08:31Brad | AI & Automation · Tutorial
This Skill Can INSTANTLY Watch any Video For Free — Here's How
A Claude Code plugin update that makes an AI watch any video 40 times faster by grabbing keyframes instead of decoding every frame.
July 2ndThe argument in one line.
A Claude Code skill turns real sales-call transcripts into a standing panel of customer-clone AI agents that argue with each other and hold their ground, replacing AI's default yes-man feedback with something closer to real buyer pushback.
Read if. Skip if.
READ IF YOU ARE…
- A solo founder or marketer who already has recorded sales calls and wants to pressure-test copy before it goes live.
- Someone who's noticed that AI feedback tools always say the idea is good and wants a second opinion that can disagree.
- A creator selling a paid offer who needs to know which specific objection is killing conversions, not just that conversions are low.
SKIP IF…
- You don't have any recorded sales calls yet — the whole system is built from real transcripts, not guesses or demographics.
- You're looking for generic AI persona tools with made-up backstories rather than ones built from your own customer data.
The full version, fast.
AI feedback defaults to agreement, so this Claude Code skill builds a panel of AI agents cloned from a business's real sales calls — composite 'customers' grouped by how they buy, each backed by 3+ real people and a bank of their actual quotes. Pasted copy or offers get a two-tier reaction (gut check, then full explanation if it's strong), a moderator finds where the panel splits, and only disagreeing agents get a rebuttal round. The payoff is a readout of exactly what would change each holdout's mind, cited to a real quote. Every avatar keeps a persistent position file so it never contradicts itself, and the panel is used to test messaging before launch, validate ideas before building, and diagnose why a launch already flopped.
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 – 00:57
01 · I cloned my customers
Cold open: Bradley argues AI's default behavior is to agree with you regardless of idea quality, then reveals a skill that spawns AI agents cloned from his own sales calls to react honestly instead.
00:57 – 01:25
02 · Run your first focus group
Typing /focus-group and pasting any landing page, offer, or price spins up one isolated sub-agent per seat on the panel; the agents can't see each other's reactions as they run.
01:25 – 01:49
03 · Round 1
Every agent reacts alone in two tiers — a fast gut check on whether it grabbed attention, then a full explanation only if the reaction is strong in either direction.
01:49 – 02:09
04 · Round 2
A moderator agent with no opinions of its own finds where the panel splits; only the disagreeing agents return, see the opposing argument, and may change their mind.
02:09 – 02:33
05 · The panel readout
The final output is a readout actionable in thirty seconds: what the panel said, what's missing, and exactly what would change each holdout's mind — every verdict cited to a real quote.
02:33 – 04:01
06 · How it works
Build process: sub-agents read the entire sales-call library, extract objections, price reactions, and decision-makers plus verbatim quotes, then group real customers by how they buy into composite avatars — minimum three people per seat — that the user can rename, merge, or cut.
04:01 – 04:35
07 · Persistent memory
Each avatar keeps a position file that updates after every session so it never contradicts a past answer, and the skill proposes new seats when fresh calls don't fit any existing avatar.
04:35 – 05:38
08 · The 3 ways I actually use it
Three live use cases: pressure-testing messaging and pricing before anything ships, validating whether a product idea solves a real problem before building it, and a post-mortem diagnosis of why a live launch underperformed.
05:38 – 07:03
09 · Set up
Two-command install works in Claude Code, Codex, and Cursor and is free; point it at exported call transcripts (Fireflies, Zoom, Google Meet) or connect via MCP, then confirm the auto-built panel.
Lines worth screenshotting.
- AI feedback tools default to flattering you, so the honest signal has to come from agents modeled on real buyers, not the model's own opinion.
- A composite avatar needs at least three real customers behind it — a seat built from a single client is disallowed by design, so no seat is a 1:1 copy of one person.
- Sales-call transcripts get mined for exact problem language, every objection raised, price reactions, and who else sits in on the buying decision.
