My Blog

Tag: www.youtube.com

  • AI Tools: I Fully Automated My Video Editing Using Claude Code (Full Walkthrough)

    Watch on YouTube: https://www.youtube.com/watch?v=HzXD4GVqXwM

    Channel: Christian Peverelli

    Introduction

    Christian Peverelli demonstrates a fully automated video‑editing workflow that relies on Claude Code as a central coordinator. By linking Claude to a suite of free and paid tools, he creates a repeatable “skill” that can edit raw footage while he sleeps, turning a 2 minute 46 second clip into a 31‑second final cut. The approach reduces editing costs from hundreds of dollars per video to roughly $27 and cuts turnaround time from several days to same‑day delivery.

    What the Video Covers

    The presentation walks through the entire system: installing Claude Desktop, enabling local mode, connecting MCP‑based tools, and building a personal editing skill file. It shows how each component contributes to transcription, animation, screen recording, music selection, and final assembly. The video also shares a cost breakdown, time‑savings metrics, and the process for iteratively improving the skill with feedback.

    Step‑by‑Step Walkthrough

    • Install Claude Desktop from Anthropic’s website and upgrade to at least the $20/month plan to unlock Claude Code.
    • Launch Claude Code, set it to local mode, and designate a working folder where all project files will reside.
    • Enable automatic permissions, switch to a newer model for better cost efficiency, and turn on internet browsing within Claude Code settings.
    • Add MCP connectors for Tella (screen and face recording) and Epidemic Sound (music and SFX) so Claude can call those services directly.
    • If Hyperframes is not automatically available, add it manually from its GitHub repository or via the MCP connectors menu.
    • Create a new Claude Code thread and ask the assistant to build a video‑editing system using the listed tools.
    • Record a short hook or any raw footage, feed the file to Claude together with your editing skill, and let the system generate a first edit.
    • Review the output, provide timestamped feedback (e.g., “cut at 00:12, keep the B‑roll”), and instruct Claude to save the correction in the skill file.
    • Repeat the feedback loop until the skill consistently produces the desired quality; each iteration refines the ~1,400‑line, 36‑rule skill file.
    • Join the We Are No Code community to download the shared skill template, get support, and see how others have adapted the system.
    • Run the skill on new footage, review the result, and give occasional feedback to keep the system improving over time.

    Tools and Numbers

    • Claude Desktop: $200/month (maximum Claude usage per video ≈ $24).
    • Tella (annual plan): $13/month (~$2 per video).
    • Epidemic Sound (annual plan): $9.99/month (~$1 per video).
    • Hyperframes: free (provides 4K animation from Claude‑generated code).
    • Parakeet: free, NVIDIA‑based transcription with timestamps.
    • FFmpeg: free (handles final cuts, black‑screen removal, audio sync).
    • Original clip length: 2 minutes 46 seconds → edited length: 31 seconds.
    • Skill file size: approximately 1,400 lines, 36 rules encoding the creator’s editing style.
    • Estimated cost per video: ~$27 (sum of the above subscriptions).
    • Time savings: editing completed same‑day versus 3‑7 days when using a human editor.

    Caveats and Hype Check

    “I save thousands of dollars a month.”

    The claim of saving thousands of dollars per month is based on the creator’s personal usage and may not reflect typical results for all users. The cost reduction from $250‑$1,000 per video to about $27 assumes the creator’s previous editing expenses and the specific subscription levels shown. No affiliate links are displayed, but the video promotes the creator’s free community and skill share, which could drive engagement to his other offerings. Users should evaluate whether the required subscriptions and time investment for initial setup align with their own budgets and editing volumes.

    Concrete Next Actions

    • Download and install Claude Desktop, then upgrade to the $20/month plan to enable Claude Code.
    • Configure Claude Code for local mode, set a working folder, enable automatic permissions, switch to a newer model, and turn on internet browsing.
    • Connect Tella and Epidemic Sound via the MCP connectors menu; add Hyperframes manually if needed.
    • Start a new Claude Code thread and request a video‑editing system using the listed tools.
    • Record a short test clip, provide it to Claude with your editing skill, and review the generated output.
    • Give timestamped feedback, instruct Claude to store improvements in the skill file, and iterate until the output meets your standards.
    • Join the We Are No Code community to access the shared skill template and obtain ongoing support.
    • Apply the refined skill to new footage, review results, and provide occasional feedback to maintain and enhance performance over time.

