DRIP.

Videos · 44 so far

Every video, and everything we mentioned in it.

A short summary of each one, the video itself, and a link to every tool, repo and source. Newest first.

On YouTube soon

Short · Oct 2026 · 1:58

Claude Code repos

10 free GitHub repos that make Claude way more powerful (Microsoft, Anthropic, Stripe)

What it covers

Ten repos for Claude, checked at the source on Oct 15, 2026 (stars as of that day), three of them run on our PC. Hyperframes (HeyGen, Apache-2.0) turns a described video into HTML and renders a real MP4 on your computer with no per-render fee; Claude usage still counts. Our test: one request, 2 min 17 s, a 5-second 1080x1920 title card. Humanizer (MIT) rewrites AI-sounding text using Wikipedia's Signs of AI writing; our made-up 85-word coffee-shop copy came back as 55 words in 15 seconds with every fact kept. It is not built to beat AI detectors. Google Workspace CLI is one command-line tool with skills for Gmail, Drive, Calendar, Docs, Sheets and Slides; it is not an officially supported Google product and needs a Google Cloud project to log in. Apify's MCP server reaches thousands of ready-made scrapers; the free plan includes $5 of usage a month and scraper runs spend it. Open SEO gives Claude keyword data, rank tracking, backlinks and site audits through DataForSEO; the hosted trial is $0.50 of credit, then $10 a month, or self-host with your own DataForSEO key. Agent Skills is Addy Osmani's personal repo (he works on Claude Code at Anthropic): /spec and /plan save files, /build does the next task. Expo's official skills include its list of 20 tells of generated React Native apps; a real iPhone needs Apple's $99 a year developer program. Supabase's and Stripe's own skills cover rules like row level security and checking webhook signatures; Claude still writes the code. Playwright CLI (Microsoft) drives a real browser; in our test the 219 KB page map of this page went to a file, not into Claude's chat. Claude Security is Anthropic's beta plugin: verifiers try to disprove each finding, it shows the token cost before a scan, and patches are never applied for you. The token savers: JetBrains tested them in July 2026 with Claude Code and Claude Sonnet 5 on SkillsBench (about 80 paired tasks each). RTK advertised 60-90% savings, never saved, and cost a median 7.6% more per task at low effort (+0.1% at high; its newer proxy mode was not tested). Caveman advertised 65%; JetBrains measured 8.5% fewer output tokens with it forced on (low effort), and its README now cites that figure. Ponytail (v4.8.4, medium effort) came out 10.3% cheaper on that workload, about half the 20% it claimed then; v5 claims more. Its Windows delay issues are closed. Claude Mem's docs say its anonymous telemetry is on by default; Claude Security also reports usage counts through Claude Code telemetry (DISABLE_TELEMETRY=1). The keyword table, Codeburn bars and user rows in the video are illustrations.

Everything we mentioned

Install lines for all ten (pick the one you need)

Install one at a time, so if something breaks you know which one did it. Lines starting with / run inside Claude Code; the rest run in a terminal. All lines are from each repo's README or docs (checked Oct 15, 2026).

  • 1 · Hyperframes (needs Node.js 22+ and FFmpeg)

    prompt
    claude plugin marketplace add heygen-com/hyperframes
claude plugin install hyperframes@hyperframes

  • 2 · Humanizer (Claude Code 2.1.142+)

    prompt
    /plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer

Then: /humanizer:humanizer

  • 3 · Google Workspace CLI (needs a Google Cloud project)

    prompt
    npm install -g @googleworkspace/cli
gws auth setup
gws auth login

Skills for your agent:
npx skills add https://github.com/googleworkspace/cli

  • 4 · Apify (Apify says to use its plugin in Claude Code)

    prompt
    /plugin marketplace add https://github.com/apify/agent-skills
/plugin install apify-ultimate-scraper@apify-agent-skills

Hosted MCP server: https://mcp.apify.com

  • 5 · Open SEO

    prompt
    Set up the MCP server: https://openseo.so/docs/mcp
Then the agent skills: https://openseo.so/docs/skills/setup

  • 6 · Agent Skills (Addy Osmani)

    prompt
    /plugin marketplace add addyosmani/agent-skills
/plugin install agent-skills@addy-agent-skills

Then: /spec, /plan, /build

  • 7 · Expo skills

    prompt
    claude plugin install expo@claude-plugins-official

  • 8 · Supabase and Stripe skills

    prompt
    claude plugin marketplace add supabase/agent-skills
claude plugin install supabase@supabase-agent-skills

claude plugin install stripe@claude-plugins-official

  • 9 · Playwright CLI

    prompt
    npm install -g @playwright/cli@latest
playwright-cli install --skills

  • 10 · Claude Security (beta; shows the token cost before a scan)

    prompt
    /plugin install claude-security@claude-plugins-official
/reload-plugins
/claude-security

  • Tip · Codeburn and Claude Mem

    prompt
    See your spend:  npx codeburn
Unused MCP servers and skills:  npx codeburn optimize

Turn off Claude Mem telemetry:  npx claude-mem telemetry disable

On YouTube soon

Short · Oct 2026 · 1:45

Claude Code plugin

Microsoft's free repo gives Claude a sleep cycle (SkillOpt)

What it covers

SkillOpt is an open-source (MIT) project from Microsoft Research. Its SkillOpt-Sleep plugin for Claude Code reads your past Claude Code sessions (read-only), finds tasks you repeat, practices them again, and drafts small edits to your CLAUDE.md and skills. An edit is kept only if it scores better on tasks it held back; nothing changes until you run adopt, and adopt backs up the old file first. It is a preview, and gains depend on the task and model: it helps most on repeated tasks with a checkable answer. The default backend is a free practice run with no AI calls; the Claude backend uses your own Claude usage and sends truncated session snippets to the model, so skip it on sensitive projects. Microsoft's own result (Claude Haiku, a deliberately weak brief-writer skill, 0.00 to 1.00 in one night) was scored on 3 held-out tasks. Our test on Oct 14, 2026 (Claude Haiku, the built-in practice tasks): 0.625 to 0.625 over 2 nights and 0.861 to 0.861 over 3, with no edits kept. The rules and task cards in the video are illustrations.

Everything we mentioned

Set it up in Claude Code in three steps

Needs Python 3.10 or newer and Claude Code. Commands are from the plugin's README (checked Oct 14, 2026).

  • 1 · Clone the repo

    prompt
    git clone https://github.com/microsoft/SkillOpt.git
cd SkillOpt

  • 2 · Add the plugin in Claude Code (run from the SkillOpt folder)

    prompt
    /plugin marketplace add ./plugins/claude-code
/plugin install skillopt-sleep@skillopt-sleep
/skillopt-sleep status

  • 3 · Preview, run, adopt (from inside any project)

    prompt
    /skillopt-sleep dry-run
/skillopt-sleep run
/skillopt-sleep status
/skillopt-sleep adopt --legacy

Nightly:  /skillopt-sleep schedule --hour 3 --minute 17
Real learning instead of the free practice run:  python -m skillopt_sleep run --project "$(pwd)" --backend claude

On YouTube soon

Short · Oct 2026 · 1:54

Who has really seen what AI can do

Karpathy's AI funnel: only about 5,000 people use what the labs keep behind closed doors

What it covers

On Oct 11, 2026 Andrej Karpathy (one of the co-founders of OpenAI, now at Anthropic) posted that the gap in understanding of AI capability has become "a sharp funnel". His napkin math: about 6 billion people have barely come in contact with LLMs, 1 to 2 billion use free chatbots now and then, about 20 million use frontier models professionally for code and math, and about 5,000 have access to frontier systems inside the labs. His three reasons: the impact grows with the size of the problem, the models are strongest where answers can be checked (and where the money is), and access (free, paid, internal). On the map one dot is 800,000 people, so the 5,000 don't fill a single dot: one person in every 1.6 million (8 billion / 5,000, our arithmetic). The chatbot answers, test results, agent timer and swarm in the video are illustrations, not real outputs. Views checked Oct 12, 2026.

Everything we mentioned

A peek at the narrow end: three things to try

These follow Karpathy's three reasons for the funnel: size, answers that can be checked, and access. They are our suggestions, not his instructions.

  • 1 · Give it a big problem you care about

    prompt
    I want to build [the thing you actually need, for example a tool that turns my receipts into a monthly budget]. Before writing anything, ask me questions until you understand it. Then write a plan with milestones and build it one milestone at a time, checking with me after each one.

  • 2 · Try something checkable, like code

    prompt
    Write tests for this first and show me that they fail. Then write the code until every test passes, and finish with the test results.

  • 3 · Use this year's models, not an old free one

    prompt
    Pick the newest model your app offers (check the model menu). For code, Karpathy points to agent tools that can run your code, such as Claude Code and OpenAI Codex, as where the jump has been biggest.

On YouTube soon

Short · Oct 2026 · 1:31

Open-source office apps

Someone cloned Word, Excel and PowerPoint for free

What it covers

WordCraft, GridCraft and DeckCraft are free, open-source (Apache-2.0) rebuilds of Word, Excel and PowerPoint by ArtCraft, written in Rust, for Windows, Mac and Linux. Their makers call them clean-room builds: made from public specifications and by observing how the apps behave, not from Microsoft's code. Not made by or affiliated with Microsoft. All three are alpha: by their makers' own estimates about halfway to replacing the originals, and file fidelity is still rough, so keep a copy of anything important. App screens in the video are the projects' official screenshots; the letter, budget and slides are our own test files. Stars checked Oct 11, 2026.

