10:07Austin Marchese · Tutorial
How Claude Code's Creator ACTUALLY Automates his work
A 10-minute reverse-engineering of Boris Cherny's skill selection system, agent strategy, and the discipline that keeps his setup lean.
April 28thThe argument in one line.
Six specific phrases typed into Claude Code — from launch sub agents to automate this — collapse output ambiguity, enable parallel execution, and create a self-verifying build loop that ships cleaner work without additional tooling.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code daily and keep hitting the same ceiling: wrong outputs, slow sequential builds, or repeated manual work.
- You work solo or in a small team and want to extract more from each Claude session without adding subscriptions or complexity.
- You have shipped at least one project with Claude and want a structured system rather than ad-hoc prompting.
SKIP IF…
- You have never opened a terminal — the video assumes Claude Code familiarity and skips onboarding entirely.
- You are looking for a broad AI overview; every phrase here is Claude Code-specific.
The full version, fast.
Most Claude Code users run one session at a time, skip implementation specs, and automate tasks without any quality filter — all three habits compound into slow, wrong output. This video teaches six corrective phrases in sequence: launch sub agents enables parallel execution; write me an implementation spec collapses the combination space before a line is written; interview me extracts the spec through questions rather than assumptions; verify before you build closes the feedback loop so Claude can self-correct; build me a skill packages repeated work into reusable instructions; and automate this is treated as a warning, not a default, filtered through a taste test and 80/20 output analysis to avoid operational debt.
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 – 02:33
01 · Phrase 1: Launch Sub Agents
Sequential Claude sessions are a single-lane highway. Three sub-agent use cases: multiple perspectives, new possibilities, and pure speed. Exact phrase and how lazy you can be with it.
02:33 – 05:35
02 · Phrase 2: Write Me an Implementation Spec
Plan mode helps but does not replace a spec. The 5^N combinations argument. Prompt that forces Claude to reveal key decisions before building anything.
05:41 – 08:31
03 · Phrase 3: Interview Me
Claude asks questions you did not know to ask, then synthesizes answers into a spec. Blank page to complete spec in under five minutes. Trap: only works if you think before answering.
08:31 – 10:44
04 · Phrase 4: Verify Before You Build
Three-layer approach: CLAUDE.md bias line, tool selection matched to what is being built, human validation zones by cost-of-error. Boris Cherny quote on 2-3x quality improvement.
10:44 – 12:57
05 · Phrase 5: Build Me a Skill
Never build skills abstractly. Only package work you just did manually. Gotchas section from Anthropic's internal playbook makes skills compound over time.
12:57 – 15:12
06 · Phrase 6: Automate This (with caution)
The most powerful and most dangerous phrase. Taste test and 80/20 output analysis as filters. Distinction between automation and augmentation. Warning against creating operational debt.
Lines worth screenshotting.
- Claude defaults to sequential, single-session execution even when parallel sub agents would finish the same work 5x faster.
- Typing 'launch sub agents' forces parallel execution; Claude defaults to sequential unless told otherwise, which is the single biggest throughput bottleneck.
- With 3 build steps and 5 options each, there are 125 possible outputs without a spec and exactly 1 with one — the math makes writing specs non-negotiable.
- Plan mode is a spec aid, not a spec replacement — Andrej Karpathy explicitly said he would rather work with the agent to write a detailed spec than rely on plan mode alone.
- The 'interview me' phrase inverts spec-writing: Claude asks the questions you did not know to ask, then synthesizes your answers into a complete implementation spec in under five minutes.
- Embedding one bias line in CLAUDE.md — asking Claude to state its verification plan before doing any work — changes behavior across every future session in that project.
- Boris Cherny found that giving Claude a tool to see its own output and then telling Claude about that tool produces 2 to 3x better final quality.
- Skills built from abstract brainstorming rarely get used; skills built from conversations you just had have already-validated use cases.
- A gotchas section in a skill captures edge cases and stylistic quirks from real use, tightening the feedback loop so the skill compounds faster.
- The taste test is a binary: if judging output quality requires human taste, the task must be augmented, not automated.
- Running the 80/20 output analysis before committing to automation prevents the single biggest AI productivity trap: operational debt from non-ideal automations.
- Most YouTube AI productivity advice assumes the ideal scenario where automations work perfectly; the realistic yield is far lower, and each broken automation adds permanent management overhead.
- Five sub agents analyzing the same input in parallel produce more diverse output than asking the same question five times in one chat, because a single chat anchors on its prior responses.
Six phrases that wire Claude into a self-correcting build system.
WHAT TO LEARN
Tacking specific trigger phrases onto your prompts changes Claude from a sequential task-runner into a parallel, spec-driven, self-verifying agent without adding any new tools.
- Claude defaults to sequential single-session execution even when parallel sub agents would handle the same work faster — typing 'launch sub agents' overrides that default.
- With three build steps and five options each, there are 125 possible outputs without a spec and one with it; the math makes spec-writing non-negotiable regardless of model quality.
- 'Interview me' inverts the spec-writing burden: Claude asks the questions you did not know to ask, then synthesizes your answers into a complete implementation spec in under five minutes.
- Adding one line to CLAUDE.md asking Claude to state its verification plan before doing any work changes behavior across every future session in that project at zero marginal cost.
- Boris Cherny found that giving Claude a tool to see its own output — and then telling Claude about that tool — produces 2 to 3x better final quality through self-correction.
- Skills built from abstract brainstorming rarely get used; skills built from conversations you just had have already-validated use cases, because you literally just did the work manually.
- A gotchas section in a skill captures edge cases and stylistic corrections from real use, tightening the feedback loop so the skill compounds and makes fewer errors over time.
- The taste test is binary: if judging output quality requires human taste, augment the process; if quality is entirely quantifiable, it is a genuine automation candidate.
- The 80/20 output analysis asks whether 80% quality is acceptable for this specific task — maintaining explicit lists for both categories prevents defaulting everything to automation.
- Every non-ideal automation adds permanent operational debt to manage; the realistic yield on AI automations is far below what YouTube productivity content implies, so filtering before committing is the critical move.
Terms worth knowing.
- Sub agents
- Separate Claude sessions launched in parallel from a single parent session, each with its own context window, prompt, and permissions — they do not see each other's outputs.
- Implementation spec
- A document written before any code is generated that outlines each build step and the key decisions at each step, collapsing ambiguous output space to a single intended path.
- CLAUDE.md
- A project-level configuration file that Claude Code reads and inserts at the start of every prompt in that project, making it the primary lever for persistent behavior changes.
- Human validation zone
- A task or system component where the cost of an AI error is high enough to require a human sign-off before any change is accepted, as opposed to low-cost zones where fast iteration is acceptable.
- Taste test
- A filter that asks whether evaluating the quality of a task's output requires human judgment; if yes, the task should be augmented rather than fully automated.
- Augmentation
- Using AI to streamline or accelerate a human-run process without removing the human from the evaluation loop, as opposed to full automation where the human is removed entirely.
- Gotchas section
- A section added to a Claude skill file that documents edge cases, stylistic corrections, and recurring mistakes from real use, so the skill does not repeat the same errors across sessions.
Things they pointed at.
01:06channelBoris Cherny (creator of Claude Code)
03:40channelAndrej Karpathy (former head of AI at Tesla)
05:38toolWhisper Flow (voice transcription)
09:09toolHostinger MCP
05:25linkFree 5-day AI email series
Lines you could clip.
01:14
“The main agent just thinks, this is easy, I'll just do it myself, and keeps working in its own context when a sub agent would have served you better.”
05:09
“It literally doesn't matter how good these models get. It doesn't know what you want if you don't tell it.”
08:11
“Give Claude a way to verify its work. If it has a feedback loop, it will two to three x the quality of the final result.”
12:20
“Concrete beats abstract every time.”
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.
00:00Today, we're talking about how to actually build 10 times faster with code. Most people think it's better prompts, a smarter model, or installing more plugins. But none of that is the actual unlock.
00:09The unlock is six simple power phrases you type into Claude that change how every project starts, scales, and runs without you. And in this video, I'll walk you through all six, the exact thing to type into Claude, and at the end, I'll share a cautionary phrase that ninety nine percent of people get wrong, which cost them time and money.