  • AI Tools: Stop Copying Faceless YouTube Channels! Build Their System Instead

    Watch on YouTube: https://www.youtube.com/watch?v=xkK3PTBolvo

    Channel: The AI Garage

    Stop Copying Faceless YouTube Channels! Build Their System Instead

    The video from The AI Garage outlines a repeatable workflow for creating faceless YouTube videos that mimics the storytelling style of a reference channel while pulling factual content from independent web research. By separating narrative structure from subject matter, the process lets creators produce an endless stream of original videos in a consistent format.

    What the Video Covers

    The presenter walks through a end‑to‑end system that:

    • Ingests the last five videos of a chosen reference channel into Notebook LM to learn hook, pacing, information density, section flow, curiosity restarts, transitions, and resolution.
    • Uses a custom prompt that tells the model to apply YouTube‑derived storytelling style only, while pulling factual information from web sources.
    • Generates original topic ideas with keywords, runs quick web research in Notebook LM to collect reliable sources, and writes a ~750‑word script that follows the learned narrative structure.
    • Creates a visual planner in ChatGPT that defines recurring visual types (kinematic infrastructure, 3D maps, sectional reconstructions, mechanical diagrams, infographic movement) and assigns each script segment a visual type, framing, camera movement, and optional text overlay.
    • Batch‑generates scene clips in Google Flow using independent prompts that already contain subject, style, framing, and animation.
    • Assembles the final video in CapCut (or any editor) by syncing voice‑over from Google AI Studio, background music from the YouTube Studio audio library, and the pre‑generated scene clips according to the visual plan.

    Step‑by‑Step Walkthrough

    1. Set up Notebook LM: Create a new notebook and add the URLs of the last five videos from the reference channel as sources.
    2. Configure the chat: Switch to custom mode and paste the scripting‑engine prompt that instructs the model to use YouTube sources for story structure only.
    3. Generate topic ideas: Ask Notebook LM for five original topic ideas with associated keywords; select one.
    4. Gather factual sources: Run quick web research within Notebook LM to collect reliable information for the chosen topic.
    5. Write the script: Use Notebook LM to produce a ~750‑word script, letting the model apply the learned narrative structure while grounding the content in the web sources.
    6. Design the visual plan: In ChatGPT, run the visual‑planner prompt to define the recurring visual types and their usage rules, then generate a visual batch plan that maps each script segment to a specific visual type, framing, camera movement, and optional text overlay.
    7. Batch‑generate scenes: Copy the visual batch plan into Google Flow, activate the agent, and send the scene prompts for parallel generation. Each scene is independent, with no reliance on previous clips.
    8. Produce audio assets: Generate the voice‑over in Google AI Studio (choose a voice that matches the tone) and download background music from the YouTube Studio audio library.
    9. Assemble the video: Import the voice‑over, music, and pre‑generated scene clips into CapCut (or another editor). Align them according to the visual plan, syncing cuts to the narration and music beats.
    10. Repeat for future videos: Keep the same notebook and visual planner; only replace the web‑researched topic and adjust the visual‑type mix if you switch to a different reference channel.

    Tools and Numbers Mentioned

    • Notebook LM – for ingesting video sources, web research, and script generation.
    • ChatGPT – to create the visual planner and batch plan.
    • Google Flow – agent‑based batch generation of independent scene clips.
    • CapCut – final video assembly (any editor works).
    • Google AI Studio – voice‑over synthesis.
    • YouTube Studio audio library – source for background music.

    The presenter notes that organizing seven scenes took about five minutes, implying that a full seven‑scene batch could be edited in roughly thirty minutes. The script target length is approximately 750 words.

    Caveats / Hype Check

    YouTube sources shape how the story is told. Web sources determine what factual content can be said.

    Each scene is completely independent. There are no character reference images. None of these clips depend on the previous scene.

    The video promotes a paid “Faceless YouTube Engine” guide, claiming a tenfold increase in speed and quality and a ready‑to‑publish first video by the end of the tutorial. These benefit statements are presented as exaggerated claims without supporting evidence.

    Concrete Next Actions

    • Create a new Notebook LM notebook and add the last five video URLs of your chosen reference channel.
    • Open the chat configuration, switch to custom, and insert the scripting‑engine prompt that limits YouTube use to story structure.
    • Request five topic ideas with keywords, pick one, and run quick web research to collect factual sources.
    • Use Notebook LM to write a ~750‑word script, applying the learned narrative structure to the gathered sources.
    • In ChatGPT, run the visual‑planner prompt to define visual types and generate a visual batch plan for the script.
    • Copy the batch plan into Google Flow, activate the agent, and send the scene prompts for parallel generation.
    • Download the generated clips, produce the voice‑over in Google AI Studio, and fetch background music from the YouTube Studio audio library.
    • Assemble the final video in CapCut (or your preferred editor) by aligning voice‑over, music, and scene clips according to the visual plan.
    • Repeat the workflow for future videos, updating only the web‑researched topic and adjusting the visual‑type mix if you study a different reference channel.

