3 Claude Code AI Workflows Nate Herk Uses Every Day (and How to Copy Them)

Key takeaways

• The video editor learned its style by studying 27 of Nate’s best-performing reels, all edited by humans
• Start from proven work (a “golden dataset”), not from what the AI guesses is good
• The Roast Room runs AI versions of well-known business thinkers as sub-agents, each owning one question
• Nate’s Instagram account is about 98–99% automated and reached almost 2,000 followers in 30 days on roughly four to five hours of his time
• Every workflow checks its own work, using screenshots, waveforms and a final fact-check before posting

Most AI demos show a single clever prompt. This one is different. In a new episode of his “Show me your workflow” series, Matt Wolfe sits down with builder Nate Herk, who walks through three AI systems he says he uses every day.

What makes the conversation useful is not the tools themselves but the thinking behind them. Nate repeatedly comes back to the same ideas: copy what has already worked, turn it into a reusable skill, and keep feeding your corrections back in. Below we explain each workflow in plain English, what it does well, where it still falls short, and how you can borrow the ideas for your own business.

What are Nate Herk’s three Claude Code AI workflows?

Nate showed three systems, all built around Claude Code, Anthropic’s AI coding agent that can run on your computer, use files and tools, and follow saved instructions called skills.

How does Claude Code edit videos in your own style?

By studying finished videos that already performed well and turning the patterns into a skill. Nate says earlier attempts at letting AI edit felt “cheap” and often took longer than doing it himself, because every change meant another slow render. What changed was the approach, not just the tools.

He had Claude Code analyse 27 reels that his human editors had made and that had done well. It went through them frame by frame, noting how often the scene changes, when B-roll appears, how the shimmer effects come in and even the relative volume of his voice, music and sound effects. From that it built a bank of patterns, which became a skill.

Now he can drop in a raw recording, mistakes and all, and get back an edit with the same fonts, pacing and three-scene layout his team uses. He says he recently produced about 20 reels in one morning, with the AI handling ideas, scripts, B-roll and AI-generated clips while he read from a teleprompter.

For long YouTube videos the same idea applies. Claude Code gets his face-cam and screen recordings, places them in his two standard scenes, switches to full-screen face-cam when his mouse has been still for a while, and adds small zoom-ins so viewers can read on-screen text.

Why start from your best proven work?

Because AI, left alone, tends to produce the obvious average. Nate points out that a language model predicts the most likely next word, so its default ideas are what most people would assume. If you want results that work, give it examples of what has already worked.

He borrows a term from AI testing: the golden dataset. Teams that deploy AI test it against a set of known-good answers. Nate’s point is simple: why not start with that set in mind when you build? Find something that has been done successfully, break down how it was done, then build towards it. He says this habit now shapes how he makes business decisions too, not just videos.

What makes the AI edit feel human?

It checks its own work with screenshots and sound, and it thinks about the viewer. Matt asked a fair question: Claude cannot watch video directly, so how does it learn? Nate’s answer is that it takes screenshots at regular intervals, something he first baked in when using Claude Code for web design so it could spot elements out of place.

• Word-level timing: he uses the ElevenLabs transcription API because it is quick and cheap, giving start and end times for every word so graphics land exactly as he says them. Free, local Whisper works too but is slower
• Waveform checks: cuts used to clip the ends of words, so the skill now waits until the audio waveform goes flat before cutting
• Pointing: if he points to a spot on screen, the edit places the relevant logo there
• Viewer gaps: it flags places where an explanation may be unclear, so he can re-record that bit
• An AI judge: he is building an “LLM as a judge” that scores edits, so the skill keeps improving

Nate also uses Blender, a free 3D tool, through Claude Code to create 3D scenes, despite never having opened Blender himself. Matt admitted he has made more in Blender in three months than ever before, for the same reason.

What is the Roast Room AI expert panel?

The Roast Room is a panel of AI advisers, each modelled on a well-known business figure, who take positions on your decision, argue with each other and then vote. The panel includes AI versions of people such as Alex and Leila Hormozi, Naval, Codie Sanchez, Sara Blakely, Gary Vee, Paul Graham, Jeff Bezos and Warren Buffett, and Nate keeps adding more.

Under the bonnet it is simpler than it looks. Each adviser is a sub-agent: a short instruction file describing their lens, how they talk, how they debate and what they never do. Each one owns a single question. For the Alex Hormozi agent, it is roughly “will people pay, how much, and how do we close?”

Each adviser also has a “brain” folder built from real sources: their core worldview in three sentences, eight rules they decide by, frameworks, positions and stories. Nate says the Hormozi brain alone drew on about 123 videos, books, articles and posts, adding up to more than 1.5 million words. A separate skill creates a new adviser on request, including a cartoon avatar made with an image model.