- Each simulated customer gives a two-tier reaction: a one-line gut check first, and a full explanation only if the reaction is strong enough to justify one.
- A moderator agent reads every reaction and surfaces disagreements without adding its own opinion — its only job is finding where the panel splits.
- Only the agents who disagreed in round one come back for round two, and only after seeing the opposing argument — a live test of whether a real buyer would change their mind.
- Every verdict is required to cite an actual quote from a real sales call, not a generalized impression.
- Each avatar keeps a persistent position file, so it never contradicts what it said about the same offer in a previous session.
- New sales calls that don't fit any existing avatar can trigger the system to propose an entirely new seat, so the panel evolves as the customer base does.
- The same panel gets used three ways: pressure-testing new copy before it ships, validating whether a product idea solves a real problem before it's built, and diagnosing exactly why a live launch underperformed.
- Launch post-mortems come with receipts — which specific objection got triggered and what line would have changed the outcome — instead of a vague 'it didn't land'.
- The install is framed as two copy-pasted commands and free, deliberately removing friction relative to the value just demonstrated.
Real disagreement beats another AI opinion.
WHAT TO LEARN
The value of an AI feedback tool comes from where its opinions are anchored, not from how confident it sounds — anchor it to real customer data and structure it so agents can disagree, or it just tells you what you want to hear.
- AI feedback tools default to flattering you, so an honest signal has to come from agents deliberately modeled on real buyers, not the model's own judgment.
- Naming the failure mode — 'the world's biggest yes-man' — before presenting the fix makes the audience feel understood before they're sold a solution.
- A fast gut-check reaction followed by a full explanation only when warranted mirrors how real people actually respond, which is more useful than forcing every reaction to be equally detailed.
- Isolating reviewers from each other before they react prevents groupthink and surfaces genuine disagreement instead of a converged, watered-down consensus.
- A neutral moderator whose only job is to find disagreement — never inject an opinion — keeps the debate focused on the real split instead of a new argument.
- Requiring three or more real people behind any composite persona prevents a research tool from overfitting to one loud or unrepresentative customer.
- Extracting exact problem language, objections, and price reactions from transcripts is what makes a persona sound like a specific customer instead of a generic stereotype.
- Letting the user confirm, rename, merge, or cut personas before finalizing them builds trust in a system whose credibility depends on feeling accurate, not automated.
- Giving each persona a persistent memory of its own past answers keeps repeated feedback consistent, which matters more for a standing tool than a one-off report.
- Letting new data propose an entirely new persona — rather than forcing it into an existing bucket — keeps a customer model honest as the customer base actually changes.
- The same research asset serves three different moments: before writing (pressure-test messaging), before building (validate the idea is wanted), and after launching (diagnose what went wrong).
- A useful post-mortem names the specific objection that got triggered and the specific line that would have fixed it, not just a verdict that something underperformed.
- Testing a new offer against the same standing panel every time makes results comparable over time, rather than every test starting from a blank slate.
- Building feedback tooling from data you already own (call recordings) costs nothing extra and is more defensible than any generic, demographic-based AI persona.
Terms worth knowing.
- Composite avatar
- An AI persona built by merging traits from three or more real customers who buy in a similar way, rather than modeling any single real person.
- Moderator pass
- A review step where a non-opinionated AI agent reads every customer-agent's reaction and flags only where they disagree, without adding its own judgment.
- Position file
- A saved record of what a customer-avatar has previously said, so its stance stays consistent across every future test instead of contradicting itself.
- MCP
- Model Context Protocol — a standard that lets an AI agent connect directly to an outside tool, like a meeting-notes app, to pull in live data instead of reading exported files.
Things they pointed at.
00:00toolClaude Code
05:50toolCodex
05:50toolCursor
06:20toolFireflies
06:26toolZoom
06:26toolGoogle Meet
06:36toolMCP (Model Context Protocol)
Lines you could clip.
00:08
“Most people think that AI is objective, but in reality, it's the world's biggest yes man.”
