13:10Austin Marchese · Talking Head
Stop Prompting Claude. Start Graph Engineering Instead.
A Stanford-adjunct framework for treating every Claude Code workflow as a graph you can redesign, not a prompt you keep tweaking.
August 20thThe argument in one line.
The builders getting the most from AI aren't optimizing prompts, they're building a personal domain harness: rules, skills, context, and an interface wrapped around a base tool like Claude Code.
Read if. Skip if.
READ IF YOU ARE…
- You already use Claude Code or a similar AI tool daily and want more out of it than better prompts.
- You have deep expertise in one specific domain, marketing, sales, ops, and want to encode that expertise into a reusable AI system.
- You're comfortable directing Claude Code conversationally to build a small internal tool, even without a traditional dev background.
SKIP IF…
- You're looking for prompt-engineering tips, not a system-building approach. The video deliberately skips prompting advice.
- You want a finished product to buy, not a framework to build yourself inside Claude Code.
The full version, fast.
The video argues the real AI advantage isn't the best model or the best prompt, it's the personal harness wrapped around a base tool like Claude Code: rules, skills, MCP connections, context, hooks, and an interface. It breaks this into four layers: understand why a harness matters, pick the one domain you already know, build the harness by extending existing tools instead of the model's raw API, then design the interface, which it argues is the most overlooked layer of all. It walks through building a real example, an internal knowledge-base harness, using Claude Code's rules and skills plus a simple local web interface, and closes by arguing the long-term winners won't be whoever has the biggest model but whoever figures out the most valuable thing to build on top of it.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:30
01 · What is harness engineering?
Cold open stating the thesis: the fastest AI builders aren't optimizing prompts, they're building a personal harness around the model.
00:30 – 02:52
02 · Layer 1: the foundation
Introduces the AI output equation (model + prompt) x base harness, and argues the domain harness added on top is the one layer you actually own.
02:52 – 05:07
03 · Layer 2: select your domain
Applies McKinsey's makers/takers/shapers framework and argues anyone can out-build Zuckerberg or Musk at a harness built for just one user: themselves.
05:07 – 07:01
04 · Sponsor break: Anthropic's unified Claude workspace
Sponsor segment covering Claude Cowork and Chat merging into one tool, Claude Design inside conversations, and shared context across chats.
07:01 – 09:10
05 · Layer 3: build the harness, don't reinvent it
Lays out two build rules, don't reinvent and keep it simple, and the eight parts of any harness: rules, skills, MCP servers, context, hooks, subagents, interface, feedback.
09:10 – 11:57
06 · Worked example: build an internal knowledge harness
Walks through building a real example harness, an internal knowledge base that only answers from documents you feed it, using one structured prompt inside Claude Code.
11:57 – 12:31
07 · The anti-slop agreement and Claude Max giveaway
Channel disclosure segment plus a subscribe ask and a Claude Max subscription giveaway.
12:31 – 14:54
08 · Layer 4: the interface
Argues the interface, not the tooling underneath, is usually what makes one AI product feel better than another, and explains why a local web interface beats cloud-based alternatives.
14:54 – 15:49
09 · The Karpathy Curve
Recaps the four layers, then explains the Karpathy Curve: a future where the model matters less than data curation, task decomposition, and organizational adaptation.
15:49 – 16:19
10 · Outro: matching the harness to the use case
Points to a follow-up video comparing Claude and Grok Bot across 12 daily use cases, for viewers who'd rather use an existing product than build their own harness.
Lines worth screenshotting.
- Your AI output isn't just model plus prompt, it's (model + prompt) multiplied by the harness wrapped around them, and most of that multiplier goes unclaimed.
- A domain harness is the one layer of the AI stack you can actually own, because frontier labs will never build it for your specific job.
- McKinsey's 2024 gen AI research splits adopters into makers, takers, and shapers, and the shapers who customize AI around their own workflow report the outsized returns.
- You don't need to out-build Anthropic or OpenAI to win with AI, you just need to design for one user, yourself, instead of thousands of customers.
- The first rule of building a harness is don't reinvent what Claude Code already does at 90 to 95 percent quality.