Two practical details stand out. First, Nate used to run the panel as an agent team but switched back to sub-agents because the team version cost more for little extra value. Second, the animated room is just a visual layer; the same debate runs in Claude Code without it.

What happened when the panel judged Matt’s newsletter?

The panel recommended rebuilding the newsletter for human readers, while also changing how sponsorships are sold. Matt’s live question was whether his AI newsletter, with about 260,000 subscribers and funded entirely by sponsorships, will stay worthwhile if AI agents end up reading newsletters instead of people.

Nate’s one rule: the panel needs options to vote on. Claude turned the context into five options: keep doing what works, rebuild for humans, resell sponsorships on outcomes, design for agents and sell “agent reach”, or shift effort to other channels. It also swapped two advisers out and brought in Naval, whose question is whether something builds leverage you own or just rents out your time.

After opening positions, one-on-one debates (run one at a time so others can see what was said) and a final vote, the verdict was to rebuild for humans and run outcome-based sponsorship pricing alongside it. The panel’s reasoning was that only the human-focused option tests whether readers actually want Matt, and that pricing sponsors on opens is risky if agents are doing the opening. It put roughly $78,000 a year at risk, but stressed that figure rested on assumed numbers. The result came as a short PDF with scores, risks and an action plan.

Want to try a smaller version yourself, with any AI chatbot? Our guide How to Build an AI Advisory Panel to Test a Business Decision walks you through it.

How does the Instagram autopilot work?

Claude Code turns a link into a researched, branded carousel and posts it, about every four hours. Nate’s AI news account had been running for about a month and was close to 2,000 followers, with Nate estimating he had spent only four to five hours on it in total.

He can send Claude a link to an X post from his phone, even from the gym, and it reads the replies, finds related articles, downloads and slices any video, and builds a carousel using his brand guidelines and tone of voice. In the demo this took around five to six minutes and produced a preview video of the whole carousel before posting.

Posting, multi-account management and comment-to-DM automation run through Blotato, a tool built by Sabrina Ramonov. One giveaway post gathered 559 comments, with people asked to follow the account before the resource is sent by DM. Every post is logged in a Google Sheet so the system does not repeat a topic.

How does it avoid posting fake news?

It double-checks facts twice: once when it researches and again just before posting. Nate notes that X is full of fake news, so the skill confirms stories are real. When he asks it to post, it checks again, because he has seen original posts deleted or edited between drafting and publishing.

Matt suggested a next step: pull performance stats back into the sheet so the system learns which posts work. Nate said he has prepared for this but not fully set it up yet, so treat the “self-improving” part as a plan rather than a finished feature.

How can you set these workflows up yourself?

The easiest route, both agreed, is to point your AI coding tool at Nate’s public GitHub repo and ask it to set things up. Matt’s suggestion was to paste the repo link into Claude Code, Codex or a similar agent and say “build this out for me”. Nate says the repo contains most of the skills shown, and he links to it from his YouTube channel.

1. Pick one workflow, not all three
2. Gather your best examples first, such as top videos, posts or past decisions
3. Install or adapt the skill in a test folder, never on live accounts first
4. Run it on a small batch and write down every correction you make
5. Save those corrections back into the skill so the next run is better
6. Only automate posting once you trust the output

What does this mean for a small business?

The biggest lesson is the method, not the tools. You do not need to edit 20 reels a morning to benefit. The same pattern works for emails, product descriptions or sales pages: collect what has already worked, describe why it worked, turn that into reusable instructions and improve them every time you correct the AI.

• Start from proof: your best-performing content is your golden dataset (see our style guide method)
• Get a second opinion: a simple AI panel can poke holes in a plan before you spend money (see our advisory panel guide)
• Keep a human in the loop where the AI is weakest, such as music, tone and fact-checking
• Choose carefully: pick the workflow with the clearest payoff first, as in How to Choose the Next Business Workflow to Automate With AI

This fits what we believe at Prodify Digital: simple systems you own, built step by step, usually beat chasing every new tool. Your own content library and documented way of working are what make any AI tool useful.

Frequently asked questions

The bottom line

Nate Herk’s Claude Code AI workflows work because they start from proven results, live as reusable skills and get better every time he corrects them. You can borrow that approach today without copying his whole setup: gather your best work, turn it into instructions and keep a human check on anything that goes public.

Related posts

• Claude Skills Clean-Up: Which Skills to Keep, Merge or Delete
• How to Load AI Skills Only When You Need Them, Not Every Session
• How to Choose the Next Business Workflow to Automate With AI

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio