DEV Community: Peter Kim Frank

Converting old home movie DVDs into a private streaming site

Peter Kim Frank — Wed, 08 Apr 2026 19:46:46 +0000

A relative sent me a few DVDs containing really old home videos that were at one point saved to DVDs. The problem is, I don't own a DVD player, and neither do most people in 2026. These discs were just sitting in a box, slowly becoming unplayable, holding memories nobody could actually watch (which is a real shame).

So I bought a cheap USB DVD drive on Amazon (~$25, an Amicool A11), and figured I'd rip the discs and put the videos somewhere my family could get to them. I'm a vibe coder at this point so I used Claude Code and put together a workflow to rip the discs and upload the videos to private streaming site my whole family can pull up on their phones.

Here's how we did it.

What you need

On the software side, everything is free CLI tools installed through Homebrew: ddrescue for ripping, ffmpeg for video conversion, and wrangler for uploading to Cloudflare. Claude Code handled the commands — I mostly described what I wanted and swapped discs.

Step one: make a safe copy of each disc

The first thing Claude Code suggested was making a bit-perfect copy of each disc before doing anything else. That way even if a disc gets scratched later, we have an exact digital duplicate.

The tool for this is ddrescue (not dd). The difference: if a disc has scratches or bad spots, dd just fails. ddrescue is smarter — it copies what it can, then goes back and retries the problem areas.

brew install ddrescue

There were a couple of macOS-specific gotchas that Claude Code figured out along the way. The disc drive doesn't show up in diskutil list like a normal drive — you need drutil status to detect it. And ddrescue needs to be told the exact disc size, otherwise it thinks there's 9,223 petabytes of data to read (basically infinity). The size comes from the block count that drutil reports.

We wrapped all of this into a script so I could just run ./rip.sh disc-01, wait about 10 minutes, swap the disc, and repeat. Each disc produced a ~4.3 GB .iso file — a perfect clone that I can keep forever as an archive.

Step two: convert to something playable

Mount the ISO and you find a VIDEO_TS folder full of .VOB files. That's the old DVD-Video format — fine for DVD players, useless for phones and browsers.

ffmpeg converts each clip to a modern H.264 MP4:

ffmpeg -i VTS_01_1.VOB \
  -c:v libx264 -crf 22 \
  -c:a aac -b:a 128k \
  -movflags +faststart \
  clip-01.mp4

I didn't know any of these flags — Claude Code picked the settings and explained why each one matters (the -movflags +faststart one moves metadata to the front of the file so browsers can stream it without downloading the whole thing first). My discs had 13 to 21 clips each.

Step three: put it online

You could upload these to YouTube (unlisted), Google Drive, or Dropbox and call it a day. But I liked the idea of having my own thing — files on storage I control, a simple site I can customize, no platform deciding to compress my videos or change their sharing features.

Claude Code set up the hosting on Cloudflare, which turned out to be free for this use case:

Cloudflare R2 stores the video files. It's object storage, like a hard drive in the cloud. You upload files with a CLI command and they just sit there.
Cloudflare Pages hosts a simple website — one HTML file with a grid of video thumbnails.
A small Pages Function connects the two, serving videos from R2 when you click play.

The upload was just a loop running wrangler r2 object put for each clip. Claude Code handled the commands while I did other things.

The one step I had to do manually in the Cloudflare dashboard was connecting the R2 bucket to the Pages site. Under the project's Settings > Bindings, I added an R2 binding so the site's code could actually read from the bucket. Took about 30 seconds but it's not something the CLI can do yet.

To lock it down to family only, there's Cloudflare Access — you add email addresses of people who should have access, and everyone else gets a login screen. Also free for small groups.

The gallery

The site itself is a single HTML page. No React, no build step. Just a dark grid of thumbnails.

Claude Code added a couple of nice touches:

Lightweight thumbnails — instead of loading 47 video players on page load, each clip shows a tiny JPEG thumbnail (~5 KB) with a play button. Click it and the actual video loads. The page loads in under a second.