Everything we mentioned

  • WordCraft

    repo

    Word-style documents (.docx), free and open source

  • GridCraft

    repo

    Excel-style spreadsheets (.xlsx), free and open source

  • DeckCraft

    repo

    PowerPoint-style presentations (.pptx), free and open source

  • ArtCraft apps

    site

    The makers' page for all their Crafting Apps

  • Claude Code MCP docs

    docs

    How claude mcp add and /mcp work

Connect them to Claude Code in three steps

We tested on Oct 11, 2026 on our Windows 11 PC (i7-10700F) with v0.5.0 and Claude Code, connected to each app's MCP server: WordCraft wrote a one-page letter in 23.5 s, GridCraft built a six-month budget with real formulas (=B2*C2, SUM) and a column chart in 23.3 s, and DeckCraft made a five-slide deck with a speaker note on every slide in 31.1 s. The apps are free; Claude Code needs a paid Claude plan or API usage.

  • 1 · Download the app from GitHub releases

    prompt
    WordCraft: https://github.com/storytold/wordcraft/releases
GridCraft: https://github.com/storytold/gridcraft/releases
DeckCraft: https://github.com/storytold/deckcraft/releases

On Windows, most PCs want the x64 .msi (for example wordcraft-0.5.0-windows-x64.msi). Install it like any app.

  • 2 · Windows: paste in a terminal (one line per app)

    prompt
    claude mcp add --scope user wordcraft -- "C:\Program Files\WordCraft\wordcraft-cli.exe" mcp
claude mcp add --scope user gridcraft -- "C:\Program Files\GridCraft\gridcraft-cli.exe" mcp
claude mcp add --scope user deckcraft -- "C:\Program Files\DeckCraft\deckcraft-cli.exe" mcp

  • 2 · Mac or Linux instead

    prompt
    Mac: download the separate command-line zip from the same release (for example wordcraft-cli-0.5.0-macos-universal.zip), unzip it, then:
claude mcp add --scope user wordcraft -- /full/path/to/wordcraft-cli mcp

Linux (.deb or .rpm put the tool on your PATH):
claude mcp add --scope user wordcraft -- wordcraft-cli mcp

Same for gridcraft-cli and deckcraft-cli.

  • 3 · Restart Claude Code, check /mcp, then ask

    prompt
    Restart Claude Code, type /mcp and check wordcraft shows as connected. Then try:

Use the wordcraft tools to write a one-page letter asking a supplier for a sample of their new coffee beans. Give it a bold title, a date line, three short paragraphs and a sign-off. Save it in this folder as letter.docx.

Use the gridcraft tools to build a six-month budget with a formula for each month's cost, a SUM total row and a column chart. Save it as budget.xlsx.

Use the deckcraft tools to make a five-slide presentation with a speaker note on every slide. Save it as deck.pptx.

Then open the file it saves.

On YouTube soon

Short · Oct 2026 · 1:34

Open-source creative tools

7 free open-source Adobe-style creative tools

What it covers

Seven independent open-source ArtCraft recreations, not Adobe source code. PhotoCraft supports layers, masks, editable type and PSD, and exposes CLI/MCP automation. Our PhotoCraft v0.5.0 + Claude Code run created an editable PSD and PNG with a Type layer and adjustment, then reopened the PSD; the original image was preserved. Other editor footage is credited official project demos. These are early projects, not a professional-parity test. PrintCraft now redirects to PdfCraft. Source checked Oct 9, 2026. Updated Oct 10: simpler Claude-assisted setup independently verified; opening COSMIC PSD Type layer created in PhotoCraft over the credited NASA/ESA/CSA/STScI Pillars of Creation photograph. Neon Drive is an official VectorCraft example.

Everything we mentioned

Let Claude Code set up PhotoCraft in three steps

The Windows v0.5.0 setup was tested in an isolated folder: Claude downloaded the official 67,717,284-byte portable ZIP, verified the release checksum, and registered project-scoped MCP. Independent CLI and MCP handshake checks passed. PhotoCraft is free; Claude Code uses your plan allowance or API billing. Setup permissions and a restart may be required. Manual Windows/macOS commands are available below.

  • 1 · Open Claude Code in your photo folder

    prompt
    Claude Code must be installed and signed in on your computer. Put a copy of photo.jpg in a project folder and open Claude Code there. This local setup uses its shell and file tools; ordinary Claude/ChatGPT web chat does not install a desktop CLI.

  • 2 · Ask Claude Code to download and connect PhotoCraft

    prompt
    Set up PhotoCraft for this photo folder on my computer. Use only the official GitHub releases at https://github.com/storytold/photocraft/releases. Choose the matching portable build for my operating system and architecture that includes photocraft-cli; on macOS use the separate CLI ZIP. Download and extract it into a tools folder without admin installation. Verify its checksum against the release SHA256SUMS.txt. Connect photocraft-cli to Claude Code as a project-scoped stdio MCP server named photocraft. Pass --automation-read-root and --automation-write-root with the absolute path of this photo folder. Preserve existing MCP entries and my original photos. Check the connection. Tell me to restart Claude Code and use /mcp to confirm PhotoCraft is connected before editing.

  • 3 · Approve, restart, check /mcp, then edit

    prompt
    Approve the download/connection actions within your photo project. Restart Claude Code after registration. If prompted, approve the project-scoped MCP server. Type /mcp and confirm photocraft is connected. Allow the relevant editing tools for the task. Then paste:

Use PhotoCraft to brighten this photo. Add Weekend Mode as an editable text layer. Save an editable PSD and a PNG.

Open the exported PNG and reopen the PSD to check its editable text. Keep the original photo.

On YouTube soon

Short · Oct 2026 · 1:34

Web design

Hairline: a free Claude Code skill that draws line animations that react to your mouse, plus 27 ready-made ones

What it covers

Hairline (lucasmarkes/hairline, MIT, by Lucas Marques, about 1,000 GitHub stars as of Oct 7 2026) is two things. A library of 27 isometric line figures that answer the pointer (a vault you crack by turning its dial, a keyboard whose keys sink, a laptop whose lid follows your mouse), with no dependencies, for React 18+ or any web page. And hairline-create, a skill for coding agents: give it an idea and it offers two or three concepts, draws the one you pick to Hairline's ten rules on the same engine, checks it with a script and in a browser, and hands over one HTML file that opens from disk. Every figure on a page shares one animation loop, and a figure off screen does no work. Our test (Oct 7 2026, Claude Code with Opus 5, run hands-free so it picked the concept itself): "a coffee cup" gave three concepts and a carrier where the cup under your mouse rides up, passed its checks, one 37.7 KB file, in about 10 minutes. "A mailbox for a contact form" gave five mailboxes on a rail whose doors swing open and flags go up, in about 17 minutes. The library and the skill are free; Claude Code needs a paid Claude plan or API credits.

Everything we mentioned

The prompt for each one

Without the -y flag, npx skills add asks which agents and where to install (your project's .claude/skills/ by default). The skill runs node scripts to check its work, and its browser check installs playwright-core once and uses Chrome. The figures need a DOM: on a page that isn't white or near-black, set --hairline-plate to your background, because the plate colour fills every surface. The npm package is ESM only. MIT: keep the licence notice.

  • 1. In your project folder (a terminal)

    prompt
    npx skills add lucasmarkes/hairline -a claude-code -y

  • 2. Open Claude Code in that folder

    prompt
    claude

  • 3. Type the command and your idea

    prompt
    /hairline-create a coffee cup

  • Tip: name the object and what the mouse does, and it skips the concept step

    prompt
    /hairline-create a mailbox; under the mouse its door swings open and its flag goes up

  • The ready-made 27 (in your web project)

    prompt
    npm i @lucasmarkes/hairline

  • Then, in React

    prompt
    import { Vault } from "@lucasmarkes/hairline/react";

export function Hero() {
  return <Vault />;
}

On YouTube soon

Short · Oct 2026 · 1:40

Vibe coding

4 free Claude Code add-ons to switch on before you vibe code, and what two of them did in our tests

What it covers

Four free add-ons for Claude Code. Superpowers (obra/superpowers, MIT, about 296,000 GitHub stars as of Oct 6 2026) stops Claude jumping straight into code: it asks what you're building, writes a plan you approve, then works through small tasks, tests first. It's in Anthropic's official plugin marketplace. The Karpathy-inspired CLAUDE.md (multica-ai/andrej-karpathy-skills, by Forrest Chang, about 217,000 stars) is one file with four rules derived from Andrej Karpathy's post on how AI models overcomplicate code: think before coding, simplicity first, surgical changes, goal-driven execution. In our small test (the same job, add discount codes to a made-up order module, 2 runs each, Claude Code on our PC) Claude added 64 and 77 lines and a new file both times without it, and 54 and 47 lines with no new file with it. The counts include tests, and every test passed. I Have ADHD (ayghri/i-have-adhd, MIT, about 54,000 stars; its README says no diagnosis is needed) stops Claude burying the answer: the next action first, no preamble, no recap. It stays off until you type /i-have-adhd and on until you say "stop adhd mode". Asked the same Netlify question once each, the answer went from 375 words to 184, with the same four fixes. Claude Octopus (nyldn/claude-octopus, MIT, about 4,000 stars) does nothing until you type /octo. Then it sends one task to other AI tools, like Codex or Copilot, and flags where they disagree. It uses those tools' own accounts or keys, or free local models through Ollama, and on Windows it needs Claude Code running inside WSL. The add-ons are free; Claude Code itself needs a paid Claude plan.