00:26Power phrase number one is launch sub agents. Before I explain how this single phrase can 10 x your productivity, let me outline an issue everyone watching this has run into. You ask Claude to do something, you wait ten minutes, you review the output, you ask it again, then you wait again, and you do this over and over again.
00:41Now, are a number of reasons why this happens, which we'll address throughout this video. But one of the biggest reasons is you're running one Claude session at a time, sequentially like a single lane highway. And the whole time you're sitting there waiting, you could have had five Claude's doing five different tasks in parallel like a five lane highway.
00:57This is what Borgeschurney, the creator of Claude Code, swears by. Multiple Claude's at once, each focused on one task end to end. And that's why power phrase one is launch sub agents.
01:06Sub agents are separate Claude sessions running in parallel. Each has its own context window, prompt, and permissions. And my favorite part is you don't have to set them up manually.
01:14You'll just say launch sub agents and Claude splits the work for you. And yes, Claude can do this without you saying launch sub agents, but Claude out of the box underutilized sub agents. There are a number of business reasons why they would do this, but for the sake of this, the main agent just thinks this is easy, I'll just do it myself, and keeps working in its own context when a sub agent would have served you better.
01:34So knowing when to force the issue is a critical skill. But what kind of task should you use this on? There are three things that I use them on in my own business as well as in my client's 7 and 8 figure business.
01:43The first scenario is multiple perspectives. Five sub agents analyzing the same thing in parallel give you a diverse opinion. If you ask the same thing five times in one single chat, the model anchors on its previous's response and it all converges to the same thing.
01:58In practice, for a non technical output, you could use something like this prompt to analyze a social media post across three different lenses. Or for a technical output, you could review your code like this where each sub agent look at the same input through a different lens. For both of these, each sub agent runs with its assigned task without seeing the others.
02:16Meaning, the results of the individual sub agents won't impact the others, which is a massive unlock. Now, scenario two is it opens a door for new possibilities. Having agents at your disposal opens a door to complete tasks that you would have never done before.
02:27For example, to help me find buildpartner.ai domain, I launched 10 different sub agents to check 10,000 plus domains, each agent focused on different groups.
02:36This is something that I just would have never done because it would have taken too long for a single agent to do and it just would have been too much back and forth. So it opens up the realm of possibilities of things that you would consider doing. Now, scenario three is you can just work quicker.
02:47If your tasks are independent of each other, just launch sub agents. To get started with all of this, it's as simple as writing this at the end of your prompt. Launch five sub agents to handle this.
02:55And this is one of my favorite of all of the power phrases because of how lazy you can be with it. There's one clear thing you have to avoid. No matter how many sub agents you launch, if you aren't crystal clear about what they should be doing, you're going to be disappointed with the output.
03:08And that brings us to power phrase number two, which is the only thing that will let these sub agents be extremely powerful. Phrase two, write me an implementation spec. The creator of Call of Code is on record for saying that before you build anything, you need to create a plan or a spec so the AI knows how to build it.
03:25And you may have heard this before, but let me address one misconception you may be thinking. Doesn't Claude already have plan mode for this? And sort of.
03:31Right? So plan mode, which you can get to by hitting shift tab and Claude code, helps you create the spec, but it doesn't replace it. As Andrea Carpathi, the former head of AI at Tesla, put it Actually, I don't even like the plan mode.
03:42I I would I mean, obviously, it's very useful, but I think there's something more general here where you have to work with your agent to design a spec that is very detailed, and maybe it's a maybe, basically, the docs, and then get the agents to write them. Okay. So that's from Carpathi.
03:55Now why is creating a spec so important? Everyone watching this has had this happen to them. Type in a Claude, you wait ten minutes, you review the output, and it isn't even close to what you wanted.
04:04And this happens because unless you're extremely specific, Claude has to make assumptions. And statistically, those assumptions are just unlikely to be correct.
04:12So to show this, let's say there are three steps in a build and there are five potential options at each step in the process. The total number of combinations is five times five times five, so there are a 125 potential ways to build it. Now, if you build a proper spec before Claude does anything outlining these three steps and how to build it, there's now one times one times one option.