Hover preview — move your mouse across any thumbnail and it scrubs through the video, YouTube-style. This is done with sprite sheets — one image per clip containing 20 frames tiled horizontally (~40 KB each). No video data loads until you actually click play.

There's a small progress bar at the bottom of each thumbnail as you scrub so you can tell where you are in the clip.

The video manifest is just a JavaScript object that maps disc names to clip counts. When I rip a new disc, I add one line and redeploy.

What it costs

Essentially nothing. Cloudflare R2 doesn't charge for bandwidth — when family streams a video, there's no per-GB fee. Storage is free up to 10 GB, and my 3-disc archive is about 8 GB of converted video. Even if I rip every disc in the box I'd be well under $1/month. For a family casually rewatching old home videos, this is basically a free setup.

The rip script

Here's the script Claude Code and I ended up with. It auto-detects the disc, figures out the size, and runs a two-pass rip:

#!/usr/bin/env bash
set -euo pipefail

if [[ $# -lt 1 ]]; then
  echo "Usage: $0 <disc-label>"
  exit 1
fi

LABEL="$1"
RAW_DIR="$(dirname "$0")/raw"
LOG_DIR="$(dirname "$0")/logs"

DISC=$(drutil status 2>/dev/null | grep "Name:" | awk '{print $NF}')
if [[ -z "$DISC" ]]; then
  echo "No disc detected. Insert a disc and try again."
  exit 1
fi

BLOCKS=$(drutil status 2>/dev/null \
  | grep "Space Used:" \
  | sed 's/.*blocks:[[:space:]]*//' \
  | awk '{print $1}')
DISC_SIZE=$(( BLOCKS * 2048 ))

diskutil unmountDisk "$DISC" 2>/dev/null || true

ddrescue -b 2048 -s "$DISC_SIZE" -n "$DISC" "$RAW_DIR/${LABEL}.iso" "$LOG_DIR/${LABEL}.log"
ddrescue -b 2048 -s "$DISC_SIZE" -r 3 "$DISC" "$RAW_DIR/${LABEL}.iso" "$LOG_DIR/${LABEL}.log"

echo "Done! $(ls -lh "$RAW_DIR/${LABEL}.iso" | awk '{print $5}')"

Pop a disc in, run ./rip.sh disc-04, swap, repeat. About 10 minutes per disc.

The whole project took about an hour of actual thinking work, and a few hours of waiting for discs to rip and videos to upload while I did other things. Claude Code handled the parts I didn't know — the ffmpeg flags, the Cloudflare Pages Function for video seeking, the sprite sheet math — and I handled the part it couldn't: physically swapping discs.

My mom can now watch her old home videos on her phone. That alone was worth it. And I can rest easy knowing these family memories are safely stored in the cloud.

Sorting algorithms visualized

Peter Kim Frank — Mon, 23 Feb 2026 19:48:11 +0000

Wanted an excuse to play with the new Gemini 3.1 Pro so I vibe-coded this sorting visualization playground and deployed it to Cloud Run.

I always thought these sorts of videos were fun so it felt cool to sort of build my own.

Setting up a public URL that flashes my office lights

Peter Kim Frank — Mon, 05 Jan 2026 18:38:28 +0000

I've been running Home Assistant on a Raspberry Pi for over two years. It controls my Hue lights, Zigbee devices, the usual stuff. It's always been a local network thing.

As every vibecoder knows, you've got it working on https://localhost:3000 but that doesn't mean you can show your friends type of thing.

I had this little home network but I never wanted to expose it to the broader internet.

This past weekend, I spun up a $4/month DigitalOcean droplet to run ntfy, a self-hosted push notification service. The idea was to give various automations a way to send push notifications to my phone.

Then I realized: now I've got a droplet which can serve as my internet-facing endpoint. That means I can probably gateway to trigger things on my home network via HomeAssistant

So, I now have a URL that anyone with a token can hit to flash my office lights red - pretty much an IRL ping that something needs my attention.

The Problem

My Pi is behind my home router. I don't want to port-forward or expose Home Assistant directly. But I want to trigger it from the internet.

The droplet solves this. It's public-facing. The question is: how do I get from the droplet to my Pi securely?