Everything we mentioned

  • obra/superpowers

    plugin

    Superpowers: brainstorm, plan, then small test-first tasks. MIT.

  • multica-ai/andrej-karpathy-skills

    file

    The Karpathy-inspired CLAUDE.md, four rules in one file. Also installs as a plugin (then it loads as a skill when Claude judges it relevant).

  • ayghri/i-have-adhd

    plugin

    I Have ADHD: action first, numbered steps, no preamble or recap. Off until /i-have-adhd. MIT.

  • nyldn/claude-octopus

    plugin

    Claude Octopus: one task to several AI tools, disagreements flagged. 12 providers, each with its own account or key; Ollama runs locally. WSL on Windows. MIT.

Install all four

Start a new Claude Code session after installing so the plugins load. We tested every install line on Oct 6 2026 with Claude Code 2.1.220. If you already have a CLAUDE.md, paste Karpathy's rules into it instead of overwriting it. Octopus sends your prompt to each provider you set up, under that provider's own terms and costs.

  • 1. Superpowers (type this inside Claude Code)

    prompt
    /plugin install superpowers@claude-plugins-official

  • If it says not found, add the author's marketplace first

    prompt
    /plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

  • 2. Karpathy-inspired CLAUDE.md (in your project folder, Windows PowerShell)

    prompt
    curl.exe -o CLAUDE.md https://raw.githubusercontent.com/multica-ai/andrej-karpathy-skills/main/CLAUDE.md

  • On a Mac or Linux

    prompt
    curl -o CLAUDE.md https://raw.githubusercontent.com/multica-ai/andrej-karpathy-skills/main/CLAUDE.md

  • 3. I Have ADHD (in a terminal)

    prompt
    claude plugin marketplace add ayghri/i-have-adhd
claude plugin install i-have-adhd@i-have-adhd

  • Then, inside Claude Code: on, and later off

    prompt
    /i-have-adhd
stop adhd mode

  • 4. Claude Octopus (in a terminal; on Windows, inside WSL)

    prompt
    claude plugin marketplace add https://github.com/nyldn/plugins.git
claude plugin install octo@nyldn-plugins

  • Then, inside Claude Code

    prompt
    /octo:setup

On YouTube soon

Short · Oct 2026 · 1:40

AI tools

VoiceStudio: a free, open-source voice cloner that runs on your own computer

What it covers

The voice at the start of our video was made by VoiceStudio on our own desktop, on the processor alone, from one line of description (voice design: nobody recorded it). VoiceStudio is a free, open-source app (AGPL-3.0), built mostly by one developer, debpalash, with about 54,000 GitHub stars as of Oct 6 2026 and 3,124 commits since April. Give it 5 to 15 seconds of clean speech and it builds a voice you can reuse, then reads anything you type in it, turns a book into an audiobook or dubs a video into another language. It also goes the other way: talk, and it types for you. It has its own MCP server, so while the app is open Claude Code can make audio files in a voice you pick, with no API key and no account for VoiceStudio. What it makes carries an invisible AudioSeal watermark by default. Clone only your own voice, or one you have permission to use. Our test: their default engine, run from source on a Windows desktop (i7-10700F with an AMD Radeon RX 9060 XT, which PyTorch on Windows can't use, so processor only). It designed the voice in our video in 22 seconds and spoke a 6-second line in a cloned voice in 2 minutes 18. The voice model is k2-fsa/OmniVoice: the app is free to use for paid work, but that default model's weights are CC-BY-NC (non-commercial), so check the model's licence before you sell anything.

Everything we mentioned

  • debpalash/VoiceStudio

    app

    The app. AGPL-3.0, free. Windows, macOS and Linux installers on the Releases page; about 10 GB of free disk, and it installs its own Python runtime and model weights on first launch.

  • voicestudio.sh

    site

    The project's own site and one-line installer.

  • k2-fsa/OmniVoice

    model

    The default voice model, by Han Zhu. 3.3 GB. Its code is Apache-2.0 but its pretrained weights are CC-BY-NC: non-commercial. Check this before you sell anything you make.

  • Their language catalogue

    docs

    All 646 languages with the hours of speech behind each one. This is the table we added up.

  • Their MCP docs

    docs

    The local MCP server: generate_speech, clone_voice, transcribe and the rest, for Claude Code and Cursor.

  • Their benchmarks page

    docs

    Still says "No verified rows yet", which is why we measured it ourselves.

Set it up in 3 steps

Fine print worth knowing before you start. It wants about 10 GB of free disk on Windows and installs its own Python runtime on first launch, and your first generate downloads a few gigabytes of model, so a slow first run is almost always the download, not a crash. Judge the speed from the second run: the first one loads the model. Every clip gets an invisible AudioSeal watermark unless you turn it off in Settings > Privacy. The who-said-what transcripts need a free Hugging Face account and a one-time licence acceptance. The app is free to use commercially; the default model's weights are not. And clone your own voice, or one you have the person's permission to clone.

  • 1. Download the installer for your computer

    prompt
    https://github.com/debpalash/VoiceStudio/releases/latest

  • Or, in a normal (non-admin) PowerShell on Windows

    prompt
    irm https://voicestudio.sh/install | iex

  • 2. Open Voice cloning, drop in 5 to 15 seconds of clean speech

    prompt
    one speaker, no music, no background noise. it writes down what the clip says for you, so leave the transcript box alone unless it stays empty.

  • 3. On Windows with no NVIDIA card? Switch the engine before you judge the speed

    prompt
    Settings > Model Catalogue > OmniVoice GGUF

  • 4. Let your coding agent use it (optional)

    prompt
    keep VoiceStudio open, then point Claude Code or Cursor at the local MCP server it exposes at /mcp

On YouTube soon

Short · Oct 2026 · 1:39

Vibe coding

Replica: a new Claude skill that reverse-engineers an app and rebuilds it as your own, and what our test found

What it covers

Replica is 11 free Claude Code skills (MIT) by Jake Schincariol. Recon maps an app from its public pages and your own account (screens, flows, the data behind them); the others plan it, rebuild it screen by screen, add logins and payments, test it, score it against the original, read what its users hate in public reviews, rebrand it, write the launch page and deploy it. It rebuilds what an app does, never its code, logo or words, and its deploy step stops until the original's name and colours are gone. Our test, public pages only, on a popular scheduling app: in about 13 minutes Recon mapped 23 screens (13 of them inferred from help pages) and 10 flows, and estimated the full rebuild at about a quarter of a year, or a first slice in a weekend to a week. The Entrepreneur read 555 public reviews from 4 sources (App Store 500, Capterra 20, Hacker News 20, Trustpilot 15; G2 and Reddit couldn't be read) and ranked bugs and crashes first, then price and paywalls, with every quote linked. Good to know: the skills are free, but Claude Code needs a paid Claude plan, and the build and deploy steps use services you set up yourself (a database, payments, email, a host, a domain). The Python checks need Python 3.8 or newer. Start with one slice, the one flow people pay for. Features usually aren't protected the way code, text, images, logos and names are, but this isn't legal advice: run the trademark checks /replica-brand lists, read the terms of anything you used (including the original's, if you study it through your account), keep the original's name out of your app, ads and store listing, and talk to a lawyer before you sell if real money is involved. Reviews are research, never testimonials.

Everything we mentioned

  • Jakeschincariol/replica-skill

    skill pack

    11 skills: recon, architect, design, build, backend, test, diff, entrepreneur, brand, launch, deploy. MIT.

    # paste into Claude Code
https://github.com/Jakeschincariol/replica-skill
install skill

  • Claude Code

    product

    Anthropic's coding AI, where the skills run. It needs a paid Claude plan.

  • Claude Code plugins

    docs

    How /plugin marketplace add and /plugin install work, the other way to install Replica.

  • Python

    tool

    Replica's checks (parity score, review ranking, the rebrand sweep) are small Python scripts. Python 3.8 or newer.

Run it in 3 steps

Recon asks which app, which part, and who it is for, so answer those first. Good-to-know notes are in the summary above.

