YouTube launch video — script

Working title: "I built and deployed a paid SaaS in 8 hours with Claude Code (and almost broke at hour 7)"

Target length: ~10:30 Aspect ratio: 16:9 1080p Format: Talking head (TH) cut with screen recordings + B-roll, AI Marker brand colors (cream + red ink) used for graphics Tone: Personal, conversational, slightly self-deprecating. Not "tech bro" energy. The honesty about the deploy struggle is the real hook — lean into it. Recording tool: Tella.tv recommended (auto-zooms, auto-captions, talking-head + screen overlay built in)

Notation in this script

•

TH = talking head (you on camera)

•

SR = screen recording

•

B-roll = stock or pre-recorded footage layered over voiceover

•

OVERLAY = kinetic text/graphics on top of whatever's playing

•

CUT = hard cut to next scene

•

VO = voiceover (your face NOT on screen, you narrate)

Bracketed [stage direction] lines are NOT spoken — they're cues for editing.

Scene 1 — Cold open hook (0:00–0:15)

Shot: TH, half-body, hands visible. You're in mid-sentence, energetic. Background: home office or wherever you record. Music: starts subtle — one beat hits on the URL reveal.

"Eight hours.
[pause, hold up open palm]
That's how long it took me to build and ship this."

[CUT to SR: full screen — your live URL ai-marker-rouge.vercel.app loading. Hold for 2 seconds on the hero with "10 minutes" highlighted.]

"A real, multi-tenant SaaS. Sign up, pay, mark a paper, get a PDF back. Twelve weeks of work for a small team — solo, in a Saturday."

[CUT back to TH]

"And I almost gave up at hour seven. Let me show you what happened."

OVERLAY: title card fades in — "AI Marker / 8 hours / 1 person / one bad deploy"

Scene 2 — Who I am, what we're building (0:15–0:50)

Shot: TH, slightly wider framing. Casual.

"I'm Harry. I run an IB tutoring academy in Korea. We mark a lot of student exam papers — chemistry, physics, math, the whole IB diploma. Twenty to thirty minutes per paper, by hand."

[OVERLAY: "Harry — IBtheTOP / GuruMe academy" lower-third while he speaks]

[B-roll: stack of paper exams, red pen, a TA marking — fade to existing 'marked-body.png' showing red ticks on a real student paper]

"I had a working command-line tool that did the marking with AI — but only on my Mac. Useless to my TAs. Useless as a product."

[CUT to SR: show a quick clip of mark.py running in terminal, producing the marked PDF — silent except for keyboard sounds]

"So I gave Claude Code a two-page PRD and said: turn this into a real SaaS that I can sell."

[OVERLAY of PRD.md scrolling briefly — bullet points highlight: "multi-user, library, batch upload, review, billing, deploy"]

"What you're about to see is what actually happened. Honest. Including the bit where it almost broke."

Scene 3 — Hour 1-2: Multi-user auth (0:50–2:00)

Shot: TH transitioning into screen-share. You start on camera, then it picks up your screen.

"Hour one. I asked Claude to add multi-user authentication. Bcrypt password hashing, sign-up forms, per-user data isolation."

[CUT to SR: show the static screenshot-cli-auth.png for ~3 seconds — kinetic camera-zoom into the "Edit api/db.py" line]

"What you see here is one paragraph of plain English from me — and Claude scaffolds six file changes, runs the migration, tests the login endpoint with curl, and confirms it works. Two minutes of compute."

[CUT to SR: live walk-through — open localhost:3000/signup, fill the form, submit, dashboard appears with admin badge]

"I have not written a single line of Python or React by hand at this point."

[CUT to TH — eye contact, deliberate]

"That's the part that took some getting used to. The AI is genuinely faster than me at the boring parts of building a SaaS — auth, billing, deploy configs. I'm just the guy with the taste decisions."

Scene 4 — Hour 3-5: Library, batch upload, review screen (2:00–4:00)

Shot: VO over screen recording. Your face in a small corner.

"Hours two through five. The features that actually make this useful for a TA."

[SR: open the /markschemes library page. Hold 2s. Then drag-drop a markscheme PDF, save with a title.]

"First — markscheme library. Save once, reuse forever. The markscheme tokens get cached, so the second student in a batch is twenty-five percent cheaper to process."