The Solution: Tailscale

Tailscale creates a mesh VPN between devices. Install it on both the droplet and the Pi, and they can talk to each other using private IPs (like 100.x.x.x) - no port forwarding needed.

Internet → Droplet (public) → Tailscale → Pi (private) → Home Assistant

What Claude Code Built

I used Claude Code to wire this up. My key insight was that I could simply give Claude Code SSH access to both my Pi and Droplet and let it handle a lot of the rest.

I described what I wanted, and it:

SSHed into my Pi and queried Home Assistant to find my light entity IDs
Wrote a bash script that flashes the lights red, then restores the previous color
Installed Tailscale on both the Pi and droplet
Generated SSH keys so the droplet can run commands on the Pi
Created a Flask webhook with token-based auth
Set up nginx to route requests
Created systemd services so everything survives reboots

The whole thing took maybe 20 minutes. Most of that was waiting for apt to install packages.

The Architecture

Request: GET /flash-peter-office-lights?auth_token=xxx
                    ↓
            Cloudflare (HTTPS)
                    ↓
            DigitalOcean Droplet
            nginx → Flask (port 5000)
                    ↓
            Tailscale (100.x.x.x)
                    ↓
            Raspberry Pi
            SSH → flash_lights.sh
                    ↓
            Home Assistant API
                    ↓
            Lights flash red → restore

The Flash Script

The tricky part is restoring the lights to their previous state. Home Assistant lights can be in different color modes, so the script saves the current state before flashing:

# Save current state
STATE=$(curl -s -H "Authorization: Bearer $HA_TOKEN" \
  "http://localhost:8123/api/states/light.office")
WAS_ON=$(echo $STATE | jq -r '.state')
BRIGHTNESS=$(echo $STATE | jq -r '.attributes.brightness // 255')
XY_X=$(echo $STATE | jq -r '.attributes.xy_color[0] // empty')
XY_Y=$(echo $STATE | jq -r '.attributes.xy_color[1] // empty')

# Flash red
curl -s -X POST "http://localhost:8123/api/services/light/turn_on" \
  -H "Authorization: Bearer $HA_TOKEN" \
  -d '{"entity_id": "light.office", "rgb_color": [255, 0, 0], "brightness": 255}'

sleep 1

# Restore
curl -s -X POST "http://localhost:8123/api/services/light/turn_on" \
  -H "Authorization: Bearer $HA_TOKEN" \
  -d "{\"entity_id\": \"light.office\", \"brightness\": $BRIGHTNESS, \"xy_color\": [$XY_X, $XY_Y]}"

First version only saved brightness. When I told Claude Code "the lights aren't going back to where they were," it figured out the issue and added the xy_color handling.

The Webhook (Flask)

from flask import Flask, request, jsonify
import subprocess
import json

app = Flask(__name__)

def load_tokens():
    with open('/root/webhooks/tokens.json') as f:
        return json.load(f)

@app.route('/flash-peter-office-lights')
def flash():
    token = request.args.get('auth_token')
    if not token:
        return jsonify({"error": "Missing auth_token"}), 401

    tokens = load_tokens()
    if token not in tokens:
        return jsonify({"error": "Invalid token"}), 403

    # SSH to Pi via Tailscale and run the flash script
    cmd = 'ssh -i /root/.ssh/pi_key peter@100.x.x.x "/home/peter/flash_lights.sh"'
    subprocess.run(cmd, shell=True, timeout=15)

    return jsonify({"status": "flashed", "user": tokens[token]["name"]})

Tokens live in a JSON file:

{
  "alice-token-123": {"name": "Alice", "created": "2026-01-05"},
  "bob-token-456": {"name": "Bob", "created": "2026-01-05"}
}

Each person gets their own token. Revoke access by deleting their entry.

What's Next

Now that the plumbing exists, I can:

Different colors for different sources - blue for Slack, green for family texts, red for emergencies
Slack slash command - /flash-peter for coworkers
iOS Shortcut - one-tap button for my wife
Rate limiting - prevent abuse
Logging - who flashed and when

If you want to build something similar, the pieces are: Raspberry Pi (or another device to run HomeAssistant), a cheap VPS with Tailscale on both ends, and some basic Python/bash (in my case, written by Claude).