  • 1. Open Claude Code in a new, empty folder

    prompt
    mkdir my-app
cd my-app
claude

  • 2. Paste this into Claude Code

    prompt
    https://github.com/Jakeschincariol/replica-skill
install skill

  • Or install it as a plugin (type these in Claude Code, pick "Install for you")

    prompt
    /plugin marketplace add Jakeschincariol/replica-skill
/plugin install replica-skill@replica-skill

  • 3. Restart Claude Code, then type

    prompt
    /replica-recon https://the-app.com

  • Then run the rest in order, one at a time

    prompt
    recon > architect > design > build > backend > test > diff > entrepreneur > brand > launch > deploy
More skills in our AI tools directory →

On YouTube soon

Short · Oct 2026 · 1:30

Vibe coding

Cloudflare's free security-audit skill: what it does, how to run it in 3 steps, and our test

What it covers

Cloudflare open-sourced a free AI skill (MIT license) that hunts for security holes in your vibe-coded app and shows you how to fix them. Cloudflare built it to hunt holes in its own code. It maps your app, sends AI agents to hunt by area and attack type, and hands every flagged hole to a fresh agent whose job is to prove it wrong. False alarms are thrown out, holes it can prove come with a fix, and the rest go on a needs-checking list. The bigger system Cloudflare grew from it, with one more scanner feeding it, left 7,245 findings for its engineers across 145 repos. We tested it on Brew Notes, a tiny app we broke on purpose with 8 holes: in about 30 minutes it confirmed 6 of them, found a crash we didn't plant, threw out our cross-site scripting plant with a good reason (it could only hit your own account) and missed one (unlimited password guesses). Without "use subagents" in the request, our run did the whole audit alone, so step 3 includes it. Good to know: On Windows, use Command Prompt if PowerShell says scripts are disabled. If Claude Code was already open, restart it after installing. The skill reads your code: it doesn't touch your live site or edit your files. It uses a lot of your Claude usage: on our 3-file test app it took about 30 minutes, and about an hour with "use subagents" (same results there; the agent team is what makes the skeptic a separate, fresh agent), so start with the quick profile. Without a sandbox (for example on Windows) it can't run your app, so holes it can't prove go in NEEDS-VALIDATION.md; in our Windows test its two checker scripts also needed a workaround, which Claude handled. A quick run is partial: a clean report doesn't mean your app is safe, and Cloudflare says one run found about half of what repeat runs found.

Everything we mentioned

  • cloudflare/security-audit-skill

    skill★ 24K

    The skill itself, free (MIT). Recon, hunting, a skeptic that tries to disprove every finding, then a report with fixes.

    npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit -g -a claude-code -y

  • Build your own vulnerability harness (Cloudflare blog, June 2026)

    article

    How Cloudflare grew this skill into a fleet-wide system, and where the 7,245 findings come from.

  • Skills CLI (npx skills)

    tool

    Vercel's open installer for agent skills. The install line uses it; it needs Node.js.

  • Node.js

    tool

    Install the LTS version first if you don't have it. The install line runs with npx, which comes with Node.js.

  • Claude Code

    product

    Anthropic's coding AI, where the skill runs. It needs a paid Claude plan.

Run it in 3 steps

You need Node.js and Claude Code (a paid Claude plan). Good-to-know notes are in the summary above.

  • 1. Install it (in your terminal)

    prompt
    npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit -g -a claude-code -y

  • 2. Open your app in Claude Code

    prompt
    cd your-app-folder
claude

  • 3. Type this

    prompt
    security audit this codebase, quick profile, use subagents

  • Where the report lands

    prompt
    ~/security-audit-skill/<your-app>/run-1/REPORT.md
(plus FINDINGS-DETAIL.md and NEEDS-VALIDATION.md)
More skills in our AI tools directory →
slide 1 of 9slide 2 of 9slide 3 of 9slide 4 of 9slide 5 of 9slide 6 of 9slide 7 of 9slide 8 of 9slide 9 of 9

Swipe through the slides →

Carousel · Oct 2026 · 9 slides

AI tools

AI tools: hype vs reality (Seedance, Claude Code, GPT Image vs Nano Banana, ElevenLabs, Whisper, Jev)

What it covers

Drip is a cartoon made with AI, so he rated the tools by how they hold up for us. Hype is how loud the internet is; reality is our take, not a lab test. Seedance 2.5 (ByteDance's video model): a video model once gave Drip no legs, AI video has got much better since, and Capp says Seedance earns the hype. Claude Code builds every video Drip is in, over rounds of fixes, not one prompt. GPT Image 2.5 vs Nano Banana Pro, same prompt and the same reference picture: GPT Image drew every picture in the post (2 of 32 came back off-model) at 8x fewer credits on Higgsfield; Nano Banana Pro got one of two takes right and gave Drip eyebrows in the other. ElevenLabs makes both voices; its newest model (Eleven v4) made Drip squeaky, so he stays on the older one. Whisper is free, runs on our computer and checks every word we record, and it also hears words in a coffee grinder. Jev ranks the 201 tools on our site for 27 jobs in about 12 seconds for under 2 cents.

Everything we mentioned

  • Seedance 2.5

    product

    ByteDance's AI video model. Seedance 2.5 is on Higgsfield.

  • Claude Code

    product

    Anthropic's coding AI. We build every Drip video with it, as code.

  • GPT Image 2.5 on Higgsfield

    product

    OpenAI's image model. It drew every picture in this post, with a picture of Drip attached.

  • Nano Banana Pro

    docs

    Google's image model (Gemini 3 Pro Image). Same prompt as GPT Image in our test.

  • ElevenLabs

    product

    Drip's and Capp's voices.

  • Eleven v4 announcement

    docs

    The newest ElevenLabs model (Sept 28, 2026). Drip stays on the older one.

  • Whisper

    code

    OpenAI's free, open-source speech-to-text (MIT license). We run it locally with whisper.cpp.

  • Jev by TypeSafe

    docs

    A decision model that returns typed answers. It ranks the tools on our site.

Every AI tool we use and rate →
slide 1 of 7slide 2 of 7slide 3 of 7slide 4 of 7slide 5 of 7slide 6 of 7slide 7 of 7

Swipe through the slides →

Carousel · Oct 2026 · 7 slides

Free AI tools

5 free AI apps that run on your own computer

What it covers

Five free, open-source apps that run AI on your own computer instead of in the cloud, with no monthly bill. Jan is a ChatGPT-style chat that works with no internet (smaller models than ChatGPT; about 8 GB of memory runs the small ones). Handy types what you say into any app, offline, and is fast on most laptops. ComfyUI makes images on your computer; it looks scary at first because it's a board of wires, and it's best with a graphics card. Upscayl makes small, blurry photos bigger and sharper; it needs a graphics card, and the new detail is the AI's best guess. Buzz turns recordings and videos into text and subtitles using Whisper, the same speech model we use. Checked on each project's own GitHub page and docs, Oct 3 2026.

Everything we mentioned

  • Jan

    code

    Chat with an AI on your own computer, offline. Apache 2.0. Windows, Mac, Linux.

  • Handy

    code

    Press a key, talk, and it types into any app. MIT. Windows, Mac, Linux.

  • ComfyUI

    code

    Make images on your own computer. GPL-3.0. Best with a graphics card.

  • Upscayl

    code

    Upscale and sharpen photos. AGPL-3.0. Needs a Vulkan-capable graphics card.

  • Buzz

    code

    Transcribe recordings and videos into text and subtitles, offline. MIT.

More free and open-source AI tools →
Watch on TikTok ↗

Short · Oct 2026 · 2:11

How AI works

How graphics cards actually work (and why AI needs them)

What it covers

Inside a graphics card are 2,048 tiny workers; the processor next to it has 8 big ones. A chatbot writes by guessing the next word, and every guess is math on numbers it learned in training: printed one per page, a big model's numbers would wrap around the Earth. On our own PC, the same chunk of AI math took the processor 10 seconds and the graphics card 1. Capp explains why with a classroom: 8 math professors (each juggling 16 problems) against a stadium of 2,048 students, a teacher calling one step to a whole row (that's lockstep), worksheets close by (the card's own 16 GB of memory), and the catch: when half a row needs a different step, they take turns. Then why AI's demand for chips and memory is a big part of why gaming cards cost more right now.

Everything we mentioned

Watch on YouTube ↗

Episode · Oct 2026 · 4:55

MCP

MCP Explained in 5 Minutes (How Your AI Talks to Your Apps)

What it covers

Your AI says "I don't have access to your email." Drip and Capp explain MCP with a restaurant called Gmail: the waiter is the MCP server, the menu is the tools, the kitchen window is the API, and the key you allow at sign-in is what lets it all happen. Then the code that's really sent, why the standard exists, and what you're trusting when you click Allow.

Everything we mentioned

  • Model Context Protocol

    docs

    The official site: what MCP is and how to use it.

  • MCP specification

    docs

    The exact messages an AI app and an MCP server send each other.

  • Introducing MCP

    source

    Anthropic's announcement of the open standard (Nov 2024).

  • Gmail API

    docs

    The kitchen in our restaurant: what a Gmail MCP server calls behind the scenes.

Watch on YouTube ↗

Short · Oct 2026 · 63s

Claude Code

4 Claude Code plugins that work like superpowers

What it covers

Capp shows Drip four Claude Code plugins that work like superpowers, all listed in Anthropic's official plugin marketplace. Taste: Frontend Design, by Anthropic, gives Claude an art director's rulebook, so it plans a look on purpose instead of falling back on a generic default. A playbook: Superpowers, by Jesse Vincent, asks what you're really building, writes a plan, writes the tests first and gets the work reviewed. Stamina: Ralph Loop, by Anthropic, hands Claude the same task again every time it tries to stop, until it says your stop word or hits your limit. There's no limit unless you set one, so always set one. X-ray vision: the Chrome DevTools plugin, by Google, lets Claude see the console errors, network calls and a speed trace behind your web page, so it can track down what's broken and fix your code. On October 1, 2026, Frontend Design and Superpowers were the two most installed plugins in the marketplace, by claude.com's own counts. Plugins run code on your computer, so check one before you install it. Not sponsored.