- The second rule is keep it simple: borrow existing tooling and extend it, rather than hand-rolling a custom system against a model's raw API.
- A harness is made of eight parts: rules, skills, MCP servers, context, hooks, subagents, an interface, and a feedback mechanism.
- Running a local web interface against Claude Code lets you use your existing Claude subscription instead of paying separately for API credits.
- The interface layer, not the tools underneath it, is usually what makes one AI product feel better than another with identical capabilities.
- The Karpathy Curve reframes AI progress: the model becomes the least interesting part of the stack, and the real work shifts to data curation, task decomposition, and organizational adaptation.
- The AI-doomer fear assumes a model smart enough to solve everything leaves no room for human value; the harness-engineering view says the value just moves to whoever shapes the system around the model.
Your harness matters more than your prompt.
WHAT TO LEARN
Winning with AI now comes down to the system you build around a tool like Claude Code, not the model you pick or how cleverly you prompt it.
02Layer 1: the foundation
- Meta's Muse and xAI's Grok Bot both took off because they wrap AI models in an easier interface, not because the underlying models changed.
- Your AI output is (model + prompt) multiplied by the base harness, and most people never touch that multiplier.
- The domain harness, the layer you build on top of a tool like Claude Code, is the only part of that equation you actually own.
03Layer 2: select your domain
- A 2024 McKinsey report splits AI adopters into makers (build the model), takers (use the vendor's tool as-is), and shapers (customize it around their own data).
- Shapers are the group reporting outsized returns, because they adapt AI to one specific, already-understood domain instead of using it generically.
- Pick the domain you already work in, marketing, sales, ops, and build toward that instead of a harness that tries to do everything.
- You don't need to out-build Zuckerberg or Musk here, because you're designing for one user, yourself, which no enterprise product can match.
05Layer 3: build the harness, don't reinvent it
- Rule one of building a harness: don't reinvent what Claude Code already does at 90 to 95 percent quality.
- Rule two: keep it simple, extend existing tooling instead of building straight against a model's raw API from scratch.
- A complete harness has eight parts: rules, skills, MCP servers, context, hooks, subagents, an interface, and a feedback mechanism.
06Worked example: build an internal knowledge harness
- The example harness is an internal knowledge base that only answers from documents you feed it, instead of guessing from general training data.
- It's built from two reusable skills, one that answers only from local context and one that syncs in outside knowledge from tools like Google Drive or Notion.
- The entire harness, rules, skills, MCP connection, and interface, gets built by handing Claude Code one structured prompt and letting it walk through setup.
- A thumbs up/down button on every answer is what lets the harness get sharper the more you use it.
08Layer 4: the interface
- The interface is the most overlooked layer of a harness, and often the real reason one AI tool feels better than another with identical capabilities underneath.
- A local web interface run against Claude Code uses your existing subscription instead of billing separate API credits for every request.
- Running the harness locally also lets you iterate fast: click what you want changed and have Claude Code update the interface in real time.
09The Karpathy Curve
- The AI-doomer fear is that a sufficiently smart model leaves no room for human value; the Karpathy Curve argues the opposite.
- Future advantage shifts away from model size and benchmark scores toward data curation, task decomposition, and organizational adaptation around the model.
- The eventual winners won't be whoever has the biggest model, they'll be whoever figures out the most valuable thing to point it at.
Terms worth knowing.
- Base harness
- The off-the-shelf tool that wraps an AI model for you, such as Claude Code, before you customize anything yourself.
- Domain harness
- The custom layer of rules, skills, and context you build on top of a base harness to make AI work specifically for your job or workflow.
- Makers, takers, shapers
- A framework from a 2024 McKinsey report describing three ways people use generative AI: building models from scratch, using vendor tools as-is, or customizing tools around proprietary data.
- MCP server
- A connection that wires an AI tool to an external data source or service, such as Google Drive or Notion, so it can pull in outside information.
- Karpathy Curve
- A concept, attributed to former Tesla AI lead Andrej Karpathy, suggesting future AI progress depends less on model size and more on data curation, task decomposition, and organizational adaptation.
Things they pointed at.