Welcome Thread

Peter Kim Frank — Fri, 19 Sep 2025 19:17:15 +0000

Welcome! What have you built recently?

The new PostHog.com is pretty amazing ... most distinctive landing page I've seen in a long time.

Peter Kim Frank — Thu, 11 Sep 2025 13:08:04 +0000

PostHog – We make dev tools for product engineers

All your developer tools in one place. PostHog gives engineers everything to build, test, measure, and ship successful products faster. Get started free.

posthog.com

What tools are you using to vibe code?

Peter Kim Frank — Tue, 12 Aug 2025 00:07:53 +0000

Share the tools, platforms, technologies that you're using as you vibe code.

Tip: ask your agent to create a /debug page

Peter Kim Frank — Fri, 11 Jul 2025 23:11:33 +0000

Have been using this "trick" more and more frequently.

When you're running into trouble articulating what's going right or wrong when working with a coding assistant/agent, consider asking it to build a /debug page.

Doesn’t need to be fancy. Just a simple live and visual canvas that shows:

what the agent thinks it’s doing
what inputs it’s working with
what tools are firing (or not)
any recent errors or outputs

There are other ways to do this sort of thing (browser/server logs, MCP with tighter feedback loops, etc), but this has been a quick/dirty trick that has saved me some frustrating debugging sessions.

Especially good when something’s almost working and you can’t quite tell why.

Join the World's Largest Hackathon: $1 Million in Prizes 💸

Peter Kim Frank — Fri, 30 May 2025 19:27:41 +0000

Our friends at Bolt just officially kicked off the World's Largest Hackathon, an awesome competition with over $1,000,000 in prizes!

We love that this hackathon is breaking down barriers and welcoming new participants into the ecosystem. If you’re not yet familiar with Bolt, it’s a powerful AI-platform that enables anyone with an idea to build production-ready applications without traditional coding skills.

Whether you're a seasoned developer or a first-time entrepreneur with a vision, we hope you give this hackathon a try!

How To Participate

Click the link below for all the details on how to get started.

Join the World's Largest Hackathon

Fun Fact: Everyone who signs up for the hackathon will receive a builder pack with thousands of credits to different services - AI voice, UI gen, infra, you name it!

Check out the Hackathon Kickoff Video:

Community Support

The main Hackathon entry process will take place via DevPost, but we’ll be supporting the World’s Largest Hackathon in a variety of ways!

Livestream Events

Starting next week, we'll be simulcasting as many hackathon-related livestreams as possible directly on the DEV homepage. Here are the first three sessions:

Domain Expertise with Entri - June 2, 2025 at 9:00AM PT
So what is an API anyways? With Pica - June 3, 2025 at 9:00AM PT
Building Conversational Video Agents with Tavus and Bolt - June 4, 2025 at 9:00AM PT

Judging

@jess will be bringing her judging experience from DEV Challenges to the event and will be serving as an official judge for the hackathon! She can't wait to check out all the innovative projects that emerge from this event.

Writing Contest Coming Soon

We'll also be launching a writing contest specifically for World’s Largest Hackathon participants later this month, so stay tuned for details on how you can share your building journey and compete for exclusive DEV badges!

What are you planning to build? Drop a comment and let the community know! 👇

Apple Magic Keyboard (USB-C) function keys not working

Peter Kim Frank — Wed, 02 Apr 2025 10:16:00 +0000

Note: I'm basically writing this post for Google because I couldn't find useful/relevant results when I was searching myself...

Tech Specs:

Macbook Pro running Sonoma 14.6.1 on Apple Silicon
New keyboard is model A3118 with USB-C
Old keyboard is model A2449 with Lightning

I've been using a Magic Keyboard with Touch ID (lightning cable) for many, many, many years. It’s been rock-solid the entire with time — no issues, always reliable. But recently, a key on my Lightning version stuck/broke so I decided to get a replacement. I chose a new model (A3118) with a USB-C connector instead of Lightning.

Seemed like a minor upgrade—same keyboard, just with the modern connector, right?