Everything we mentioned

  • Frontend Design

    plugin

    Anthropic's design skill: Claude commits to a clear look and picks type and color on purpose, instead of a generic AI default. 1,134,112 installs on claude.com (Oct 2026). See it on claude.com.

    /plugin install frontend-design@claude-plugins-official

  • Superpowers

    plugin★ 294K

    Asks what you're really building, writes a plan, works test-first and reviews the work between tasks. More steps can mean more tokens. 1,009,371 installs on claude.com (Oct 2026). See it on claude.com.

    /plugin install superpowers@claude-plugins-official

  • Ralph Loop

    plugin

    A Stop hook hands Claude the same prompt every time it tries to finish, until it prints your completion phrase or hits your limit. There's no limit by default and every lap uses tokens, so always set one: /ralph-loop "your task" --completion-promise "DONE" --max-iterations 20. Needs bash and jq (on Windows, Git for Windows). See it on claude.com.

    /plugin install ralph-loop@claude-plugins-official

  • Chrome DevTools

    plugin · MCP★ 53K

    Google's DevTools for coding agents: Claude reads console errors, network requests and performance traces from a live Chrome, and takes screenshots. Needs Node.js and Chrome, starts its own Chrome profile by default, and sends Google usage statistics unless you turn that off (its README says how). See it on claude.com.

    /plugin install chrome-devtools-mcp@claude-plugins-official

  • Install and manage plugins

    docs

    Anthropic's guide to /plugin: marketplaces, install scopes, updating and removing.

  • Plugin security and trust

    docs

    What a plugin can do on your computer, and how to review one before you install it.

  • Ralph, by Geoffrey Huntley

    post

    Where the loop technique behind Ralph Loop comes from.

More skills and plugins →
slide 1 of 8slide 2 of 8slide 3 of 8slide 4 of 8slide 5 of 8slide 6 of 8slide 7 of 8slide 8 of 8

Swipe through the slides →

Carousel · Oct 2026 · 8 slides

Prompting

What's a prompt, really? It's a coffee order

What it covers

Capp told Drip his prompts were too short. A prompt is whatever you type to the AI, and it's most of what the AI knows about the job. It's like ordering coffee: say "a coffee" and you get a coffee, say "a large oat latte" and you get the one you meant. The four things Drip adds now: the task, who it's for, an example (one good one often beats a paragraph) and the shape (a list, three lines, a table). Then a made-up before and after (a text to the boss), a trick for when you're stuck (ask it to ask you three questions first), and fixing one part in the same chat instead of starting over. Anthropic's own prompting guide gives the same core advice: be clear and direct, give context, use examples and say the format you want.

Everything we mentioned

Free courses and prompt libraries →
Watch on TikTok ↗

Short · Sep 2026 · 2:03

Vibe coding

The real risks of an app built with AI, and four fixes

What it covers

Capp's point: an app built with AI can cost you money before its first sale, just mostly not the way the scary videos say. The real risks: database rules left off (in March 2025, researchers at Replit, a Lovable rival, found 170 of 1,645 Lovable showcase apps with data anyone could read; Lovable now scans every publish) and runaway bills (a startup's test code kept calling itself and ran up about $72,000 on Google Cloud despite a $7 budget alert; Google forgave it once). Then the US rules that count per message: promo emails need an unsubscribe link and a mailing address, and a person can sue for $500 per marketing text sent without the required consent. Two scary claims corrected: COPPA covers kids under 13, and a kid signing up to an app not aimed at kids isn't automatically a violation (once you know, get a parent's OK or delete their data); the German Google Fonts case was 100 euros to one visitor, followed by thousands of letters demanding 170. It ends on a four-fix checklist (below). Not legal advice: for your own app, ask a lawyer.

Everything we mentioned

The checklist from the video

The end card, to screenshot. Not legal advice: for your own app, ask a lawyer.

  • Four fixes

    prompt
    1. Table rules on, for every table
   secret keys on a server, never in the page
2. A hard spending stop, not just an alert
   if your host has one
3. Unsubscribe link + address in promo emails
   written consent before marketing texts (US)
4. Self-host your fonts
   if your app targets or tracks people in the EU
App builders in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 1:22

AI tools

How you type a feeling into ElevenLabs v4

What it covers

It opens on a whispered emergency call, and Capp says nobody said it: it was typed into Eleven v4, ElevenLabs' new voice model, every breath of it. The trick is typing the words and then the feeling in brackets, so one voice says "I can't believe you did that" delighted, furious, heartbroken and stunned: the words never change, the shape does (the highs, the lows, the pauses), and ElevenLabs says the model can copy that shape in more than 90 languages. Then the payoff promised at the start: exactly what we typed for that call, tags and all, plus three more scenes made the same way (a nervous confession, a goodbye and a coach's speech), each with its typed text on screen. Not sponsored.

Everything we mentioned

  • ElevenLabs

    tool

    The voice app every line was typed into, on its new Eleven v4 model.

  • ElevenLabs: Introducing Eleven v4

    announcement

    The launch post (Sep 28, 2026): direction typed in brackets, like [laughs], and more than 90 languages.

  • Eleven v4 page

    product page

    Where the bracket tag examples on screen come from.

Voice tools in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 1:39

AI tools

Four movie scenes, every voice typed: ElevenLabs v4

What it covers

Four short movie-style scenes play back to back: a terrified whispered phone call, a nervous confession, a goodbye and a coach's speech. Drip asks who all these people are, and Capp says nobody said those lines: each was typed into Eleven v4, the voice model ElevenLabs released on Sep 28, 2026, with the feeling written in brackets like [shaky breath] or [voice cracking], and the exact text we typed is on screen (ElevenLabs stock voices, best of 3 takes each). Then one voice says the same six words, "I can't believe you did that," four ways: the words never change, the shape does (the highs, the lows, the pauses), and one study found listeners hear at least 24 kinds of emotion in wordless laughs, sighs and cries. ElevenLabs says the model works in more than 90 languages, shown with one voice in English, Spanish and Japanese. Not sponsored.

Everything we mentioned

Voice tools in our AI tools directory →
slide 1 of 14slide 2 of 14slide 3 of 14slide 4 of 14slide 5 of 14slide 6 of 14slide 7 of 14slide 8 of 14slide 9 of 14slide 10 of 14slide 11 of 14slide 12 of 14slide 13 of 14slide 14 of 14

Swipe through the slides →

Carousel · Sep 2026 · 14 slides

Claude Code

Claude can animate now: 12 motion graphics, all code

What it covers

12 motion graphics, each one playing on its own slide: words that slam in on the beat, 608 dots that turn into Drip, one shape that becomes a dashboard, goo that bounces, a word tunnel, a coverflow, a chat that gets stamped, a bar chart race, a shelf where books vanish, a balance scale, a data center that draws itself, and a zoom from a chip to Manhattan. Claude Opus 5.5 built all of them in code in Claude Code, and each took a few rounds of fixes. The prompts below describe each motion so you can make your own.

Everything we mentioned

The prompt for each one

Prompts to try. They describe each motion (every state, the springs, the loop) the way we asked Claude for them; expect a few rounds of fixes, the way we needed them.

  • 01 Words → slam

    prompt
    Make a 6-second looping motion graphic in code (HTML/SVG or Remotion, 30 fps, every frame a pure function of time). Dark background. Four words stacked: EVERY / FRAME / IS / CODE. They slam in one per beat, every 0.32 s: each word starts at 3x scale and a 12° tilt and springs to 1x with a little overshoot, with a quick orange glow on impact and a ring that pulses out from the centre. Last word orange, the rest cream. After the last word, the whole stack punches once and an orange underline sweeps in. Fade out in the last half second so it loops cleanly.

  • 02 Dots → Drip

    prompt
    Make a 6-second looping particle animation in SVG. About 600 dots start scattered and drifting. Give every dot a target point on the outline of a simple character (a coffee cup with two dot eyes, a small mouth and wavy steam above it). From 0.45 s each dot springs to its target with a small random delay, so the character assembles in a wave; hold it breathing slightly; at 4.5 s every dot springs back to where it started. Use one spring function (damped, closed form) for every move. Dark background, cream dots, orange steam.

  • 03 Button → dashboard

    prompt
    Make a 6-second looping 'one shape, never cut' animation. A single rounded rectangle changes state on springs, never swapping for a new element: a dot, then an orange 'Render ▶' button, a drawn cursor clicks it, it shrinks into a cream circle with a spinning loader ring, turns orange with a check mark drawing in, then grows into a white dashboard card with a number counting up to 12,480 and six bars growing in. Then it shrinks back to a dot and disappears. Name every state with its size, radius and colour, and animate position, size, corner radius and colour on springs.

  • 04 Blobs → bounce

    prompt
    Make a 6-second looping SVG animation using a gooey filter (Gaussian blur plus a color matrix alpha threshold). Seven orange blobs orbit the centre, then spring inward and melt into one ball. The ball drops and bounces three or four times with decaying height, squashing on each landing and stretching as it rises, with a soft shadow that grows as it lands. At 4.5 s it splits back into the seven orbiting blobs. Light paper background, a small caption: SQUASH · STRETCH · SPRINGS.