05:07toolAnthropic Claude (video sponsor)
06:05toolClaude Docs and Slides
03:13linkMcKinsey: The takers, shapers, and makers of gen AI
Lines you could clip.
00:59
“It is not about who has access to the best models. It is about who can shape them.”
07:03
“The simplest solution that solves the problem is better than any complex solution, especially when you're first starting.”
15:09
“The winners aren't the ones who built the biggest brain, they're the ones who figured out the most valuable thing to think about.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
I've spent the last four months studying how the fastest AI builders actually work, and I found something that I wasn't expecting. The people who are actually winning with AI aren't thinking about the model. Instead, they're focused on the harness that interacts with AI.
So in this video, I'm going to break down and outline what harness engineering is and how you can make your first harness today, which can be broken down into four layers. And at the end, I'll outline what the Karpathy curve is, a concept described by the former head of AI at Tesla, and how it reframed how I think about the future.
So layer one. the foundation. why this matters so much.
So recently, Meta and Mark Zuckerberg released Muse, while SpaceX and Elon released Grokbot. And both these tools are harnesses that simplify how you interact with AI. And they've gotten so much traction because of how easy these harnesses make it to use AI for daily tasks.
Here's Peter Levels, a famous entrepreneur, talking about the concept of how important harnesses are in reference to Y Combinator, one of the biggest startup investors on the planet. He said, I saw this Y Combinator's batch literally only has harness startups and hardware. startups.
So . So why is the harness so important? Well, in today's world, it's not about who has access to the best models.
It's about who can shape them to provide the most value for a specific use case. So clearly the smartest people are focusing on this, but should you? The answer is obviously yes.
That's why I'm making this video, but let's break down in terms of how you actually use AI today. Think about your AI output like an equation. Your AI output is equal to model plus prompt multiplied by the base harness.
And in practice, in your day -to -day, when you use cloud code, which is a base harness. The output that you produce is a combination of the model, call it Opus, the prompt, what you say to it, and the base harness, which you chose to be Clawed Code.
In plain English, the harness is whatever wraps around the model and the input you provide it to improve the actual output you get. And tools like Clawed Code, Muse, Grokpot, they all wrap the AI model to make interacting with it easier. And I'm saying this term wrapper because a lot of people have heard the term an AI wrapper, and that's really no different than an AI harness.
And the one question you have is i like using cloud code i don't want to recreate it and that's totally fine because there's a finer point here so let's go back to the equation i mentioned 99 of people like using ai like this right that's your model plus prompt multiplied by base harness that's straightforward but the leaders in this space are thinking about it like this its output is equal to model plus prompt multiplied by base harness plus domain hardness where the domain harness is what you can build on top of tools like claude and claude code and anyone watching this video can build their own hardest because the actual advantage is deep specific knowledge of a person's exact workflow yours the most productive people i know have effectively built their own harness hyper specific to their job and their day -to -day without even thinking about it or even calling it that so this domain harness is the one thing you actually own here the next question is obvious where do you even start and where should you focus your energy which brings us to layer two select the domain at a foundational level a harness is a way to alter ai's output so it works specific for your situation so naturally you have to pick which domain to help it with and this will naturally lead to two different types of people they'll be the person that wants the harness that does everything or they're the type of person that they can't figure out
what they want to focus on and for both of these people the answer is actually the same and so this pulls from a concept popularized by a mckinsey report where it outlined three types of people there's makers takers and shapers a maker is someone that builds the model itself that's open ai anthropic google and almost nobody else should try and do that that isn't you takers are the people that use ai the way the vendor built it so that's people that take the products that are out there and just pull in value to their systems this is fast easy and relatively affordable but there's no difference between you and anybody else who are using these tools.
And the third is shapers, which build their own layer on top of someone else's model. This is essentially the process of harness engineering. And the key here is the term shaper.
You have to be able to shape something you actually do. And in order to shape, you need the confidence and the domain expertise to have precision in that specific domain. And this mindset of being a shaper, for people who fall in the category of do everything, this will help them be more precise because they have to shape specific things.