[SR: cut to /new page. Drop in 3 student PDFs. Fill names. Hit submit. Dashboard refreshes with three "marking" rows.]

"Multi-paper batch upload. Three students, one click."

[CUT to TH briefly — emphatic point]

"But the feature I actually care about is this one."

[SR: cut to a /review/<id> page. Use the composited screenshot we have, OR record fresh. Hover over the question panel.]

"Side-by-side review. The AI marks the paper, then I — the teacher — go question by question and either agree or override. One click to override any score. The PDF re-renders instantly with my changes."

[OVERLAY: pull-quote text appears over the review screen — "AI is a draft. The teacher is the verdict."]

"This positioning matters. The product that replaces teachers is the product nobody buys. The product that saves them twenty minutes per paper — that one they pay for."

[CUT to TH]

"And every override I make gets logged to a feedback table. I can come back next month and see exactly which questions the AI gets wrong, retrain the prompt, and ship a better marker. The product gets sharper the more it's used."

Scene 5 — Hour 5-6: Stripe — turning it into a business (4:00–5:30)

Shot: TH then VO over screen.

"Hour five. The part where most 'I built an app with AI' tutorials end. They show you the demo and stop. I wanted real money to actually flow."

[SR: open the /billing page. Hold on the four pricing cards — Free / Starter / Pro / Academy. Caption pill briefly: "Stripe — Free / $19 / $49 / $149"]

"Stripe. Four plan tiers. Free for trial, then nineteen, forty-nine, one-forty-nine a month."

[SR: click Subscribe on Starter. Stripe checkout opens. Use test card 4242 4242 4242 4242.]

"Real Stripe checkout. Test mode for now, but switching to live is a thirty-minute job."

[SR: payment completes. Redirect back to /billing?checkout=success. Page reloads — plan badge flips to "Starter".]

"Webhook fires, plan flips, quota updates from three papers per month to thirty. Five seconds end-to-end."

[CUT to TH]

"And again — I didn't write the webhook handler. I described what should happen — when a subscription completes, update the user's plan field — and Claude built the four endpoints, including signature verification. About fifteen minutes."

"By hour six, I had a real SaaS. Working sign-up, working payment, working data isolation. Now I just had to deploy it."

[Beat — pause for emphasis]

"And that's where it almost broke."

Scene 6 — Hour 7-8: The deploy hell (5:30–7:30)

Shot: TH, slight intensity bump. This is the dramatic act break.

"I picked Render for the backend. Render's a great service — I use it for other stuff. Should be a forty-minute deploy."

[CUT to SR: show pushing to GitHub via terminal, Render auto-deploys, status shows "Live ✓"]

"Build succeeded. Render says live. Container's running. Logs look clean."

[SR: cut to Render dashboard showing "Deploy live" then a curl command in terminal returning 404]

[OVERLAY: "Status: 404 Not Found" stamps on screen, animated with a slight wobble]

"Every. Request. Returns 404."

[CUT to TH — frustrated, slightly amused]

"I lost an hour on this. Suspended the service. Resumed it. Cleared the cache. Rebuilt from scratch. Same result."

[SR: cut through 3-4 of the actual Render screenshots from your Desktop — Settings tabs, log tabs — fast cuts at 1.5s each]

"And then I found this."

[CUT to SR: the screenshot-render-empty-health.png — full screen, hold for 3 seconds. Slow zoom into the empty Health Check Path field. Add a red circle around it.]

[OVERLAY: "Health Check Path: ___ ← empty" with kinetic text]

"Health check path. Empty. The render-dot-yaml in our repo had it set, but Render's dashboard wizard didn't read the YAML field. So Render's load balancer was probing a path that didn't exist, getting 405 Method Not Allowed, marking the service unhealthy, and refusing to route traffic. Even though the container was perfectly fine."

[CUT to TH]

"I had two options. Keep fighting Render, or switch platforms. I picked switch. Fly.io. Same Docker image, totally different host."

[CUT to SR: show terminal — fly deploy command. Build runs. Show the existing screenshot-fly-deploy.png as the final beat, then a curl on the new URL returning 200.]

[OVERLAY: "It just worked." — handwriting font, red color]

"Three minutes. Live. End to end."