  • 05 Words → tunnel

    prompt
    Make a 6-second seamless loop: a z-tunnel. Streaks of light fly outward from the centre (speeds chosen so the loop is seamless), cream with every fifth one orange. Every 1.5 s a word flies from far away straight at the camera, scaling up exponentially and fading as it passes: SPRINGS, SEEK(t), BEATS, CODE. (last one orange). Dark background.

  • 06 Cards → coverflow

    prompt
    Make a 6-second looping coverflow of 7 portrait cards (use your own screenshots or frames). The centre card faces the camera; the others rotate away in 3D (rotateY about 44°) and shrink as they move out. Every 6/7 s the row springs one card to the left, wrapping around so the loop is seamless. Light paper background, soft shadows, a small caption under it.

  • 07 Chat → stamp

    prompt
    Make a 6-second looping chat animation. A user's orange bubble types out 'Is this book real?' next to a small avatar. Three typing dots pulse on the left, then a grey reply types out 'Yes! It's a bestseller.' A red rubber stamp slams down on top, reading IT ISN'T., starting at 2.5x scale and springing to 1x, while the whole card shakes briefly. Fade out and loop. Dark background, label it MADE-UP EXAMPLE.

  • 08 Bars → race

    prompt
    Make a 6-second looping bar chart race with four racers (a round avatar or initial, a name, a bar and a number). Each racer's value moves between keyframes on springs, and the bars re-order smoothly by value (use a soft rank so they slide past each other instead of jumping). The leader gets a small '▲ 1ST' tag. At the end all bars shrink to zero and the loop restarts. Dark background, bars in four warm colours, label it ILLUSTRATION.

  • 09 Shelf → vanish

    prompt
    Make a 6-second looping animation of a bookshelf. Fifteen book spines in warm colours drop onto the shelf one after another with a springy landing; a headline reads 15 BOOKS. Then ten of them vanish one by one, each tilting, lifting and breaking into a few flying specks, while the headline counts up 1 OF 15 GONE ... 10 OF 15 GONE in red. At the end the rest drop off the shelf and it loops.

  • 10 Scale → tilt

    prompt
    Make a 6-second looping balance scale. A beam on a post with a pan hanging from each end (the pans stay level while the beam tilts). A character stands in each pan. Four small weights drop in one at a time, alternating sides; after each landing the beam springs to its new angle with an underdamped wobble. After the fourth weight it settles level and LEVEL. pops in. Then the weights lift away and it loops. Dark background, cream scale, orange pivot.

  • 11 Blueprint → live

    prompt
    Make a 6-second looping 'blueprint comes alive' animation of a building in plan view on a faint grid. The outline draws itself with a stroke-dashoffset sweep. Then, in stages with a label popping in for each: racks light up in a diagonal wave and flicker (COMPUTE), dashed cable trays start flowing over each row (NETWORK), two power lines pulse into a substation box and battery cells glow (POWER), rooftop fans spin and a cool pipe and a hot pipe start flowing through the rows (COOLING). Fade out and loop. Label it ILLUSTRATION.

  • 12 Chip → city

    prompt
    Make a 6-second 'powers of ten' zoom in one continuous camera, with the world drawn in real metres. Start on one chip, then pull out to the rack it sits in (drawn as its 72 chips), the hall of racks, the building, the campus, and finally a city outline drawn to its real land area with the campus inside it. Animate the camera as a view width in metres, eased in log space on springs, and only draw each layer while it's big enough on screen to read. A label names each level. At the end, rewind fast back to the chip so it loops.
Claude Code in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 1:48

How AI works

Why doesn't AI just say "I don't know"?

What it covers

In 2023 a lawyer asked ChatGPT if a court case it gave him was real. It said yes; it wasn't, and neither were at least five others. In 2025 two big newspapers printed a summer reading list made with AI: 10 of the 15 books don't exist. Why doesn't AI just say "I don't know"? It learns by predicting text, and a fact it saw once has no pattern to lean on. Then it gets graded like a multiple-choice test, and 9 of the 10 popular AI tests OpenAI's researchers checked give zero points for "I don't know", so guessing wins. Inside a Claude model, "I can't answer" is the default, and a familiar-feeling name can switch it off. What helps: tests that take points off for wrong guesses, telling the AI "I don't know" is allowed, and handing it the source.

Everything we mentioned

On YouTube soon

Short · Sep 2026 · 2:09

How AI works

What's really inside AI data centers?

What it covers

A zoom out from one AI chip to a campus built to use as much electricity as about a million homes, then inside one building: rows of racks, over 5,000 cables in one AI rack, grid lines, batteries and backup generators, and liquid piped to the chips. Water: Google says a typical Gemini text question uses about five drops at the data center; all US data centers together used about 17 billion gallons on site in 2023, plus an estimated 12 times that at power plants. The IEA says text chat is a thin slice of the electricity; the bulk must be for training, video and AI agents. Then both sides, each with a source. Costs: data centers caused 38% of the largest US grid operator's latest capacity auction cost; one study says power costs could rise 6-29% by 2030; about a third of one Oregon town's water; tax breaks in most states. The other side: AI used by hundreds of millions every week; utilities making them pay for most of the power they reserve; about 40% of Loudoun County, Virginia's budget; about 1,500 workers to build a building, about 50 to run it. Your call.

Everything we mentioned

Watch on YouTube ↗

Short · Sep 2026 · 2:02

Claude Code

Did Claude make that in one prompt? How the Opus 5.5 animations are made

What it covers

It opens on animations people made with Claude Opus 5.5, every one of them code, then explains how. Claude takes in text and images and puts out text, so to get a video it writes a program that draws any frame on demand, a browser draws every frame, and ffmpeg stitches them together. Some really came from one prompt: the orange showreel was one line in Claude Code, and the Steve Jobs film's creator says one prompt made about 8,700 lines of code. The most-watched one we found had a 9,554-character prompt and, its creator says, 12 hours of Claude's work. We make Drip & Capp the same way: this whole video took Claude a little over three hours of working time and 437 model calls, research included. The creators' posts and every source are below.

Everything we mentioned

Claude Code in our AI tools directory →
Watch on TikTok ↗

Short · Sep 2026 · 2:02

Claude

Chat vs Cowork vs Claude Code: which Claude to use

What it covers

It isn't six products. It's one Claude on one account, and three questions. What do you want back: an answer (chat), the finished thing (Cowork), or software (Claude Code)? Where are you: your phone, a browser, or your computer? What does it need to reach: another app (a connector), something on your computer, or a web page (Claude in Chrome)? Plus what each plan includes, and the one usage pool.

Everything we mentioned

Claude in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 65s

Voice to text

How voice to text works: it hears numbers, not words

What it covers

Drip asks how voice to text knows what he's saying. It doesn't get words: his one question arrives as 45,600 numbers (2.85 seconds at 16,000 a second), gets chopped into slices and stacked into a grid of how loud each note is, and a model guesses the words that most likely made it. Every number and grid in the video is measured from the video's own audio. Then the part newer AI dictation apps add: punctuation, and cutting the slip when you correct yourself.

Everything we mentioned

  • The full film

    video

    How voice to text works, the 9:51 episode this short comes from.

  • Whisper paper

    paper

    OpenAI, 2022. Source for 680,000 hours of training audio and the 16,000-a-second, 80-band setup.

  • Whisper

    repo

    OpenAI's open speech-recognition model.

  • Wispr Flow

    app

    An AI dictation app that removes filler words and fixes mid-sentence corrections.

Dictation apps in our AI tools directory →
slide 1 of 8slide 2 of 8slide 3 of 8slide 4 of 8slide 5 of 8slide 6 of 8slide 7 of 8slide 8 of 8

Swipe through the slides →

Carousel · Sep 2026 · 8 slides

Claude

Bigger model or think harder? The two knobs in Claude

What it covers

Drip thought the model and the effort were the same knob. They're not: the model is which brain answers, and effort is how hard that same model works on this one job. We tested it on September 24 in Claude Code with one puzzle (walks across a 5x5 grid, answer 34), 5 runs per setting, counting the median output tokens. Sonnet 5 used 99 tokens on low effort, 233 on medium, 314 on high and 525 on max, and got all 20 right. On low effort, Haiku 4.5 used 594, Sonnet 5 99 and Opus 5 71, so the smallest model used the most tokens. That's why the cheapest price per token wasn't the cheapest answer: per 1,000 answers (output only, September 2026 prices), Haiku 4.5 came to $2.97, Sonnet 5 $0.99 and Opus 5 $1.78. Capp's rule of thumb: fix depth with effort, fix know-how with a bigger model, judge by cost per finished task, and test it on your own task.

Everything we mentioned

  • Effort

    docs

    The five levels (low, medium, high, xhigh, max), which models support each, and each model's default.

  • Claude pricing

    pricing

    Price per million tokens, in and out, for every Claude model. Prices change; the slides use September 2026.

  • Choosing a model

    docs

    Anthropic's guide to picking the right model for a job.

  • Model and effort in Claude Code

    docs

    How to switch the model (/model) and the effort level (/effort) in Claude Code.