And for those who overthink, you'll just think about shaping what you already do don't overthink it just focus on where you spend your energy right now so if you're in marketing focus on shaping ai to get better at marketing if you're in sales focus on shaping ai to be better at sales and focus on specific things that you're an expert at already so that you can better tweak and shape ai to get better outputs and so you're thinking about the domain that you want to work in and one limiting belief that you may have is why would you be better at building a harness than mr zuckerberg or mr elon musk there's a number of reasons but primarily you get to design for one user not thousands not a whole department just you and that's value no company building a product for thousands of customers will ever provide so you can build something that's more valuable for yourself than these enterprise companies now if you're still not sure what to focus on and what domain you want to build your harness in here's a prompt that you can run that will help you better identify the niche or domain that you should focus on so now that you've selected the domain that you want to focus on
It brings us to layer three, which is building the harness. But before we get to that, one harness that I've mentioned that I know everybody has used is Cloud Code, which brings us to today's video sponsor, Anthropic, the team behind Cloud, which you guys know that I love their tooling and I'm hyped to be working with. Now, there are a couple of product features that they launched that are important to understand as you position yourself to be productive with AI.
Cloud Cowork and chat have now been combined into the same tool. And the reason that I like this doesn't necessarily have to do with the specific features, although I will cover that in a second. But instead, I say this all the time.
Moving the decision of what tool to use makes you more productive because it lets your brain focus on high leverage decisions. And this is cloud step in that same direction. And this is also an example of them creating a more effective harness for you to use exactly what we're talking about in this video.
And by combining chat and cowork, you don't have to think about which one you have to use because they're the same thing. Now, the second is you guys know that I love cloud design and now it works right inside your conversations. So you don't have to context switch to a different tool.
Not only that they've. launch Claude docs and slides, which are specifically designed for those use cases. See, one of my clients recently asked me, how should I use AI to create proposals that I'm just struggling going back and forth?
And so I told him, set up Claude connections to wherever relevant conversations are about the proposal. Take that and bring in a reference of a past proposal and then tell Claude, make me a template proposal that I can use for future clients. And then say, make a proposal for this client using context from our connections.
And I'll then go from there, creating a proposal template. And all of this back force lives in the same interface which removes context switching that i mentioned earlier and the final feature in this launch that i like is now all chat conversations share one context your connected skills and projects they're accessible in any conversation so now you don't have siloed projects with context it's all one thing now this specific feature will be rolling out to pro and max users automatically so to make sure you don't miss it check the first link in the description so coming back to layer three which is building your harness the first rule layer three Don't reinvent things.
Rarely do you actually want to rebuild something from scratch. And most of the time, the highest and best use of your time is extending what's already there. So I don't want you to spend the weekend hand -building a harness from zero only to realize that Claude already does 90 to 95 % of what you needed.
Going back to our equation before, right? We're not building the base harness. We're just building the domain harness.
All right, so enough of that. What actually makes up a harness? Well, there are a handful of variables.
Let's just go through them one at a time. First, rules. These are direct instructions.
for how AI should behave. Skills are reusable capabilities you build once, and then you can call them whenever you need them. MCP servers, they lay the wiring to talk to other tools.
These are basically external connections. There's context. This is the actual knowledge you feed it specific to you.
And if you're trying to scale this out, a knowledge graph helps a lot here. Then there's hooks. These are scripts that run on specific events, such as a block and command or running a test after an edit.
There are sub -agents, which are how you delegate work. And this is how you can partition tasks to streamline. the output the interface which is how you actually interact with the whole thing and we're going to do a deep dive on that in the next layer but at a high level you can run a harness through a chat window through documents or my personal favorite is a simple dashboard that you can run locally and the final step to any good harness is a feedback mechanism this lets you shape the harness over time which is so important again from earlier the harness wraps the existing functionality of ai to force it to do something else and if a lot of what i just said sounds like a lot of of what you're already doing, that is the point.
This concept is actually quite simple and you're likely doing a lot of these things. Anyway, you have the foundational skillset, but it's this mindset of harness engineering that we're building. So the first rule is don't reinvent the wheel.
And the second rule is keep it simple. So a complex harness would be an example where you go straight to the models API and you build everything from scratch. A simple harness means borrowing what Claude code already gives you and building on top of that.
