This is adapted from a talk I gave at Smart Marketer Mentor Tables in Nashville. If you'd rather watch than read, hit play above.
Most AI Users Are Bad at AI. I Can Prove It.
I use AI daily. I have used it daily for two years. I recognize AI output every time I see it. And it makes me want to scream.
It feels lazy. If you're responding to me with ChatGPT, I could have just gone and asked ChatGPT myself. That's why AI is so annoying for the regular users of it.
NBC News ran a poll in February. 51% of people say they use AI daily. But only about half of those people actually like it. Everyone's using it, and most of them think it kind of sucks.
Then I saw a tweet that reframed the whole thing for me: "What most people call AI slop is just AI generated content that's not good enough yet."
So maybe the problem isn't the tools. Maybe it's the users. Probably a little of both. But I know for a fact that with better prompting, we can get you better results.
The Future Is Here, It's Just Not Evenly Distributed
There's a William Gibson quote I love from 1993: "The future is already here, it's just not evenly distributed." (He used an em dash in the original. You can't post that online today without people accusing you of being ChatGPT. Punctuation used to mean you were fancy. Now it means you're lazy.)
The pitch with AI was that you'd do less work, produce better output, and then you could go outside, touch grass, spend time with the kids. You know what actually happened? I produce 50% more and work 20% harder. Productivity is way up. Free time is not.
My neighbor Andrew is a high school social studies teacher. He said something that stuck with me: "Kurt, AI works for you because you can string a sentence together. Some of my students can't do that." If you don't have the underlying skills, the tool can't save you. Garbage in, garbage out.
So let me give you the input side. A framework, a system prompt tweak, and the 10 tools I actually use.
The RTF Framework: Role, Task, Format
This is the one thing to take away if you take nothing else. It's called RTF. Role, Task, Format.
Researchers figured out that ChatGPT has roughly 90% of all books stuck in its head. They ripped entire libraries into it. If I put 80 gigs of data into your head, you'd have a nosebleed and need direction on where to go next. AI is the same. It needs you to point it somewhere.
AI loves to role-play. So start with the role:
- Role: Usually a job title. "Your role is Shopify developer." "Your role is expert-level marketer." "Your role is skeptic, challenge me." When I don't want a conversation I'll say "your role is software utility," because a software utility doesn't have opinions.
- Task: The thing you want it to do. AI is a golden retriever. It wants to please you, but you have to tell it how.
- Format: Optional but powerful. "Output a 250-character tweet." "Give me an HTML table." "Return a bulleted list." If you don't specify, it'll guess.
Use RTF on every prompt. Your output gets better immediately, regardless of the tool.
The Promptimizer: Use AI to Write Better Prompts
Here's the hard-to-swallow part: sometimes AI is better at prompt engineering than I am. So for anything I'm saving (a custom GPT, a Claude project, a reusable prompt) I run it through what I call a promptomizer. I read the official documentation from Anthropic and OpenAI on how to prompt their models, synthesized it into a prompt that optimizes prompts, and now I just pass my drafts through it.
For one-off tasks: RTF. For anything you're bookmarking and reusing: RTF plus promptomizer.
Get it here: The Prompt Engineer Prompt
Stop Letting ChatGPT Tell You You're Absolutely Right
Right now the dumbest person you know is being told they're absolutely right by ChatGPT. Two-year-old tweet. Still true.
Why? Because these companies optimize for time on site. You're the product. If you're not using the product, they can't raise their next round. So ChatGPT doesn't want to disagree with you. Ever. Unless you're asking about something illegal, it'll back whatever you say.
Worst tell: ask it something subjective ("should I take this job?"), get an answer, then say "are you sure?" About half the time it'll reverse itself. It's not reasoning. It's appeasing.
Fix it in your system settings. Keep them short, but add these three rules:
- It's okay to say "I don't know."
- Push back on me.
- Give direct critiques when ideas are flawed.
I recently told Claude I wanted to spend $10,000 on Apple Mac Studios. It pushed back. It was right. I still bought them.
The 10 AI Tools I Actually Use
Before the list, one warning: avoid AI wrappers. An AI wrapper is software that's just a fancy UI for a prompt running on top of Claude or ChatGPT. The tell is credits-based pricing. If you're paying for credits, you're paying a markup on tokens you could buy directly. None of the tools below are wrappers.
10. Live Chat (Shopify Inbox, Crisp, Tidio)
This is more category than tool. Customers want to solve their own problems if you'll let them. Returns, order edits, and "where's my order?" are 90% of support tickets. An AI chat tool gets them started. We use Crisp in our apps. I've heard good things about Tidio. If you're on Shopify you already have Shopify Inbox.
9. Microsoft Clarity
Clarity is free. I have no idea why. Microsoft clearly wants to be Google Analytics, and they're giving the tool away to get there. You get heatmaps, session recordings, and now AI analysis baked in. Export the AI summary, drop it into Claude with your brand context, and ask it to identify PDP optimization opportunities. Works incredibly well.
8. Topaz Labs
Got B-grade UGC? Topaz fixes it. Photo or video. Upscaling, sharpening, cleanup. Topaz Video takes hours to process but it can rescue a mediocre webcam recording. Both easy to use and professional-grade, which is a rare combination.