04:31There's only one possible solution and Claude will get it right almost every single time. That's a simplified example, but in reality, you may be building something with thousands of options and hundreds of steps. So without a proper spec, it's statistically impossible for Claude to get it right on the first try.
04:46And it literally doesn't matter how good these models get. It doesn't know what you want if you don't tell it. In practice, here's what this could look like.
04:52Write me an implementation spec to build a daily overview automation. Claude will then write a spec doc for how it plans to do it. You can then review it, and now you can be confident that it'll build the right thing.
05:02If you want a more robust prompt, you can type this in or just take a screenshot of it and drag it into Claude. The most important line of this is for each step, show me the key decisions you'd make. This forces Claw to break down the build so you can override decisions before it builds anything.
05:16Now, something you may be thinking to yourself is how do you know what should be in the spec and how do I know if that's even the best strategy? And we'll cover that in power phrase number three, but before that, I'm moving pretty quickly. And if you want a more in-depth breakdown, you can do at your own pace.
05:29I have a free five day email series that covers the AI system I use to help build a $25,000,000 plus startup. Over 4,000 people just like yourself have gone through it and have absolutely loved it, but if you don't like unsubscribe at any time.
05:41So to check that out, click the first link in the description. Again, it is entirely free. Power phrase number three is interview me.
05:47So you've created the spec which sounds great in theory, but there are some real limitations. You have AI write the spec for you and then you realize you don't know what the you're looking at. You know what you want to build at a high level, but you don't know the details that make or break the project.
06:00But I think you do know the details. You just don't know the right questions to ask. And that's why power phrase three is interview me.
06:06Instead of you trying to use AI to give it enough context to write a perfect spec from scratch, Claude flips the switch. Claude then asks you the questions that you didn't know to ask. And the reality is when Claude's outputs aren't great, it usually boils down to a thinking problem.
06:19And by having Claude interview, very quickly you'll realize, man, I never thought about that. This Claude thing, it's pretty smart. And this will save you hours of building the wrong thing.
06:28Here's a prompt you can use, and within it, there are a couple of key meta points you have to consider. The first is work with me to identify the core problem we're solving, who it is and isn't for. I cannot stress how important this is.
06:39One of the reasons Anthropic, the creator of Claude, has been able to grow so quickly is they have a clear focus on who they are initially building for and who they are not building for. For them, they were building specifically for developers. And you need to think about this in the same way because that impacts the final product.
06:55The second is work through any key decisions together. This invokes Claude to ask questions about each step. You may not know all the steps to building a thing, which is totally fine, but Claude does.
07:05And work through those steps with Claude, and if you don't know the answer, just say use your best judgment. That's basically the same as just YOLO ing it, but at least you now understand what impacts the actual final product. The third is summarize it back to me as an implementation spec.
07:19The interview me pulls the info out of you, and then it'll create a spec from it. You go from a blank page to a complete spec in less than five minutes. Now, the trap with interview me is it only works if you actually think before you answer.
07:30And don't just rush through it saying yes or no because you might as well just have Claude auto answer for you then. So with each question and answer with full context. And the pro tip here, use hex or whisper flow to record your voice and have it write the answer for you so it feels a little bit more like a real interview.
07:45Now, you've gotten the spec, the interview, and parallel sub agents launched. You're moving extremely fast, but there's a massive problem that catches everyone off guard. And that's where power phrase number four, verify before you build comes in.
07:56Unfortunately, Claude hallucinates, tells you task is complete when the thing isn't actually complete or it doesn't work as you want it. Boris, the creator of Claude code is on record for saying, give Claude a way to verify its work. If it has a feedback loop, it will two to three x the quality of the final result.
08:11But Boris actually took it one step further. He said the real move is to give Claude a tool to see the output of its work, then tell Claude about that tool. Once he can actually see its own output, it can self correct.
08:21So if you're building a website, give Claude the ability to open a browser and check the work. If you're creating a report, give Claude a way to check it against your brand guidelines. The verification rule tells Claude what to verify.
08:31The tool tells Claude how to verify it. So to actually verify before you build, take a three layer approach. Layer one is update your Claude MD.
08:39Have Claude add one line to your Claude MD file. Before you do any work, mention how you could verify that work. The Claude dot MD file gets inserted at the start of every prompt you write.