  • AI Tools: he made $1,000,000 in 30 days as an AI data broker

    Watch on YouTube: https://www.youtube.com/watch?v=8DAyeNK6VOg

    Channel: Corey Ganim

    Introduction

    The video from Corey Ganim’s channel features Ryan Locke of Polyshares describing how his AI data‑brokerage generated roughly $1 million in revenue during its first month of operation. The discussion centers on the mechanics of sourcing private company data, preparing it for AI labs, and the financial structure that enables referrers to earn a 6 % commission on closed deals.

    What the Video Covers

    Locke outlines the end‑to‑end workflow of Polyshares, positioning the firm as a procurement partner that:

    • Acquires data from private, remote‑first, white‑collar organizations.
    • Removes personally identifiable information (PII) through vetted third‑party de‑identification services.
    • Packages the cleaned data into full‑workflow datasets (e‑mail threads, Slack/Teams messages, ticketing systems, CRM entries) that frontier AI labs require for model training.
    • Negotiates deal terms, including exclusivity periods and licensing structures, and connects sellers with appropriate buyers.

    He also explains the economics of the business, the typical deal size, the referral commission model, and why demand for this type of data is expected to grow as more companies train proprietary models.

    Step‑by‑Step Walkthrough of the Data‑Brokerage Process

    1. Identify qualified sellers – Target companies that are US‑based, English‑language, remote‑first, employ at least 50 people, and have used the same software stack for three or more years. These characteristics simplify de‑identification and increase data consistency.
    2. Engage a third‑party de‑identification vendor – Locke stresses that PII removal must be performed by vetted providers; incorrect de‑identification destroys the dataset’s value and can expose both seller and buyer to legal risk.
    3. Assemble full‑workflow context – Rather than delivering isolated files, Polyshares aggregates emails, chat logs, support tickets, CRM records, and any other relevant digital artifacts that illustrate how work actually unfolds inside the organization.
    4. Structure the deal – Typical transactions are around $300,000, though some reach seven figures. Locke cites a notable Spirit Airlines‑Google deal valued at $10 million as an outlier. Exclusivity periods are commonly set at 24 months; perpetual licenses are discouraged because they limit the ability to monetize the same data later.
    5. Close and compensate – Once a buyer signs, the seller receives payment, and any referrer who supplied the lead via a tracked link earns a 6 % commission on the closed deal.
    6. Monitor market shifts – As more firms train their own models, Locke anticipates demand moving toward specialized, hard‑to‑obtain data such as egocentric or manual‑labor video captured from head‑mounted cameras.

    Key Numbers and Tools Mentioned

    • Revenue claim: ≈ $1 million in the first month.
    • Average deal size: ≈ $300,000.
    • Largest cited transaction: $10 million (Spirit Airlines‑Google).
    • Referral commission: 6 % of closed deal value.
    • Standard exclusivity term: 24 months.
    • Company name: Polyshares.

    Notable Quotes from the Video

    “The average would be about 300,000.”

    “We’ve made seven figures in a single transaction.”

    “If you want to send your deals to Ryan, let him close them and then you get 6%.”

    “Remote first and white collar.”

    “Exclusivity terms norm is probably 24 months.”

    Caveats and Hype Check

    The presentation of the $1 million first‑month revenue figure lacks independent verification. The video ties this claim to an affiliate‑style referral program that offers a 6 % commission, which can create incentives to overstate earnings. Viewers should treat the revenue number as a self‑reported outcome rather than an audited result.

    Concrete Next Actions for Interested Parties

    • Use the tracked referral link provided in the video description to submit qualified leads to Polyshares and earn a 6 % commission on any deals that close.
    • Download the free 80/20 summary manual referenced in the show notes for a quick reference guide on the brokerage model.
    • Begin a lead‑generation side hustle focused on identifying private companies that meet the criteria: >50 employees, ≥3 years on uniform software, US‑based, English‑language, and remote‑first.
    • Research and vet third‑party data de‑identification vendors, ensuring they comply with relevant privacy regulations (e.g., GDPR, CCPA) before engaging in any data‑brokerage activity.
    • Stay informed about emerging high‑value data types, such as egocentric or manual‑labor video captured via head‑mounted cameras, as these may become sought‑by AI labs training proprietary models.