01:18
“Marcus has no idea what Dana said, and they can't see each other's reactions.”
03:17
“Every avatar needs at least three real people behind it, so no seat is ever a one to one copy of an actual client.”
04:20
“They actually remember what they said last time, and they never contradict themselves between sessions.”
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.
00:00I cloned my best customers in Claude code, and now I can test any idea on them before it goes to market. Most people think that AI is objective, but in reality, it's the world's biggest yes man. And no matter how good or bad your idea is, it finds a way to justify exactly what you want to hear just to keep you happy.
00:18So I built a skill that makes Claude honest. It lets you see exactly how your real customers would react without being a yes man at all. All you have to do is paste your landing page, copy, or offer, and it spawns a full focus group of independent AI agents to react and analyze it.
00:35Now these aren't generic AI personas either. Each of the focus group agents are avatars based on composites of your real customers. That means they think how they think, react how they react, and even use the same words and phrases.
00:49When I ran it, it built these five agents based on 97 sales calls over the last couple of months, and each of them reacts completely differently to the offer. So before I let you in on how this skill lets you read your customers' minds, I wanna show you how the focus groups actually run. All you have to do is type slash focus group and paste in whatever you want to test, literally anything.
01:10And when you hit enter, the skill will spin up separate sub agents for every seat on your own custom focus panel. All of these agents are running at the same time, but they're isolated from each other. So Marcus has no idea what Dana said, and they can't see each other's reactions.
01:25They even have two reaction levels. The first time, they'll give you just a quick gut check. Did this actually get their attention?
01:32Do they care about whatever you're trying to test with them? Then after that comes the spoken take. If the gut reaction says they'd skim it, you literally get a two line brush off just like a real person would give you.
01:43If the agent has a strong reaction, whether that's positive or negative, then you get the full explanation as well. Then a moderator agent comes over the top and reads every reaction and pulls out the sharpest disagreements.
01:55Doesn't add any opinions of its own. It just finds where the panel splits. In round two, only the agents that disagree get called back into the room.
02:03Each one sees the opposing argument. The point here is to see if a person in their position would genuinely change their mind. Once that's done, you get a readout that you can act on in thirty seconds.
02:12It gives you a summary of what they said and what's missing. Every verdict is cited from a real quote from your customers. Then there's a section that's actually worth this entire skill.
02:22What would change the mind of each holdout? It tells you exactly what to change to get your customers to say yes. And from there, you can grill any of the agents directly in the conversation.
02:33I So keep saying that these agents react like your real customers, but I wanna explain exactly how this skill builds them. Each of these agents is an avatar, a composite of your real customers built from your own business data. But here's how that works in reality.
02:47The first time the skill runs, it reads your entire sales call library, and it fans out a whole fleet of sub agents. Each one of those reads a few transcripts and pulls out the stuff that actually predicts buying behavior.
02:59That might be the exact words they use to describe their problems, including every objection they raised, how they reacted when price came up, and who else is involved in the buying decision. Plus, it extracts their actual lines, word for word, which go into a quote bank.
03:14Then it takes all of those real people and groups them by how they buy. From there, it comes up with the composite avatars. And every avatar needs at least three real people behind it, so no seat is ever a one to one copy of an actual client.
03:28If your calls only support two clean archetypes, you'll only get two. The skill is never going to pad out the panel with thin avatars just to look impressive. From there, it actually shows you what it found, and you can confirm the panel.
03:41So you can rename the seats, merge them, or cut the ones that you don't want. Then it builds each avatar into a full dossier and saves it into a markdown file on your computer. Who they are, how they buy, their objection map, a bank of 30 plus real quotes that anchor how they talk, and a coverage map of what they can, can't, and can speak to.
04:01And every one of these agents on the panel has persistent memory across sessions, so they remember their past decisions.
04:08They each keep a position file that gets added to after every single focus group session that you run, So they actually remember what they said last time, and they never contradict themselves between sessions. They stay current too because every new sales call that comes in, the skill will refresh the dossier with new data.