Instead, the simplest solution that actually solves the problem is better than any complex solution, especially when. first starting and 99 of people watching this should just start with a simple version build it end to end and then go from there so for this video we're going to build a harness that is both very helpful and also shows the power of harnesses so what we'll build is an internal knowledge harness where it only answers questions based on the context you provide it and we're intentionally choosing this harness because it's valuable for everyone watching this so for example let's say you're at your job and you're to ask claude a question about how to onboard a client it will pull from everything it's training data it might get you the answer it might not But using this harness, it will restrict the information based on knowledge you provided.
So if you were to ask it a question, it'll just pull from the documents you provided to provide the answer. You could see how this is altering how AI responds to you. You're building a harness or a wrapper around it.
So let's go through all of the components that I mentioned earlier. So the rules for this, we want to only pull context from the project we're working on, the skills. We're going to create two reusable skills, one called KB answer that only pulls its answer from local context provided, and then KB sync, which will use external provider.
providers to bring in context to help answer those questions. For MCP servers, we're going to connect to Google Drive, Notion, or wherever your current knowledge base is, and this will help us sync the data. For context, this is all going to be specific knowledge that you want AI to pull from.
This is the part of the harness that is specific to you. The hooks and sub -agents, we're going to use Cloud Code to decide when this is needed. For the interface, we're going to create a simple chat interface that is a one -page website where when you hit enter in the input, it'll answer specifically using this KB answer skill.
And for the feedback, mechanism on the interface itself, we're going to create an easy way to provide feedback so that if you're missing context, it's easy to know and then add it to the system. So to build all this, I'm going to package all these features up into this prompt that creates everything I listed above with the interface being a simple single page website.
And now if you are starting from scratch, go to cloud desktop app, hit code in the top right, select open folder, select new folder, and just name it harness demo. Then once you do that, your system's all set up, run this prompt, which will walk you through creating the harness. Now, if you have questions on this prompt or you want to go through it more step by step, I cover each step in detail and how to build it in my buildpartner .ai plugin, which is linked below for free.
You just run slash buildpartner colon build space harness, and it'll automatically run you through it step by step. But the prompt I shared here is enough to get you going. Once you run it, you just let AI do its thing.
And the interface for the harness will look a little bit like this with two buttons. The sync button will run KB sync skill to get additional context. And then the send button will.
trigger the KB answer skill on your machine. And then after you receive the response, there's a feedback mechanism for you to say if the response was good. And if it wasn't, you can use the tool to ingest additional data.
Over time, you can imagine training and sharpening this harness. Now, this general structure is what a lot of these smartest people are doing, and they're creating workspaces that let them build 10 times faster. So at this point, we have the architecture set up.
But before we get to the last layer, as well as the Karpathy curve, which changed how I think about all of this, you know the drill. This is our anti -slap agreement. Everything here is built by humans for humans.
So as part of this agreement, subscribe to this content so it can keep reaching more people. And if you already subscribe, make sure to sign up for post notifications. It helps my channel and I appreciate every one of you.
These drawings and my terrible handwriting, this is for you guys. And as a thank you to everyone watching, I give away a Claude Max subscription every single video. And this video's winner is JeremyGlen87, who is working on various side projects.
To enter this video's Claude Max giveaway, let me know what you're building in the comments and let's... chat about any problems you're running into, I'd love to help. Now let's get into layer four, the interface.
This is a layer that most people completely neglect. And honestly, in my eyes, it's actually the most important of the four layers. This is how you actually visualize the tools that you're building.
And so the harness isn't just the tools underneath it. It's the whole experience. And that includes how a human actually uses this thing.
So I recently made a video comparing Claude and Grokbot, which I'll link at the end of this video. But the reality is that Claude and Grokbot can essentially do all of the same things, but the interface makes one better than the other at specific use cases. It has nothing to do with the model, nothing to do with the underlying guardrails or skills, but everything with how a user actually interacts with it.
This is the key to harness engineering. At the end of the day, who cares if you have an amazing tool set, if it's just hard and complicated to use it. And so for you, there are a number of interfaces for you to consider.
In this video, we outlined a webpage, which may feel the most typical to what you're used to. This is one way, but another way could be through a terminal or a command line interface. And another could be through custom plugins.
