLI: another way in

A CLI (command line interface) is a tool you run by typing commands. AI agents can run those commands for you, which is another way to connect AI to software that has no MCP connector. That's how an agent can render a video or process files on your computer.

Connect a tool in Claude

1Select + in a chat (or type /), then Connectors, then Manage connectors.
2Browse the directory, pick the tool, and select Connect.
3Sign in to the tool to approve it. Then just ask Claude to use it.
First prompt with any new connector
What can you do with the [tool] connector? List its tools in plain words, tell me which ones change things versus just read, and suggest three ways I could use it for [my topic].
Safe action prompt
Use [tool] to draft [the thing]. Show me the draft before you send, post, publish or spend anything.

Two kinds of agents

Your agent drives the tools. Claude connects to tools through MCP and runs them from one seat.
The tool's own agent. Inside HeyGen or Luma, the platform's agent picks the models and does the work. Your agent can direct that agent too.
Go deeper

5 · AI images

Midjourney

The vibe model. Beautiful and cinematic.

Open / install

Some image models make beautiful pictures. Others get the words and the layout right. Pick the one that fits the job, then tell it exactly what you want.

Two kinds of image models

Vibe models. Midjourney and Uni. The most beautiful, cinematic looks, with less control over the details.
Business models. ChatGPT's image model, Nano Banana (Gemini) and Microsoft's image model. Accurate text, layouts and edits by conversation.

Claude has no image model of its own. You connect it to one through MCP, or give it access to your browser. In ChatGPT, you can use its built-in image model.

References

What you feed an image model

Photos you tookStyle frames and mood boardsCharacter sheetsProduct shotsYour logo and brand colorsLayouts you likeYour own sketch

Tell the model what each reference is for: "Use image 1 for her face, image 2 for the light, image 3 for the color."

Every word leaves a trace

Every word in a prompt changes the picture. Here is one scene, made six times in Uni, Luma's image model. Each version changes one choice from the base prompt.

The base prompt
A cellist playing alone on a city rooftop at golden hour. Warm low sun behind her, rim light in her hair. Eye-level medium shot, 50mm lens, shallow depth of field, the skyline soft behind her. Cinematic photograph, calm and focused mood.

What goes into an artistic prompt

You don't have to master all of these. Know they exist, so you can ask for them, and so you can see what's missing when an image feels off.

Subject and action. Who or what, doing what.
Setting. Where, and what surrounds them.
Time of day and weather. Golden hour, blue hour, noon, rain, fog.
Lighting. Backlight, rim light, soft window light, hard flash, neon, candlelight.
Camera angle. Eye level, low angle, high angle, overhead, Dutch tilt.
Shot size. Extreme wide, wide, medium, close-up, extreme close-up.
Lens and focus. 24mm wide, 50mm natural, 85mm portrait, shallow or deep focus.
Composition. Centered, rule of thirds, empty space, leading lines, foreground framing.
Style and medium. Photograph, watercolor, oil painting, 3D render, ink, collage.
Film and texture. 35mm film, Portra, black and white, grain, polaroid, VHS.
Color palette. Warm, cool, muted, neon, your brand colors by hex code.
Mood and emotion. Calm, tense, joyful, lonely, triumphant.

For video, add camera movement: push in, pull back, pan, orbit, handheld, static.

Let the AI be your cinematographer
Here's my idea: [one sentence]. Act as a cinematographer. Turn it into a full image prompt that names the subject and action, setting, time of day, lighting, camera angle, shot size, lens, composition, style, film or texture, color palette and mood. Then give me three variations that each change one of those choices, and tell me which one you changed.

How to prompt

The banana rule
A phototype of a bananasubject in tiny sunglassesflair

Then turn up the detail with six dials: who, where, what, light, mood, style. The more specific, the closer it lands.

One dial at a time. Keep the prompt, change one thing, regenerate.
Words in the image. Write the text first, put it in quotes, and proofread.
Prompt libraries. Save the prompts that work, and reuse them.
Ask the AI. Have it write the detailed prompt for you from your brain.