  • AI Tools: Level Up Your AI Agent – 7 steps to a smarter Muse

    Watch on YouTube: https://www.youtube.com/watch?v=K4ZyKu9TXcs

    Channel: Humanoid

    Introduction

    The Humanoid channel presents a practical framework for turning a personal AI agent—referred to as Muse or Q—into a more reliable and context‑aware collaborator. The video outlines seven concrete steps that move the agent from a generic responder to a tool that learns from corrections, maintains a consistent voice, and applies structured review processes before any output reaches the user.

    What the Video Covers

    The presentation begins with a TL;DR that summarizes the core idea: interview the agent to build deep context, train it on your writing style, turn every manual correction into a permanent lesson, add automated self‑review stages, schedule weekly system checks, teach the agent patience, and apply a four‑step tiered plan review guided by ten engineering principles. Optional adversarial or virtual‑boardroom reviews are suggested for higher‑stakes projects.

    Step‑by‑Step Walkthrough

    • Interview your Muse for context and success criteria. Ask the agent any questions needed to understand the business or task better than you do, and explicitly define what success looks like before assigning work.
    • Build a content library of your past writing. Upload representative samples, extract voice lessons, test generated drafts, and correct mistakes to refine the agent’s style.
    • Treat every correction as a teachable moment. Instruct the Muse to retain the lesson from each manual edit and, when necessary, perform a root‑cause analysis to convert the error into a lasting rule stored in memory.
    • Implement automated multi‑stage review. Configure the agent to first check facts and assumptions, then evaluate voice and quality, so you are never the first reviewer of its output.
    • Schedule a weekly systems review. Set a recurring audit (the video suggests Wednesday morning) where the Muse reviews all scheduled tasks, flags failures, and reports on workflow reliability.
    • Teach the Muse patience. Distinguish capability questions from execution commands, making clear that a query about ability does not trigger immediate action.
    • Apply a four‑step tiered plan review. Follow the sequence Frame → Approach → Foundation → Choose Tier, evaluating the plan against ten core engineering principles to decide whether a Light, Standard, or Deep review is required. For high‑stakes work, add an Adversarial Reviewer (a separate Q instance) or a Virtual Boardroom of five Q instances with competing perspectives to stress‑test the plan.

    Tools and Numbers Mentioned

    • Muse agent (the primary personal AI)
    • Q agent (used for adversarial or boardroom reviews)
    • Gemini API (free tier) as the underlying model interface
    • Ten core engineering principles that guide the review tiers
    • Four‑step tiered plan (Frame, Approach, Foundation, Choose Tier)
    • Three review levels: Light, Standard, Deep
    • Five distinct Q instances for a virtual boardroom
    • Recurring weekly check scheduled for Wednesday morning
    • Example task used in the video: drafting YouTube comment responses

    Caveats / Hype Check

    The video focuses on workflow improvements and does not make revenue promises, affiliate promotions, or guarantees of specific outcomes. Effectiveness depends on how consistently the steps are implemented and the capabilities of the underlying AI model. The presenter notes that results will vary across users and use cases.

    Concrete Next Actions

    • Interview your Muse about your current business or task to capture context and define success criteria.
    • Collect a representative sample of your past writing, upload it to the Muse, extract voice lessons, test outputs, and correct errors to refine its style.
    • Whenever you correct the Muse, instruct it to treat the correction as a permanent lesson and, if needed, discuss the root cause to store a lasting rule.
    • Set up two automated review stages: first a fact/assumption check, then a voice/quality check, so you never see raw output before review.
    • Create a recurring weekly audit (e.g., every Wednesday morning) where the Muse reviews all scheduled tasks and reports any failures.
    • Add explicit rules that capability questions are not execution commands, teaching the Muse to wait for a clear command before acting.
    • Adopt the four‑step tiered plan review (Frame → Approach → Foundation → Choose Tier) and evaluate each plan against the ten core engineering principles to select the appropriate review depth.
    • For high‑stakes projects, deploy an Adversarial Reviewer (a separate Q instance) or assemble a Virtual Boardroom of five Q instances with competing viewpoints to stress‑test the plan.