08:48So now Claude will bias towards telling you its verification plan first before anything is built. Layer two is enable the necessary tools. Ask Claude, based on what I'm building, what are some tools that could help with verification?
09:00And these tools can be both external or internal and technical or nontechnical. Two examples, the first one, an external technical example. Let's say you're deploying an app to Hostinger and you wanna make sure that it's live.
09:11You can set up the Hostinger MCP to verify the deploy was successful. And then the other example, an internal nontechnical example. Let's say you're writing a social media post.
09:19You can make a brand voice validator skill that checks output against your voice. And I'll cover later in this video the best way to create these skills. Now layer three is understand your human validation zones.
09:30Murphy's law states, anything that can go wrong will go wrong. And in the age of AI, this has never been more true. AI agents doing the majority of the work now means that humans are removed from the loop for almost everything that relates to execution.
09:43So you can just guarantee that things will happen that you didn't want to happen. So to actually verify agents outputs, you have to identify where cost of error is high. Those then become your human validation zones.
09:54Let's say you have an app and the website itself that the users interact, let's consider that low cost of error. But the payment system behind all of it, that's high cost of error. So you would identify the payments as a human validation zone, which you would then require to sign off on on any changes.
10:08And then on the other hand, the low cost of error, the website itself, you're okay with building fast and breaking things there. Now there's a whole rabbit hole on how to handle those human validation zones, but the simplest move is to require human sign off on changes, remove permissions entirely. I go a bit deeper into this in this video, but the 8020, you can update your Claude MD to say this.
10:27Now, to integrate all three layers that I covered as quickly as possible, type this into your Claude code. This is a way to eighty twenty the whole verification process. Now the first four power phrases set the foundation, but the last two are the most important ones that you can't ignore.
10:40But before we get to those, if this is your first video, welcome to the channel. If it's your second or more, here is our anti slop agreement. The visuals, testing, the time I put into this is for humans, you watching this, not for AI clanker robots.
10:53So all I ask is that you subscribe as part of this agreement so that I can keep making videos for humans and not AI robots. Also, I do want to congratulate Floki Crypto for winning our Claude Max giveaway. For everyone else watching, comment below with what you're building to enter the next giveaway.
11:08Power phrase number five. Based on this conversation, build me a skill.
11:12So now you've got sub agent specs interview verified. You're shipping clean work fast. But here's a problem you'll notice within a week.
11:19You're gonna start doing things repeatedly. And that's where you're gonna need the phrase based on this conversation, build me a skill. To summarize what a Claude skill is, it's a folder with instructions that Claude can call automatically.
11:29Think of it like a set of directions that allows Claude to do the same thing every single time. But here's where most people get this wrong. They sit down and they think abstractly.
11:37They say, what skill should I build? And either they never build one because they just can't think of anything, or they build one and they never use it. Which is why you should never build skills abstractly.
11:46You have to build them from conversations that you've already had. When you say based on this conversation, build me a skill, you're forcing the skill to come from something you just did. The use case is already validated because you literally just did the work manually.
11:59Claude is just packaging it up to make it repeatable for the next time. So once you do this, you've created a skill, but there's one move that'll 10 x its quality. Anthropic just published an article on how they use skills internally.
12:09And the best thing you can do to these skills is add a gotchas section. What exactly is this? Well, when you first make a skill, you're gonna realize that it's not perfect.
12:17And the small issues that come up when you're using it, these are what we call gotchas. These are edge cases, stylistic quirks, anything you had to go back and forth to get the final output correct. To enhance your skill over time, you'll include these gotchas as you use it so that it doesn't make the same mistake twice.
12:31You can say, based on this conversation, enhance any skill I use to include a gotcha section so we don't make this mistake again. The biggest takeaway from all of this is instead of asking what skill should I create, you have to say, should I turn this process into a skill?
12:45Concrete beats abstract every time. The faster you can identify gotchas and update the skill, the faster it'll improve. Tightening that feedback loop is what makes skills compound.
12:54And that's why this next power phrase is the most powerful, yet the most dangerous of all of the phrases. Power phrase number six, automate this.