Prompt modifier cheat sheet

Lighting: golden hour, soft window light, cinematic rim light, volumetric god-rays, softbox studio.
Framing: extreme close-up macro, eye-level portrait, wide establishing shot, bird's-eye flat lay.
Aspect ratio: 4:5 or 9:16 vertical (social), 1:1 square (feed), 16:9 wide (decks and sites).
A loaded prompt
Subject: a single artisan coffee bag, kraft paper, label facing camera.
Label text (render exactly): "RIVER & OAK" / "Single-Origin Ethiopia" / "Whole Bean 12 oz".
Scene: warm walnut countertop, scattered beans, a softly blurred cafe window behind.
Lighting: soft morning key from the left, gentle shadows. Camera: 50mm, shallow depth of field.
Style: premium editorial product photography, photoreal. Palette: deep navy, cream, a hint of copper.
Do NOT include: extra text, gibberish, watermarks, people, plastic surfaces. Output: vertical 4:5.
Edit by talking
Keep everything the same, but: make it wider (16:9), move the product to the left third, and change the background to [color]. Don't change any text.
Let the AI write the prompt
Using my brain and brand guide, write a detailed image prompt for [what I need]. Include subject, scene, lighting, camera, style, palette, exact text in quotes, and a do-not-include list.
Unslop it

Spot the hallmarks: garbled text, plastic sheen, the generic stock look. Then add reference images, change one dial and iterate.

Exercise 3: make an infographic

1Make the slop version first with a one-line prompt.
2Run the detailed prompt below, filled in from your brain.
3Unslop it. Check every number and word against your sources, fix the hierarchy, cut clutter, apply your brand and proofread.
Advanced: the detailed infographic prompt
Create a detailed vertical infographic (4:5, 1080 x 1350).

Goal: explain [topic] to [audience] in under 30 seconds.

Headline (render exactly): "[headline, 8 words max]"
Subhead (render exactly): "[one sentence]"

Layout, top to bottom:
1. Headline band across the top
2. Hero number in the center: "[number]" with the label "[label]"
3. Three sections in a row, each with an icon, a short title and one line of text:
   - "[title 1]": "[line 1]"
   - "[title 2]": "[line 2]"
   - "[title 3]": "[line 3]"
4. A simple bar chart comparing [A] and [B]: [values]
5. Footer: source line "[source]" and logo space at bottom right

Style: clean editorial infographic, flat vector icons, generous white space,
clear visual hierarchy, one focal point.
Brand: palette [hex 1], [hex 2], [hex 3]. Bold sans-serif headlines,
light sans-serif body.
Text rules: use only the words in quotes, spelled exactly. No other text.
Do NOT include: gibberish text, made-up numbers, stock-photo people,
3D clip art, heavy gradients, drop shadows, watermarks.
Infographic: How to Spot AI Slop

This prompt, filled in for "How to Spot AI Slop" with the Talentless AI colors, run once in ChatGPT's image model. Every word came out right on the first try. Now unslop it: check it, then change one dial at a time.

Homework

Make the same infographic in a vibe model and a business model. Compare what each one got right.

Go deeper

ChatGPT

Built-in image model, easy on-ramp.

Web + iOS + Android.

chatgpt.com

Nano Banana / Gemini

Google's image model: text, layouts, edits by talking.

Web + iOS + Android.

gemini.google.com

Ideogram

The typographer. Posters, logos, signage.

Web + iOS + Android.

ideogram.ai

Magnific

Upscale to print resolution. Do it last.

Web.

magnific.ai

6 · Good video starts with audio

ElevenLabs

Voice, sound effects and music.

Open / install

Finish the voiceover and music first. They set the pace and the mood, and you cut the video to them.

ElevenLabs: voice, sound and music

Voice. Narration in a stock voice, a designed voice or your own clone.
Sound. Sound effects from a description.
Music. A score or bed for your piece.
References

What you feed an audio model

A clean recording of your voice, to clone itA voice description: age, accent, pace, moodA sample track for the style you wantReal sounds you recordedThe script, with how to say tricky names

Clone only your own voice, or a voice you have permission to use. Record in a quiet room: the model copies the room too.