  • Grok Bot: GrokBot Just KEEPS Getting Better (4 New Use Cases)

    Watch on YouTube: https://www.youtube.com/watch?v=q7lekwOXXqk

    Channel: Kacper Rutkiewicz | AI Made Simple

    Introduction

    The video from Kacper Rutkiewicz | AI Made Simple showcases four recently added capabilities of Grok Bot. Each feature is demonstrated in a short, hands‑free session that ties together search, file creation, messaging, and voice interaction. The presenter frames the bot as part of a Cursor‑based workflow and shares concrete usage numbers from his own account.

    What the video covers

    Kacper walks through:

    • Native X (Twitter) search that does not consume developer credits.
    • Creation of real Google Docs, Sheets and Slides in the user’s Drive.
    • A dedicated bot‑owned inbox at @mail.grokbot.com.
    • A hands‑free voice mode that returns a full transcript after the session.

    He ties the four functions together in a 33‑second voice command that searches X, builds a deck, a doc and a sheet, then emails the links, and he shows the impact on his weekly quota.

    Native X search

    Grok Bot can query X directly from the chat interface without drawing on the user’s X developer or API credits. The feature is limited to 1,000 searches per day per account. In the demo the bot’s X Developer balance stayed at $3.61, confirming that no credits were spent. This allows repeated research queries—such as “top posts about X in the last 48 hours”—to be run throughout the day without affecting any paid X API usage elsewhere.

    Google Docs, Sheets and Slides

    When Google Workspace is connected, Grok Bot creates actual files in the user’s Drive rather than markdown placeholders. Users can point the bot at personal master templates so generated decks, docs and sheets follow brand styling. Slide generation typically takes 5–10 minutes. The same 33‑second voice command that produced a deck, a doc and a sheet moved the weekly quota from 32% to 35% (about a 3% increase), indicating that file‑creation consumes a modest portion of the limit.

    Bot email inbox (@mail.grokbot.com)

    Each bot instance receives a unique address of the form name@mail.grokbot.com, separate from the user’s personal Gmail. The address is claimed by typing “Claim our email and create our inbox”; allocation follows a first‑come, first‑served basis as the feature rolls out. The bot can send formatted HTML messages to any recipient and can also receive mail—useful for verification codes, receipts or newsletter sign‑ups—and reply to incoming messages. An important caveat is that new mail does not automatically wake the bot; a polling routine must be set up otherwise messages remain unnoticed until the user explicitly asks the bot to check.

    Voice mode

    Activating the waveform or call button enables hands‑free speaking; Grok Bot returns a full transcript after the conversation ends. In the headline demo a 33‑second verbal command—search X, build a deck, a doc and a sheet, then email the links—was executed successfully. Longer sessions, such as 15‑ to 20‑minute verbal brainstorming calls, barely register against the weekly quota on the Cursor plan, showing that voice interaction can support extended workflows without significantly impacting usage limits while still providing a searchable transcript.

    Tools and usage numbers

    • Cursor subscription: $60 / month (the plan used for quota measurements).
    • Weekly quota before demo: 32%; after demo: 35% (≈3% increase for the full multi‑step voice command).
    • X Developer credit unchanged: $3.61.
    • Native X search cap: 1,000 searches per day per user.
    • Slide generation time: 5–10 minutes.
    • Voice demo length: 33 seconds; longer brainstorms 15–20 minutes barely affect limits.
    • Integrated tools: Grok Bot, Cursor, X (Twitter), Google Workspace (Drive, Docs, Sheets, Slides, Gmail, Calendar), mail.grokbot.com.

    Caveats and hype check

    The presenter notes that the demonstrated usage figures are based on a single creator’s session and have not been independently verified. Claims about Grok Bot being a “premier personal AI agent” should therefore be treated as directional rather than definitive. The email inbox feature is still rolling out, so name availability varies and users may need to try several variations of @mail.grokbot.com. Additionally, new mail does not auto‑trigger the bot; a polling routine is required to avoid missed messages.

    > “Even 15‑to‑20‑minute verbal brainstorming calls barely impact usage limits.”

    Concrete next actions

    • Claim a preferred @mail.grokbot.com address by typing “Claim our email and create our inbox.”
    • Connect Google Drive via Connect Apps; optionally set a master template for Docs, Sheets and Slides.
    • Add a routine that polls the bot inbox (since new mail does not auto‑wake the bot).
    • Try a native X search prompt, for example: “top posts about AI agents from the last 48 hours.”
    • Run one multi‑step voice command (search X, create a doc, sheet and slide, email the links) and review the transcript plus the quota delta.
    • Optional: join the Skool community at skool.com/ai-automation-network-2970 for templates and guides.