Not quite.

Right out of the box, the USB-C version had issues. It connected fine, typed fine, but none of the system-level features worked. The Touch ID sensor didn’t respond. The function keys—brightness, volume, media—didn’t do anything. I toggled every keyboard setting I could find in System Settings. Nothing changed.

Everything worked as normal on my built-in native keyboard on my Macbook, so I knew it was a firmware/connection issue and not something system wide.

Yes, I tried toggling the Function Keys in Keyboard Settings

I tried everything:

Restarted my MacBook Pro)
Connected via USB-C directly
Paired via Bluetooth
Used it in wired-only mode
Reset keyboard settings and Bluetooth modules

Still broken.

To double-check it wasn’t my machine, I plugged in the old Lightning version. Everything worked perfectly—Touch ID, function keys, all the macOS integrations I’d come to expect.

I brought my Macbook Pro along with the new USB-C keyboard to the Apple Store. They tested it and saw the same issues. They then brought out their store tester keyboards (both USB-C and Lightning) and were able to confirm the issue -- demonstrating that it wasn't a specific faulty unit, but a broader issue. Lightning worked without issue, USB-C has the broken function keys.

So, yeah... if you’re considering upgrading to the new USB-C Magic Keyboard with Touch ID: maybe wait or do a bit more research. There seems to be a firmware or macOS-level compatibility issue that breaks key features.

Hope this helps the next person avoid the painful Google search + troubleshooting issues.

check out my new landing page @ https://localhost:3000 and please send me feedback

Peter Kim Frank — Tue, 01 Apr 2025 14:13:25 +0000

Claude 3.7 Released

Peter Kim Frank — Mon, 24 Feb 2025 20:09:37 +0000

Earlier today, Anthropic released their new model -- Claude 3.7.

Anthropic's Claude Sonnet 3.5 has been one of my most-used models so I'm excited to try it out.

Seems like a really strong model for SWE use cases.

Generating ~450 images for $0.50

Peter Kim Frank — Thu, 20 Feb 2025 15:23:42 +0000

Gallery link: https://pkf-ai-gallery.pages.dev/

I've been getting more and more interested in the Bittensor ecosystem, a decentralized, open-source network interconnected machine learning models.

One of the more interesting "subnets" in the network is Chutes.ai which provides Serverless AI Compute across a bunch of different LLMs, image models, etc. I had listened to them in a recent podcast interview and was looking for an excuse to play around.

I asked DeepSeek-R1 to come up with 50 imaginative prompts:

A celestial cathedral floating atop a spiraling nebula, its stained-glass windows depicting constellations come to life, gilded arches entwined with ivy made of starlight, surrounded by floating orbs of liquid mercury
An opulent steampunk airship shaped like a mechanical peacock, its feathers crafted from interlocking brass gears and glowing amber lenses, hovering above a fog-shrouded Victorian metropolis illuminated by gaslamps
A surreal garden where trees are composed of cascading sapphire ribbons, their roots embedded in pools of liquid gold, guarded by stone serpents with eyes of smoldering opal under a twilight sky streaked with auroras
47 more...

And then ran those through 9 different image models:

Dreamshaper XL 1.0
Flux 1 Dev
Flux 1 Schnell
JuggernautXL
Ostris Flex 1 Alpha
Playground v2.5
Psychedelic Trees
Realistic Vision v5.1
Shitao Omnigen v1

All in all, it cost about $0.50 (~$0.001 per image), which I paid for using $TAO, the native currency of Bittensor.

I used Cursor's new-ish Agent mode to write the Python code to make all of this possible. The entire project took about ~15-20 minutes of tinkering around in Cursor, and then an hour or two to generate all of the images (which I just let run before going to bed).

I then took the list of prompts and image directory and (again) had Cursor generate a gallery that I uploaded to Cloudflare Pages.

You can view the final gallery here: https://pkf-ai-gallery.pages.dev/

Overall, this was a fun little project that has been made possible / much easier through the advent of lower-cost AI models, and code workflow assistants like Copilot, Cursor, Windsurf, Replit Agent, Q Developer, etc.