Just want a song? Suno or Stable Audio

ElevenLabs is the class example because it does voice, sound and music in one place and connects to Claude. If you only need music, Steve uses Suno and Stable Audio all the time. Pick by what you'll do with it.

Suno. Full songs with vocals, from a prompt or your own lyrics. Fast for a finished track.
Stable Audio. Instrumentals, beds and loops. Good for scoring under a voice.

Next level: build music in layers

One prompt gives you one flat song. For real control, make each part on its own: drums, bass, synths, textures, vocals. Then cut, process and stack them, the way a producer works. Steve builds his tracks this way (think Nine Inch Nails), with a Claude skill that runs the process: it generates each stem, gives him options at every step, then assembles the mix in code.

Layered song, step one
Help me build an original song in layers. Ask me for the mood, tempo and references in words, not artist names. Then give me three options for the drum pattern only. After I pick, do the same for bass, then synths, then texture, then vocals. Keep a running list of what we chose.

Exercise 4: connect the ElevenLabs MCP

1In Claude, select + in a chat, then Connectors, then Manage connectors.
2Find ElevenLabs and select Connect. Sign in to ElevenLabs to approve it.
3Ask Claude to narrate one paragraph from your brain, then add a music bed.

No ElevenLabs account? Follow along on the screen.

Audio prompt
Using ElevenLabs, narrate this paragraph in a warm, clear voice: [paragraph]. Then make a 30-second music bed that feels [mood], and a short sound effect for the opening.
Design a voice
Design a voice for my narrator: [age range], [accent], [tone, like warm, wry, calm], speaking pace [slow or quick]. Read this line in it: [line]. Give me three takes.
Write for the ear. Short sentences, one idea each. Read your script out loud before you generate.
Spell it how it sounds. If a name comes out wrong, respell it in the script ("Rio Grand"), and keep the real spelling on screen.
Unslop it

Listen on headphones. Fix wrong words and odd stresses, cut dead air, and keep the music under the voice.

Homework

Narrate one paragraph from your brain in two different voices. Pick the one that sounds like you.

Go deeper

ElevenLabs

Voice, sound effects and music.

Web + iOS + Android. Has an MCP connector.

elevenlabs.io

Suno

Full songs with vocals from a prompt or your lyrics.

Web + iOS + Android.

suno.com

Stable Audio

Music and sound from Stability AI. Strong on instrumentals and loops.

Web.

stableaudio.com

Descript

Edit video by editing text.

Desktop + web. Has an MCP connector.

descript.com

7 · AI video

HeyGen

Video agent and avatars.

Open / install

An AI video is audio and images, put in motion and edited. What you're making decides the workflow and the tools, so start there.

References

What you feed a video model: all of it

Images: keyframes, character sheets, location platesFirst and last framesVideo clips: a camera move, a performance, a motion to copyAudio: the song, the voiceover, for timing and lip syncYour brand kit for motion graphics

Video is where references stack. One shot can take a keyframe, a character sheet, a clip of the move you want and the song it plays under.

Four kinds of video

Start with the everyday work: corporate and social video. Cinematic comes last, because it takes the most time and the most credits.

1 · Corporate

Animated explainers

Training, updates and data stories. Claude designs and builds the video in your brand.

2 · Social

Social cuts from your footage

Webinars, events and interviews, cut into clips with captions.

3 · Corporate and social

AI avatars

You on camera without the camera. HeyGen is the example tonight.

4 · Cinematic

Commercials and short films

References, keyframes and image to video. The most control and the most work.

Use case 1 · Corporate

Animated explainers: Claude designs the video

Good forTraining, updates, data stories
ToolsClaude, ElevenLabs
TimeAn afternoon

Every video starts the same way: you tell Claude what you're making. Claude already has your brain and your brand, so it writes the script, adds the voice, and builds the video as an animated web page. Then it records that page to an MP4. It can pull in your slides, charts and B-roll too.