  • Thinking

    docs

    What happens when Claude thinks first, and how effort steers how much it thinks.

Claude in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 1:40

Claude

Claude Desktop explained: what it does that the website can't

What it covers

Your chats are the same on the web, your phone and the desktop app. What the app adds is reach into your computer, and you decide how far: a one-shortcut window on a Mac, your folders, a built-in browser, and on Pro and Max (in beta) your screen itself. It asks before it touches each app. Free to download on Mac and Windows, with Linux in beta.

Everything we mentioned

Claude in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 2:01

Claude

Claude connectors explained: how Claude reaches your apps

What it covers

Claude can't see inside your calendar or inbox until you connect them. A connector is a key to one app: you sign in yourself, Claude gets the same access your account has, and it asks before it sends, replies or forwards an email. Under the hood connectors run on MCP, an open standard Anthropic created. Every claim in the video comes from Anthropic's own help pages, linked below.

Everything we mentioned

Claude in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 2:02

Claude

Claude Cowork explained: hand it the work, get it back done

What it covers

Chat answers you. Cowork does the work: put your meeting notes in a folder, ask for a ten-slide deck, and it comes back finished. It only touches the folders you connect and asks before permanently deleting anything. Since September 16, 2026, Cowork has been merging into regular chat on Pro and Max, so you might not see a Cowork button at all. All from Anthropic's own pages.

Everything we mentioned

Claude in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 2:07

Claude

Claude Code explained: it works inside your project

What it covers

Not copy and paste. Claude Code reads your codebase, edits files, runs commands and tests, and checks its own work: gather context, take action, verify results. It runs in your terminal, VS Code, JetBrains, the desktop app, the browser and even Slack and GitHub. Permission modes decide what it can do without asking, and pressing Escape twice rewinds a change. Included with every paid Claude plan.

Everything we mentioned

Claude Code in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 44s

AI tools

AI tools for making videos: voice, clips, thumbnails and music

What it covers

Drip wants to start making videos and has no idea which AI tools to use. Capp sends him to the Drip the Cup AI tools directory: tap what you want to do and Jev ranks the best fits. The video shows four tasks as Jev ranked them that day: a voiceover (ElevenLabs, then Murf and Fish Audio), cutting a long video into clips (Opus Clip, then CapCut and DaVinci Resolve), a thumbnail (Canva, Ideogram, Recraft) and music (Udio, Suno, Stable Audio). The directory is free and marks which tools are free or open source; at the time it had 27 tasks and 191 tools.

Everything we mentioned

  • The AI tools directory

    directory

    Tap a task and Jev ranks the best fits. Free, and it marks which tools are free or open source.

  • ElevenLabs

    tool

    Realistic voiceover, voice design, dubbing and music. Jev's top pick for a voiceover.

  • Murf

    tool

    Studio voiceovers for training and marketing videos.

  • Fish Audio

    tool

    Expressive voices in 80+ languages with emotion control.

  • Opus Clip

    tool

    Turns a long video into ready-to-post shorts. Jev's top pick for clips.

  • CapCut

    tool

    Quick short-form edits with auto captions.

  • DaVinci Resolve

    tool

    Pro editing and color, with a generous free version.

  • Canva

    tool

    Social graphics and thumbnails from templates. Jev's top pick for a thumbnail.

  • Ideogram

    tool

    Logos and text-heavy designs.

  • Recraft

    tool

    Vector art, icons and on-brand illustrations.

  • Udio

    tool

    Music generation with fine control. Jev's top pick for music.

  • Suno

    tool

    Full songs with vocals from a prompt.

  • Stable Audio

    tool

    Music and sound effects from a prompt.

  • OBS Studio

    tool

    Free recording and live streaming.

  • whisper.cpp

    repo

    Free, private speech-to-text on your own computer.

  • Chatterbox

    repo

    Free open-source text-to-speech you can run yourself.

Browse all AI tools →
slide 1 of 9slide 2 of 9slide 3 of 9slide 4 of 9slide 5 of 9slide 6 of 9slide 7 of 9slide 8 of 9slide 9 of 9

Swipe through the slides →

Carousel · Sep 2026 · 9 slides

How AI works

How Claude actually works, one piece at a time

What it covers

Drip's notes from Capp on how Claude writes: one piece at a time. It reads text in pieces called tokens, scores every possible next piece, picks one, adds it and repeats. It learned by reading a huge amount of text and practising that guess, and later training with feedback on its answers shaped it into a helpful assistant. On hard questions it can think before it answers, which uses more tokens. It only sees what's in its context window: up to 1,000,000 tokens on the biggest Claude models (as of September 2026). A new chat starts fresh unless memory is on, and with tools it can search the web, run code or use your apps (MCP is one way to plug tools in). The token split and the next-word odds on the slides are illustrations, not real tokenizer output.

Everything we mentioned

  • Anthropic glossary

    docs

    Tokens, context window, pretraining and RLHF (training with human feedback), in Anthropic's words.

  • Context windows

    docs

    What the window holds (your messages, files and its own replies) and what happens as it fills up.

  • Claude models overview

    docs

    Window size per model: 1M tokens on the biggest, 200K on Haiku 4.5.

  • Thinking

    docs

    How Claude works a problem through before it answers, and why that uses more tokens.

  • Chat search and memory

    help article

    What Claude's memory saves between chats, and how to see, edit or turn it off.

  • Tool use

    docs

    Claude asks for a tool, the app runs it, and Claude reads the result.

  • Model Context Protocol

    docs

    The open standard for plugging tools and apps into AI.

  • Our MCP explainer

    video

    How your AI talks to your apps, in 5 minutes.

Claude in our AI tools directory →
Watch on YouTube ↗

Short · Sep 2026 · 67s

Design

Same request, five websites: design skills for Claude

What it covers

Drip asks Claude for a site for his coffee club and gets a familiar look: cream background, fancy serif, brown cup. So he gives the same request to five copies of Claude, one with no skill and one per design skill: Taste, Impeccable, UI UX Pro Max and Open Design. Every page in the video is a real build of the same brief. Impeccable's checker flagged 9 AI tells on the no-skill page and 0 on its own. Playwright then opened every page in a real browser at phone and laptop sizes: no sideways scroll on any of them.

Everything we mentioned

  • Taste

    skill

    npx skills add https://github.com/Leonxlnx/taste-skill

  • Impeccable

    skill + CLI

    npx impeccable install. Its checker: npx impeccable detect

  • UI UX Pro Max

    skill

    npx skills add https://github.com/nextlevelbuilder/ui-ux-pro-max-skill --skill ui-ux-pro-max

  • Open Design

    design systems

    Ready-made design systems (we used its cafe one). The full app is a desktop install, see its README.

  • Playwright MCP

    MCP

    Lets Claude open and screenshot your page in a real browser.

  • Web Design Guidelines (Vercel)

    skill

    npx skills add https://github.com/vercel-labs/agent-skills --skill web-design-guidelines

Browse all AI tools →
Watch on YouTube ↗

Short · Sep 2026 · 58s

Claude

Claude Opus 5.5: what's new, and what people built with it

What it covers

Drip asks what Claude Opus 5.5 is. Anthropic says its new model, out September 22, performs at the level of Claude Fable 5.1 on most work and, at default settings on typical work, costs 40% less to run than Opus 5: $4 per million input tokens and $20 per million output tokens (Opus 5 was $5 and $25), with output more than 30% faster. Then four things people built with it, from their own posts: Ryan Sael's interactive lens lab (one shot, 1 hour 26 minutes, $25.66 in API costs), Meng To's boat ride through Japan with rain, lightning, day and night, Techartist's house that grows from a sketch to a finished home in 3D, and Higgsfield's underwater game, built head to head with GPT-6 Sol. On Anthropic's own launch chart it scores 66.4% on Terminal-Bench 4.0, the top score in that table.

Everything we mentioned

  • Anthropic: Introducing Claude Opus 5.5

    announcement

    The launch post: the Fable 5.1 comparison, the 40% cost claim, prices, speed and the Terminal-Bench chart.

  • @claudeai

    post

    Claude's launch post on X, shown at the start of the video.

  • @RyanSael

    post

    The lens lab: 1 hour 26 minutes in one shot, $25.66 API cost.

  • The lens lab (live)

    live demo

    Ryan Sael's lens lab. Move the focus ring and watch the sharp plane shift.

  • @MengTo

    post

    A playable boat scene through Japanese landscapes with weather and day and night lighting.

  • Sakura River Valley (live)

    live demo

    Meng To's boat scene, playable in your browser.

  • @techartist_

    post

    From sketch to home, built in 3D with Three.js.

  • @higgsfield_ai

    post

    Claude Opus 5.5 vs GPT-6 Sol building a 3D game in Unreal Engine.

Claude in our AI tools directory →
slide 1 of 9slide 2 of 9slide 3 of 9slide 4 of 9slide 5 of 9slide 6 of 9slide 7 of 9slide 8 of 9slide 9 of 9

Swipe through the slides →

Carousel · Sep 2026 · 9 slides

AI agents

Create your own WhatsApp agent for your business

What it covers

Give Claude a WhatsApp number and your business gets its own agent: a help desk that answers customers on WhatsApp. Kapso (which calls itself an official Meta Business Partner) gives you the number and Claude Code sets it up with Kapso's skills. Plus WhatsApp forms in the chat and the fine print: the 24-hour window, opt-in, Meta's rule against general-purpose AI assistants, and per-message pricing (replies inside the window are charged from October 1, 2026). Facts checked September 23, 2026.