04:26And if enough new calls come in that don't fit any of your existing avatars, it'll actually propose a new seat so the panel evolves as your market does. So what do you actually use the focus group skill for?
04:37There's three massive use cases that I'm using all of the time right now. The first one is for testing messaging and positioning, and I'm using this one every single day. I'll test landing pages, emails, hooks, pricing, and you can actually paste in the draft, run the panel before and iterate before anything ever goes live.
04:55This is exactly how I'm tuning my launches right now, I can because I know which lines will pull the ICP in, and I'm keeping the lines that can push the bad fit prospects away. The second one is validating a concept or product before you build it. So before you sink months into building something, you can ask the panel if this need is even real.
05:13Does the product actually solve a problem your customers said they have? And the third one is figuring out why something didn't work after you've launched. Say you sent a launch email and you got no replies, or an offer got clicks but no calls booked.
05:26You can run that exact thing through the panel and just ask why it didn't work. The panel will give you the failure causes with receipts, who scrolled past it and why, which objection was triggered, and what would have changed their mind. Okay.
05:39Now let's get this thing running on your own setup, and it takes about two minutes. If you're on Claude code, head to the link in the description below, and you literally just copy these two commands in. The first one adds the marketplace, and the second one installs the focus group skill.
05:53This skill also works in codex and cursor plus regular Claude as well. And the full install instructions for every platform are in the readme, and it's completely free. The first time you run it, it's going to ask you one question.
06:05Where does your sales data live? You've got two options here. The first one I'd recommend is just downloading your call transcripts into a local folder and pointing the skill at it.
06:15Fireflies, Zoom, Google Meet, they've all got transcript exports, and the local files are free for Claude to read. Or if you don't want to do that, you can connect your meeting tools to MCP, something like Fireflies or Google Drive.
06:28It can still go and pull those calls for you. It's just gonna be a little bit slower and a little bit more costly. And from there, it builds the entire panel exactly like I showed you earlier.
06:37It reads the calls, finds the avatars, and you can confirm the panel. And once that's done, you can have a conversation with it. You've got a standing focus group you can pitch anything to forever.
06:47Now this skill doesn't just run on its own either. It's plugged straight into my AI operating system, sitting next to the skills that I run for my content and my research. So if you wanna see exactly how I actually run my whole business on Claude, that's the next video to watch, and it's linked right here.
The Hook
The bait, then the rug-pull.
Bradley opens with the complaint every builder who's leaned on AI feedback quietly knows: it always says the idea is good. His fix is a Claude Code skill that clones his actual customers from real sales calls into a panel of agents that can disagree with him — and with each other.
Named ideas worth stealing.
01:25model
Two-round focus-group protocol
- Round 1: every avatar reacts alone — gut check first, full explanation only if the reaction is strong
- Moderator pass: a non-opinionated agent reads every reaction and finds the sharpest disagreements
- Round 2: only the disagreeing avatars return, see the opposing argument, and may change their mind
A structured way to simulate genuine customer debate rather than a single averaged AI opinion — isolation prevents groupthink, and the rebuttal round tests whether a real objection would actually hold up.
Steal forAny AI critique or research tool that needs disagreement to feel earned rather than simulated.
02:33list
Composite avatar construction
- Read the full sales-call library
- Extract buying signals per transcript — problem language, objections, price reactions, decision-makers — into a quote bank
- Group real people by how they buy, not by demographics
- Require 3+ real people behind every seat — never a 1:1 clone of one client
- Let the user confirm, rename, merge, or cut seats before the dossier is saved
The pipeline that turns raw call transcripts into defensible AI personas instead of invented ones.
Steal forBuilding any customer-research agent from a business's own call recordings.
How they asked for the click.
VERBAL ASK
05:38link
“head to the link in the description below, and you literally just copy these two commands in”
Frames install as two copy-pasted CLI commands taking about two minutes and free — deliberately trivial friction right after demonstrating the payoff, which lowers resistance to acting immediately.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