1Tell Claude what you're making. Who it's for, how long, and what to use. Approve the script.
2Claude adds the voiceover through the ElevenLabs connector.
3Claude creates the video, scene by scene, on brand, timed to the voice.
4Add B-roll and your slides. Ask for footage, AI stills or charts in any scene.

Made from one request to Claude: a 15-second animated video written as a web page and rendered to MP4, with a B-roll still from Steve's Luma board. Silent demo.

Motion graphics video
Using my brain and brand system, write a 90-second script about [topic] for [audience], scene by scene. When I approve it, make the voiceover with ElevenLabs, then build an animated video in my brand that times every on-screen change to the voice. Use these slides for the charts: [attach].
Use case 2 · Social

Social cuts from footage you already have

Good forWebinars, events, interviews
ToolsClaude, Descript, Opus Clip
TimeAn hour

Most teams are sitting on hours of footage. This one starts with Claude too: upload the footage to Descript, then tell Claude what you want.

1Upload the footage to Descript, or give it straight to Claude.
2Tell Claude what you want. Through the Descript connector it reads the transcript, looks at the visuals, and finds the best moments.
3Get a first cut back, then shape it yourself.
4Spin out clips for social, with captions.
Opus Clip. If you only need shorts, Opus Clip turns one long video into a batch of captioned, reframed clips.
Under the hood. When Claude cuts video itself, it uses ffmpeg, a free video tool. You never have to touch it.
First pass on footage
Here's the recording of [event]. Read the transcript and watch for strong visual moments. Give me the five best moments with timestamps, a 3-minute highlight cut, and three 30-second social clips with captions.
Editing with Descript's Underlord
Remove filler words and long pauses. Add captions in my brand font. Add a soft music bed under the talking. Then make three 30-second highlight clips for social.
Use case 3 · Corporate and social

AI avatars with HeyGen

Good forUpdates, lessons, promos, translations
ToolsClaude, HeyGen
TimeA day to set up your avatar, minutes per video after that

Make an avatar of yourself to deliver updates, lessons and promos, or use stock avatars in a controlled, labeled way. Same start: connect the HeyGen connector, then talk to Claude.

1Connect HeyGen in Claude's connectors.
2Tell Claude what you're making. Claude writes the script from your brain.
3Claude sends it to HeyGen with your avatar and voice, and brings back the video.
HeyGen video agent
Make a 60-second vertical explainer about [topic] for [audience], presented by my avatar. Open with a hook in the first three seconds, cover three points, end with [call to action]. Use my brand colors [hex codes], captions on, upbeat but not salesy.
Example
The promo for this class, made with a HeyGen avatar

"AI Steve," built from a HeyGen digital twin and a cloned voice.

Record clean. Good sound, soft even light, a simple background, and look at the lens. Your avatar is only as good as your footage.
Disclose it. Label synthetic video, and never make an avatar of someone without their consent.
Use case 4 · Cinematic

Cinematic: a commercial, a short film, a story

Good forAds, brand films, series
ToolsDreamina, Higgsfield, Runway, Luma
TimeDays, and the most credits

References look like this. A turnaround sheet shows one character from every angle, so every shot keeps the same face and costume. These are from Steve's boards.

Every cinematic video starts with references and keyframes.

References. A character sheet, a location, a product, a style frame. Reuse them in every shot, and your people and places stay consistent.
Keyframes. The first frame of each shot, and the last frame when the model allows it. Approve the stills before you spend video credits.
Reference sheet
Create a character reference sheet for [character]: front, three-quarter and side views, neutral light, plain background, the same outfit in every view. No text.
Keyframe
Using my character and location references, make the first frame of shot [number]: [who], [where], [the moment just before the action], [camera framing], [lighting]. 16:9. Match the references exactly.

The motion recipe: camera move (push in, pull back, pan, orbit, static) + performance (what the subject does) + direction (where the motion goes). One move per shot.

Motion prompt
Slow push-in on the coffee bag as steam drifts past and morning light shifts across the counter. Gentle, warm, cinematic. 5 seconds. No text, no new objects.