13:02As they say in Spider Man, with great power comes great responsibility. And this power phrase is one that I'm the most skeptical of. There are two scenarios to consider when setting up automation with AI.
13:12The ideal, which it works all the time perfectly. And then the not ideal, it breaks and the output isn't exactly what you want. And the reality is unlike a lot of the YouTube AI productivity hype boys, almost nothing is the ideal scenario.
13:25You'll end up with a lot of non ideal automations. So each one adds operational debt that you have to manage. So the reality is you should be thinking about augmentation, which is using AI to streamline your processes, not just automating them entirely.
13:39So when should you automate this and when should you augment it? First, you need to run what I call the taste test. Ask yourself, does this task require taste to judge a good or bad output?
13:49Or is it entirely quantifiable? If it requires taste, then it has to be augmented. If it doesn't require taste, then it's a potential automation candidate.
13:57Then once you do that, run the eighty twenty output analysis. If the output was 80% as good, would I be okay with it? If yes, automate it.
14:04If no, augment it. For the eighty twenty output analysis, here's a list of things that I'm okay with 80% as good. And then on the flip side, here's what I refuse to lose any quality on.
14:14So these two filters are things that 99% of people don't do. Unfortunately, what happens instead of doing this, people create a ton of different automations that do nothing besides waste tokens and produce AI slop at scale.
14:26Now, Claude has a bunch of features that help with both automation and augmentation, which I'm gonna be making a whole video on, which you'll be able to find in the description once that's ready. But on screen, you can see three features the creator of Claude code calls the most powerful features. Hooks, schedule, and loops.
14:40If you bring this into Claude, it'll give you a starting point for how to think about automation and augmentation. Now, please please please be mindful anytime you say automate this.
14:50As someone who loves productivity and spends every day thinking about this, trust me, be careful using this phrase. Now, is the last of all six of the power phrases that we covered today. Now, if you like this video, you'll love this one where I build on the exact concepts we covered.
15:04I show you how to take these phrases and turn them into a system that will give you a competitive advantage for the next five to ten years. I'll see you in the next one. Peace.
The Hook
The bait, then the rug-pull.
The bottleneck in most Claude Code sessions is not the model, not the prompts, and not the plugins. It is the habit of running one session sequentially when the same work could be split across five agents running in parallel. Austin Marchese walks through six specific phrases — the kind you append to any prompt — that change that default across every stage of a build.
Named ideas worth stealing.
13:48concept
Taste Test
Binary filter before automation: does judging output quality require human taste? If yes, augment. If purely quantifiable, automate.
Steal forAny decision about which AI workflows to fully automate vs keep human in the loop
13:56concept
80/20 Output Analysis
If 80% quality output is acceptable for this task, automate it. If any quality loss is unacceptable, augment instead.
Steal forSOPs for AI-assisted content production, client deliverable workflows
09:40model
Human Validation Zones
Map all tasks by cost-of-error. High-cost zones require human sign-off; low-cost zones can build fast and break freely.
Steal forAny agentic system design, especially where Claude has write permissions
04:04concept
5^N Spec Math
- N build steps
- 5 options per step
- 5^N without a spec
- 1 option with a spec
With N steps and 5 choices per step, an unspecced build has 5^N possible outputs. A complete spec collapses that to 1.
Steal forConvincing teams to write specs before any agentic build
How they asked for the click.
VERBAL ASK
05:35newsletter
“I have a free five day email series that covers the AI system I use to help build a 25 million dollar plus startup. Over 4,000 people have gone through it. Click the first link in the description.”
Mid-video email series plug at the natural chapter break after phrase 2. Clean placement, not disruptive. Framed around the AI system behind a real outcome.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
- How Claude Code's creator starts every project ↗
- How the Claude Code's creator actually uses Skills ↗
- The Claude Code trap experts see that you don't ↗
- Use BuildPartner to build 10x faster with Claude Code (try free) ↗
- Want to build a SaaS product without hiring a CTO? ↗
- Wispr Flow (voice-to-text for the interview) ↗
- Hex (voice-to-text for the interview) ↗
- Human-validation zones deep dive (video) ↗
- Instagram ↗
- Youtube ↗
Visual structure at a glance.