Everything we mentioned

  • WhatsApp Cloud API

    docs

    Meta's official API for sending and receiving WhatsApp messages from your own code.

  • Webhooks (receiving messages)

    docs

    How new messages reach your code.

  • The 24-hour window

    docs

    Free-form replies only within 24 hours of the customer's last message; after that, approved templates.

  • WhatsApp Business policy (opt-in)

    policy

    Get people's opt-in permission before you message them.

  • Meta terms: AI providers

    policy

    General-purpose AI assistants are not allowed; AI that's incidental to a business is.

  • WhatsApp pricing

    docs

    Per-message pricing; non-template messages are charged from October 1, 2026.

  • Kapso

    tool

    One way in: a WhatsApp number on the official Cloud API, with a sandbox, webhooks and logs. Free plan: 2,000 messages a month plus Meta's fees (Sep 2026).

  • Kapso agent skills

    skill pack

    3 skills (integrate, automate, observe) that teach Claude Code Kapso's WhatsApp API. Labeled alpha.

    npx skills add gokapso/agent-skills
More AI agent tools →
Watch on TikTok ↗

Short · Sep 2026 · 58s

Claude

Claude sped up Claude: how claude.ai got about 3x faster

What it covers

The Claude team says it made claude.ai about 3x faster in a two-week sprint, with an internal research model roughly comparable to Opus 5.5 finding the slow spots, building the benchmarks, shipping the fixes and watching the deploys. Per their post, a fresh page load went from 3,085 ms to 550 ms, and sending a message in Cowork from 928 ms to 48 ms (19x). Two of the tricks: you see a plain copy of the page right away and can start typing while the real app loads underneath, and a single em dash had been making the browser store a whole reply the slow way (UTF-16), so code highlighting took 1.0 s; now it takes 0.35 s. More than 3,000 changes shipped, over 200 on the busiest days, with no customer-facing incident or rollback, and a human approved every change.

Everything we mentioned

Claude in our AI tools directory →
Watch on TikTok ↗

Short · Sep 2026 · 45s

AI tools

AI tools for a small business: meetings, email, expenses and automation

What it covers

Drip's small business eats his whole day, and he keeps his receipts in a mug. Capp hands the boring parts to AI through the Drip the Cup tools directory: tap what you want done and Jev ranks the best fits. The video shows four tasks as Jev ranked them that day: meeting notes (Otter, then Fireflies and Fathom), a faster inbox (Superhuman, then Spark and Shortwave), business expenses (Expensify, Ramp, QuickBooks) and automating work between apps (Zapier and Workato tied at the top, then Make and n8n). It won't do your taxes: it finds you the right tool, links every one, and marks which are free.

Everything we mentioned

  • The AI tools directory

    directory

    Tap what you want done and Jev ranks the best fits. Every tool is linked and the free ones are marked.

  • Otter

    tool

    Live transcripts for meetings and lectures. Jev's top pick for meeting notes.

  • Fireflies

    tool

    Searchable records of every team call.

  • Fathom

    tool

    Free call summaries for Zoom, Meet and Teams.

  • Superhuman

    tool

    An email app built for speed. Jev's top pick for clearing an inbox.

  • Spark

    tool

    A tidy, free email app with AI help.

  • Shortwave

    tool

    An AI assistant on top of Gmail.

  • Expensify

    tool

    Snap a receipt, get an expense report. Jev's top pick for business expenses.

  • Ramp

    tool

    Company cards with expenses sorted automatically.

  • QuickBooks

    tool

    Small-business accounting with AI help.

  • Zapier

    tool

    Connects thousands of apps, no code. Tied for Jev's top pick for automation.

  • Workato

    tool

    Company-wide automation and integrations with enterprise security.

  • Make

    tool

    Visual automations with branching logic.

  • n8n

    tool

    Self-hostable automations with AI steps.

Browse all AI tools →
Watch on YouTube ↗

Short · Sep 2026 · 56s

AI tools

191 AI tools, and Jev picks the right one for you

What it covers

Drip has too many AI tools and no idea which one to use. Capp walks him through the Drip the Cup AI tools directory: 191 tools in 24 categories, one line on what each is best for. Tap what you want to do and Jev ranks the best fits. The video shows three tasks: saving tokens in a coding agent, generating a video, and typing faster by talking.

Everything we mentioned

  • The AI tools directory

    directory

    191 tools in 24 categories. Tap a task and Jev ranks the best fits.

  • Ponytail

    plugin

    Makes Claude Code write less code and spend fewer tokens.

  • Graft

    repo

    One of Jev's top fits for saving tokens in a coding agent.

  • Fast Jev Compaction

    repo

    Uses Jev to compact a coding agent's context.

  • Graphify

    skill + CLI

    Maps a codebase into a knowledge graph so the agent rereads less.

  • Higgsfield

    tool

    Many video and image models in one place.

  • Kling 3.0

    tool

    Text and image to video.

  • Veo

    tool

    Google's video model.

  • Seedance 2.5

    tool

    ByteDance's video model.

  • Wispr Flow

    tool

    Dictation that works in any app.

  • Superwhisper

    tool

    Dictation with local and cloud models.

  • Willow

    tool

    Voice dictation for typing faster.

Browse all 191 tools →
Watch on YouTube ↗

Short · Sep 2026 · 65s

Claude Code

4 free GitHub repos that make Claude Code way better

What it covers

Drip's Claude Code keeps forgetting what they were building. Capp shows him four free repos that change how it works: Superpowers plans and tests before it codes, Context7 hands it current docs, Graphify maps the codebase so it rereads less, and Ponytail makes it write less code and spend fewer tokens. Three are plugins; Graphify is a skill with its own command-line tool.

Everything we mentioned

  • Superpowers

    plugin★ 290K

    Asks what you're building, writes a plan, then works through it with tests at every step.

    /plugin install superpowers@claude-plugins-official

  • Context7

    plugin · MCP★ 62K

    Current, version-specific docs so Claude uses functions that still exist.

    /plugin install context7@claude-plugins-official

  • Graphify

    skill + CLI★ 120K

    Turns your codebase into a knowledge graph Claude looks things up in. Building the graph uses no LLM credits.

    uv tool install graphifyy
graphify claude install

  • Ponytail

    plugin★ 144K

    Makes Claude work like a lazy senior developer. On the author's benchmark (12 real tasks): 54% less code, 22% fewer tokens, 20% lower cost. Read the benchmark.

    /plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail
More skills and plugins →
Watch on YouTube ↗

Short · Sep 2026 · 56s

Jev

Jev vs ChatGPT: this AI doesn't write back

What it covers

Ask a chatbot which of 200 customer emails are frustrated and it writes you a paragraph your code then has to turn back into yes or no. Jev, TypeSafe's new model, reads the same emails and hands back one answer per email with how sure it is, so your code can act on it. It's priced on what it reads, and it can't answer outside your options, though it can still pick the wrong one.

Everything we mentioned

  • Jev by TypeSafe

    early access

    Declare the allowed answers; Jev returns them with probabilities. $0.042 per million input tokens, output not charged.

  • TypeSafe docs

    docs

    The three question types (Choice, Score, Noul), confidence, and worked cookbooks.

  • TypeSafe skills

    skill pack★ 1.9K

    Agent skills for building apps on Jev.

  • 724 ads for 9 cents

    source

    Matthew Berman's post: 724 live ads from 37 brands broken down in 40 seconds.

Try it: our AI tools finder runs on Jev →
Watch on YouTube ↗

Short · Sep 2026 · 74s

Jev

What is Jev? A decision model, not a chatbot

What it covers

Capp explains Jev to Drip, who reads his notes out loud. Chat models write paragraphs; Jev picks an answer and says how sure it is. It has three answer shapes (pick one, score it, yes or no), its 'System One' name comes from the fast, intuitive thinking in Thinking, Fast and Slow, and your code stays in charge of what happens next.

Everything we mentioned

Watch on YouTube ↗

Short · Sep 2026 · 67s

MCP

What is MCP? The one plug between your AI and your apps

What it covers

Capp explains MCP to Drip with a phone charger. The Model Context Protocol is one standard plug between an AI and your apps: each app builds its connector once, and any AI that speaks MCP can use it, instead of every AI needing its own custom connection to every app.

Everything we mentioned

Watch on YouTube ↗

Episode · Sep 2026 · 9:51

Voice to text

Four words. 25,000 numbers. Zero letters. How voice to text works

What it covers

What happens between you talking and words appearing on screen. Your voice reaches the computer as a stream of numbers, around 25,000 of them for four words and not a single letter. Drip and Capp follow those numbers as they're turned into a picture of sound, how a speech model guesses the words, and how today's dictation apps clean up what you actually meant.

Everything we mentioned

  • Whisper

    repo

    OpenAI's open speech-recognition model.

  • whisper.cpp

    repo

    Whisper running fast and privately on your own computer.

  • Speech recognition

    background

    The history, from 1950s digit recognizers to today's models.

Dictation apps in our AI tools directory →