Tools: Dreamina, Higgsfield, Runway and Luma turn keyframes into shots.

Video models, in plain words

Video models take different inputs. Most real work mixes them.

Text to video. You describe the shot and get a clip. Fast, with the least control.
Image to video. You give it a keyframe and it puts the image in motion. This is how cinematic work starts.
Reference to video. You give it characters, places or products, and it keeps them consistent across shots.
Video to video. You give it a clip and it restyles it, changes the shot, or swaps what's in it.
ModelMade byKnown for
Seedance 2.5ByteDance (in Dreamina)Long, reference-heavy scenes with picture and sound together. Ranked first for image to video on Artificial Analysis in November 2026.
Veo 3.1Google (in Flow)Outdoor, atmospheric scenes and native dialogue. Clips run 8 seconds. Among the most expensive.
Kling 3.0KuaishouMulti-shot sequences and 4K. A cheaper alternative.
Runway Gen-4.5RunwayHands-on creative control, plus video-to-video editing.
RayLumaRuns inside Luma's agent and boards. We use it in the next module.

Rankings change every month. Test your shot in two or three models before you commit. In 8 · Luma, we run the whole process: make an image, turn it into a shot, ask the agent to build on it, and put it together.

Unslop it

Watch it with fresh eyes. Fix the audio first, cut anything that drags, and replace any shot that looks generic.

Homework

Pick the use case closest to your work and make 30 seconds of it this week.

Go deeper

HeyGen

Video agent and avatars.

Web + iOS + Android. Has an MCP connector.

heygen.com

Higgsfield

Cinematic video with real camera control.

Web.

higgsfield.ai

Runway

Video generation, editing and its own agent.

Web + iOS. Has an MCP connector.

runwayml.com

Descript

Edit video by editing text.

Desktop + web. Has an MCP connector.

descript.com

Claude

Your brain, design system, skills and connectors.

Web + desktop + iOS + Android. Paid plan for Cowork.

claude.ai

8 · Next level: a creative agent

Luma Labs

Images, video, audio and agents on one board.

Open / install

Luma puts images, video and audio on one board, with an agent that plans and makes the work with you. Tonight the whole room shares a real board.

Hands-on
Open the class Luma board

Make images, video and audio on the same board as everyone else in the room. The board needs an invite. Make a free Luma account first, then sign up on the Welcome tab with the same email and check "Add me to the class Luma board."

How a real project runs: a music video

This is the workflow behind Steve's educational hip-hop videos. The client sends a song and a loose shot list. Everything after that runs through Claude and the Luma agent.

INPUTPRE-PRODUCTIONPRODUCTIONSHOTSEDITClaudeLuma agentVideo modelsThe clientSong + a looseshot listThe planBack and forth with ClaudeVisual styleCamera stylePerformance styleOur shot listSong audioClaude cuts it intoclips for each shotCharactersReal-looking people,checked for slopUniLocationsLocation plates:the worldUniKeyframesPeople + places,shot by shotGPT Image SunburstShot prompts+ keyframes+ reference frames+ audio clipsPick by shot typeSeedance 2.5Seedance 2.0MiniMaxThe editDownload the shots;the editor cutsthem to the song

Claude is pre-production. Luma is production, with a different model for each job. The editor puts it together.

1Take the client's input. The song and a loose shot list.
2Pre-production in Claude. Go back and forth until the creative choices are made: the visual style, the camera style and the performance style. Rewrite the shot list so each line starts to look like a prompt. Claude also cuts the song into an audio clip for each shot.
3Production in Luma. Give the plan to the Luma agent. It makes the characters and the location plates with Uni, and you check the people for slop. Then it makes the keyframes with GPT Image Sunburst, putting the people in the places, shot by shot.
4Make the shots. Each shot prompt gets its keyframe, the reference frames and its audio clip. The video model depends on the shot: Seedance 2.5, Seedance 2.0 or MiniMax.
5The edit. Download the shots, and the editor cuts them to the song.

Different jobs, different models. Luma lets you switch models inside one board, so you pick the best one for each step.

By the numbers: Say's Law

The video in production right now is a rap battle between the economists Jean-Baptiste Say and Thomas Malthus. Here is what goes in, and what one prompt turns it into.

1song: an intro, 3 verses and 2 choruses
37rows in the shot list
144lip-sync takes: 72 shots, 2 takes each
122B-roll cuts
1prompt in the Luma agent
$328for 266 shots, about $1.23 a shot

The real output of pre-production is the prompt. All the back and forth in Claude ends in one prompt guide: the style, the references, the coverage rules and every shot. Paste that one prompt into the Luma agent, and it writes and runs all 266 shots for the first draft. Then you review, fix and regenerate the shots that miss.

Pre-production in Claude
Here's the song and the client's shot list. Before we make anything, help me decide the visual style, the camera style and the performance style. Ask me one question at a time. Then rewrite the shot list so each shot names the character, the location, the camera move and the line of the song it plays under.

House rules for the board

Play, but don't go crazy. A few generations each, so everyone gets a turn.
Build on each other. Remix what someone else made, and see where it goes.
No real people. No faces of real people, celebrities or brands you don't own.
Room-safe. Nothing you wouldn't show the whole room.

Things to try

1Make an image with Uni, using the banana rule.
2Turn it into a shot with one camera move.
3Add sound or music, and watch the shot change.
4Ask the agent to plan a 15-second sequence from your best still.
Ask the Luma agent
Take this image and plan a 15-second, three-shot sequence around it. Keep the character and style consistent, add one camera move per shot, and suggest a sound bed. Show me the plan before you generate.
Go deeper

Luma Labs

Cinematic video, Uni images and agents in one place.

Web + iOS. Has an MCP connector.

lumalabs.ai

Prompting playbook

These habits work in any AI tool, for any format.

1Context first. Who it's for, what it's for, and what good looks like. Your brain does most of this for you.
2Ask me five questions first. Let the AI find the gaps before it writes.
3Show, don't just tell. Paste an example, attach a reference image, link the brand guide.
4Name the format. Length, structure, aspect ratio, number of options.
5Iterate high. Fix the goal and the outline before you polish words. One good outline saves ten edits.
6Change one thing at a time. Each word is a lever. Move one, look, then move the next.
7Build a reviewer. Have the AI critique the draft like your toughest boss before anyone else sees it.
8Let the AI write the prompt. Describe what you want, and ask it for the detailed prompt.
The universal starter
I need [deliverable] for [audience] so they [action or feeling].
Use my brain and brand guide in this project.
Format: [length, structure, options].
Before you start, ask me up to five questions.
The tough reviewer
Review this draft as [my boss or my toughest client] would. List the five strongest objections, what's missing, and anything that sounds like AI wrote it. Then suggest fixes, but don't rewrite it yet.
Unslop pass
Read this like an editor. Cut generic phrases, check every fact and number against my sources, replace anything vague with something specific, and keep my voice. Show me what you changed.
Prompt writer
I want [result]. Ask me what you need to know, then write the most detailed prompt you can for [tool], with the exact words, style, format and a do-not-include list.
Go deeper

What it costs

It adds up. A starter stack of Claude Pro, ElevenLabs Creator, HeyGen Creator and Luma Plus runs about $101 a month.

The rule: start free where you can, pay for the one tool that unlocks your main format, and add the rest only when a real project needs it.
ToolWhat you use it forFree optionEntry paid planHigher tiers
ClaudeBrain, orchestration, design, skillsYesPro, $20/moMax, $100 or $200/mo
ChatGPTChat, built-in image modelYesGo, $8/mo; Plus, $20/moPro, $100 to $500/mo
Gemini (Nano Banana)Business imagesYesGoogle AI Pro, $19.99/moUltra, from $99.99/mo
MidjourneyVibe imagesNoBasic, $10/moStandard $30, Pro $60, Mega $120
Luma LabsImages, cinematic video, agentsTrial creditsPlus, $30/moPro $90, Ultra $300
ElevenLabsVoice, sound, musicYes, no commercial useStarter, $6/mo; Creator, $22/moPro $99 and up
HeyGenVideo agent, avatarsYes, 3 watermarked videos a monthCreator, $29/moPro $49, Business $149
DreaminaImage-to-videoYes, limited creditsBasic, $15/moStandard $36, Advanced $79
DescriptEditingYes, 60 minutes a monthHobbyist, $24/moCreator $35, Business $65
CanvaDesign, brand kitYesPro, $18/moBusiness, $25 per person

US monthly list prices as of November 2026. Prices change often, so check each tool's pricing page. Some tools are cheaper billed yearly.

The cheat sheet

Everything on this page, on one card. Screenshot it.

The console

LLM + context + MCP. The engine, what it knows about you, and the cables to your tools.

Rerun it tomorrow

1Open your brain. Start every piece in your Claude project.
2Run a skill. Ask for the asset by name.
3Pull in the tools. Images, audio and video through your connectors.
4Unslop it. Ship only what you'd sign.

The banana rule

A [type] of a [subject] with a [flair]. Then turn up the six dials (who, where, what, light, mood, style). For words in the image, write the text first and quote it.

The eight modules

1 Brain, Claude project, deep research, voice and brand layers.
2 Brand, design system, then one-pager, newsletter, deck.
3 Skills, explain it twice, make it a skill.
4 MCP, connectors and the CLI.
5 Images, vibe: Midjourney, Uni · business: ChatGPT, Nano Banana, Microsoft.
6 Audio, ElevenLabs for voice, sound and music.
7 Video, cinematic · motion graphics · footage you have · avatars.
8 Luma, a creative agent: images, video and audio on one board.

The full toolkit

Claude

Your brain, design system, skills and connectors.

Web + desktop + iOS + Android. Paid plan for Cowork.

claude.ai

ChatGPT

Built-in image model, easy on-ramp.

Web + iOS + Android.

chatgpt.com

Nano Banana / Gemini

Google's image model: text, layouts, edits by talking.

Web + iOS + Android.

gemini.google.com

Ideogram

The typographer. Posters, logos, signage.

Web + iOS + Android.

ideogram.ai

ElevenLabs

Voice, sound effects and music.

Web + iOS + Android. Has an MCP connector.

elevenlabs.io

HeyGen

Video agent and avatars.

Web + iOS + Android. Has an MCP connector.

heygen.com

Higgsfield

Cinematic video with real camera control.

Web.

higgsfield.ai

Runway

Video generation, editing and its own agent.

Web + iOS. Has an MCP connector.

runwayml.com

Luma Labs

Cinematic video, Uni images and agents in one place.

Web + iOS. Has an MCP connector.

lumalabs.ai

Descript

Edit video by editing text.

Desktop + web. Has an MCP connector.

descript.com

Canva

Brand Kit; the way ChatGPT gets your brand.

Web + iOS + Android.

canva.com

We're AI so you don't have to be

Talentless AI is Steve's studio in Austin. It makes AI avatar videos and branded content for clients, and runs trainings like this one.

I'm Steve Mudd, founder of Talentless AI, veteran of IBM Watson and Ogilvy. I run classes like this one and build the pipelines I teach.

Three ways to work with us

Talentless AI

The studio. AI avatar videos, branded content, and trainings like tonight's. A certified HeyGen partner.
talentless.ai

Video Booth AI

An AI video booth for events. Guests step up and leave with their own short video, and you get the content. Steve built one for the CMO Huddles community.
Dead Galaxy key art

Backlot74

The studio's own AI series and microdramas: Double Wide Desires (30 episodes, done), Dead Galaxy, Mister Hum and more. What we sell, we test on our own shows first.
backlot74.com · microdramas on YouTube

Connect with Steve on LinkedIn

Interested? Check the box on the sign-up form and Steve will follow up.

Work together? Email steve.mudd@talentless.ai, text 720-254-7893, or grab the Substack on the Welcome tab.