7. Adobe Enhance
Audio is table stakes. Good audio doesn't help. Bad audio definitely hurts. Adobe Enhance is drag-and-drop and free. One tip: set the strength to 40%. Past 50% it starts sounding auto-tuned and robotic.
6. WhisperFlow
WhisperFlow lets you talk to your computer. I did their typing test. It says I type 80 wpm and talk at 105. I'm offended at the 80 but sure, fine. The real use case isn't speed, it's getting unstuck. When I need to go from zero to one on an idea, I brain-dump into WhisperFlow and then iterate on the transcript. Typing and talking use different parts of your brain. Talk first, edit second.
5. Whisper Transcription
LLMs are built for text. Give one an image or video and behind the scenes it's calling a separate tool to transcribe it. So transcribe first, save the context. Whisper Transcription runs locally on your Mac. Nothing leaves your computer. Meeting notes, YouTube videos, talks, voice memos from your car, all of it becomes a text file you can feed to Claude. I use this several times a week.
4. Adobe Firefly (and Nano Banana)
Image models come and go constantly. My current favorite underlying model is Google's Nano Banana (genuinely, that's the name). But I access it through Adobe Firefly, which gives you a clean sandbox with multiple models including Adobe's own, which is trained only on commercially licensed stock. Less copyright anxiety if the output ever makes it into client work.
3. Grok
I need a backup chatbot because Claude goes down more than I'd like as they scale. So ChatGPT is my backup. And Grok is the backup to the backup because it's free and it does image and video generation. OpenAI killed Sora last week (it was reportedly costing them a million dollars a day), so Grok picked up that slack. Grok is also basically unmoderated, which is occasionally useful. I tried to make a Newman's Own parody label for my home-grown weed. ChatGPT refused. Grok rolled up its sleeves.
2. Claude Code
Claude Code is amazing if you're a nerd. Command-line interface, no mouse, ASCII art. My teenager walked in while I was using it and said "oh my god you're hacking." This is what drove the last year of public conversation around AI. Marketers shouldn't have to touch it, though. Which brings us to number one.
1. Claude Cowork
This is the one. All the power of Claude Code, but in the desktop app instead of a terminal window. It's the tool I'm actually excited about this year, and the reason is two words: agentic, and folders.
Why Working in Folders Changes Everything
Agentic just means autonomous. A regular chatbot is only "awake" during the moment it's processing your message. Agentic AI can run on a schedule, spin up sub-tasks, and hand work off to other instances of itself. That's the headline.
The real power is that it works in folders on your computer. That sounds boring. It's not.
Here's the problem with a normal chatbot: every time it responds, it re-processes the entire conversation. It doesn't actually remember anything. Your working memory is about seven items (that's why phone numbers are seven digits). Claude Cowork's context window is a million tokens, which sounds huge, but a token is two or three characters. You burn through it fast.
When it runs out of room, it summarizes the conversation and throws the rest away. That's context rot. It's why a session that started great suddenly gets dumb. The model forgot the stuff that made it smart.
Folders solve this. Instead of stuffing everything into the conversation, you drop files into a folder: CSVs, PDFs, text, images. Claude reads them on demand. Your context stays clean.
Better: drop a file called claude.md into the folder. Claude reads it every time, without fail. That file becomes persistent instructions. I have one for Matrixify work. I have one for my command center folder where I plan projects. When I'm wrapping up a session, I tell Claude to make notes for itself. A month later I zip up the folder, hand it to a teammate, and they pick up exactly where I left off.
You can go beyond folders. Claude connects to Slack (where my team talks), Google Drive (where our docs live), and Shopify directly. That last one scares me a little. Run backups. There are plenty of horror stories of Claude cheerfully deleting production data. Keep a human in the loop.
One caveat: every integration burns context. If you load up 20MB of spreadsheets plus 10 tools, you've blown the budget before Claude can do any real thinking. Be selective.
What to Actually Use It For
Once you're working in folders, the use cases open up:
- Product catalog cleanup. Export from Shopify, ask what SEO fields are missing, standardize product types.
- Email campaign builds. HTML and text, easy for Claude. Drop the analytics CSV in the same folder.
- Customer research. Dump all your product reviews and support tickets into a folder. Find common pains and unmet objections. Then cross-reference against your product descriptions.
- Content calendar and social. I exported all my past tweets and LinkedIn posts with analytics. Claude built me templates from the high-performers and told me when to post.
- Financial and inventory analysis. Anything you can export as a CSV, Claude can analyze and visualize.
Recap
The takeaways:
- RTF on every prompt. Role, Task, Format.
- Promptomize anything reusable. Let AI clean up prompts you're going to use more than once.
- Fix your system settings. Give it permission to say "I don't know," push back, and critique flawed ideas.
-
If you try one new tool, make it Claude Cowork. Work in folders, drop in a
claude.md, stop fighting context rot.
A year ago these tools couldn't do most of this. Today they can, but they still need a human doing the strategic thinking and double-checking the output. Two years from now? Probably different. Don't chase the latest and greatest. Pick the tools you're curious about, play with them in your personal life before you unleash them on a client, and get comfortable. That's it. That's the whole game.
Sponsors
- Swym — Wishlists, Back in Stock alerts, & more
- Cleverific — Smart order editing for Shopify
- Zipify — Build high-converting sales funnels