And the reason that I love the web -based local interface is one, you likely already pay for subscriptions like Claude. And if you run the app locally, it can just use your Claude membership instead of having to pay API credits for a typical cloud -based app. The second is that you can iterate extremely fast.
On my internal tool, I actually have a pencil icon that lets me hover over things that I want to change. And I just click it. I type what I want.
want and then i hit enter for example on screen i show me selecting a data source and i just say i want to add outlook and then it goes and cooks on it and this process of just like quickly iterating is something i do all the time now i'll do this on a call but me and a team member will just be like i wish response was like this and i'll just instantly change it speeding up this iterative cycle so the harness can get better faster the third is that there's a clear pathway to scale it to your team now this can be a bit more complex but when you're ready you can build a tool that scales out to other people on your team now for the this video we won't cover how to scale this harness to your broader team but if you are interested in that there is a link below where we can talk through how to set it up for your team now before we go through the carpathia curve which changed how i fundamentally think about ai we just went through the four layers of harness engineering layer one understanding the foundation layer two identify the domain
Layer three is building the actual harness, and layer four is focusing on the interface to get the most out of these tools. We have an exact idea in terms of how to build one that's valuable for you today, as well as creating your own harness through the prompts that I shared. But now we need to cover the Karpathy curve, especially in a world where AI doomerism is just so prevalent.
So simply put, what's the Karpathy curve? The Karpathy curve suggests a different future, one where the model is the least interesting part of the stack, where the real work is in data curation, task decomposition and organizational adaptation, and where the winners aren't the ones who built the biggest brain, but the ones who figured out the most valuable thing to think about.
That future explanation is different than the one that we've grown accustomed to hearing. See, the AI doomer vision is a future where AI rules the world and is smart enough to solve any problem we give it, and there's really nothing that you can provide value on. But instead, this explanation, this Karpathy curve, suggests that the winners will be the people who create a harness and a system to best interact with the models based on their specific situation.
And this process of harness engineering, you now have a foundational understanding of when it's needed, how to build it, and how it applies to you. And so we've created a strong foundation, but the reality is you shouldn't build your own custom harness for everything.
So a natural problem you'll run into is which use cases to use which harness for. So for that reason, I made this video, which is a deep dive on 12 daily use cases and breaks down which harness, either Grokbot or Claude, is better for that specific job. This video broke down making your own, but sometimes you want to use the existing products out there.
And so if you like this video, you're going to love that, and I'll see you over there. Peace.
The Hook
The bait, then the rug-pull.
Most people are still trying to out-prompt everyone else. This breaks down why that's the wrong fight: the people actually winning with AI stopped optimizing prompts and started building a personal harness, rules, skills, and an interface wrapped around a tool like Claude Code, shaped around their exact job.
Named ideas worth stealing.
01:28model
The AI output equation
Output = (Model + Prompt) x Base Harness, extended to x (Base Harness + Domain Harness) for people who customize their tools.
Steal forexplaining why tooling matters as much as the model in any AI-strategy content
03:13model
Makers, Takers, Shapers
- Makers
- Takers
- Shapers
McKinsey's three-way split of how organizations adopt generative AI, from building models to customizing them around proprietary data.
Steal forpositioning content around AI adoption maturity
00:00list
Four layers of harness engineering
- Layer 1: Foundation
- Layer 2: Domain
- Layer 3: Building
- Layer 4: Interface
The video's own structure for building a personal AI harness.
Steal forstructuring any 'build your own AI tool' tutorial
07:27list
Eight components of a harness
- Rules
- Skills
- MCP servers
- Context
- Hooks
- Subagents
- Interface
- Feedback mechanism
The building blocks that make up any AI harness, according to the video.
Steal fora checklist for scoping any internal AI tool
How they asked for the click.
VERBAL ASK
15:49next-video
“if you like this video, you're going to love that, and I'll see you over there”
soft close pointing to a related video comparing Claude and Grok Bot across 12 use cases, layered on top of a mid-roll sponsor read for Anthropic rather than a hard sell.
MENTIONED ON CAMERA
Visual structure at a glance.