[CUT to TH — exhale, relax]

"And that's the part of the day I wanted to share. Because every 'I built an app with AI' video skips this. Real building has these moments. The AI doesn't get tired or frustrated, but I do. The right move at hour seven-thirty wasn't to keep going on Render. It was to switch. The AI saw that calmly. I needed the calm."

Scene 7 — The reveal (7:30–8:15)

Shot: VO over the polished demo video.

[Cue the demo-polished.mp4 — let it play through with its own captions and timing. ~44 seconds.]

"What I ended up with is this."

[Demo plays through, captions appear in sync — "Public landing", "Real marked exam papers", "27 papers already marked", etc.]

"Real product. Real users could sign up right now. Real Stripe checkout. Real marked PDFs."

[At end of demo, hold on the outro card "Try it now — ai-marker-rouge.vercel.app"]

Scene 8 — What it costs to run (8:15–9:00)

Shot: TH with a clean overlay graphic showing a cost table.

"Cost to run, before customers."

[OVERLAY: animated table fills in row by row]

Vercel (frontend)            $0
Fly.io (backend, idle-stops) ~$3/mo
Stripe                       $0 fixed
Anthropic API                ~$0.15 per paper marked
─────────────────────────────────
TOTAL FIXED                  ~$3/mo
Plain Text
복사

"Under five bucks a month. Goes up only when paying users mark papers — and that's already covered by the subscription pricing."

[CUT to TH]

"If I get one paying customer at nineteen dollars a month, that's eight thousand percent gross margin on the fixed costs. The unit economics are obvious."

Scene 9 — What I learned (9:00–10:00)

Shot: TH, looking directly into the camera. This is the takeaway segment.

"Three things I learned.

[Pause briefly between each]

One. The AI is a great pair-programmer for known patterns. Auth, billing, deploy configs — these have been written a thousand times. Claude's solutions are tighter than what I'd write because it knows all the gotchas. Idempotent migrations. Webhook signature verification. I never had to think about bcrypt."

[Beat]

"Two. The AI is a worse pair-programmer for novel product decisions. Should the review screen be spot-check or thorough? Should we soft-delete or hard-delete markschemes? Those are taste calls. The AI offers options. I have to decide. The PRD I wrote up-front made the decisions explicit before we started. That's probably the highest-leverage hour of the whole project."

[Beat]

"Three. The AI doesn't get frustrated. The Render deploy disaster would have eaten a junior dev's evening. Claude just kept methodically diagnosing, then suggested switching platforms when it became clear something was structurally off. The right move wasn't to keep trying. It was to switch. The AI saw that calmly."

[CUT to a slow zoom on you, slight tilt]

"I needed the calm."

Scene 10 — Outro / CTA (10:00–10:30)

Shot: TH, energy back up.

"If you're an IB teacher or tutor — link's in the description. Three papers free, no credit card."

[OVERLAY: URL ai-marker-rouge.vercel.app with subscribe-button styling]

"If you build software — I'm packaging the entire scaffold I used into a Claude Code skill. Same stack, same deploy flow. PRD in, working SaaS out, in roughly an hour. Subscribe and I'll cover that next."

[CUT to outro screen — your branding + "Subscribe" + "Like" + next-video thumbnail]

"And if you've shipped something built with AI — drop the URL in the comments. Curious what people are making."

[OVERLAY: end-card with subscribe button + your other videos]

[Music outro — fade]

Production checklist

Recording day

Tella.tv account set up (free tier is fine to start)

Lighting: face a window in daytime, OR single softbox key light

Audio: AirPods Pro or Lavalier mic ($30 RØDE if you want to invest)

Wear something that contrasts with your background (no plain white shirt on white wall)

Have water nearby; recording 10 min straight is more tiring than you think

Record each scene 2-3 times — pick the best take in editing

Talk slightly slower than feels natural. You can always speed up in editing.

What to have on screen for each SR moment

•

The deployed site at ai-marker-rouge.vercel.app (use incognito for the landing scenes)

•

The Render screenshots already on your Desktop (for the deploy-hell sequence)

•

The polished demo video demo-polished.mp4 (for Scene 7)

•

The CLI mockups screenshot-cli-auth.png and screenshot-cli-debug.png

•

The marked-body.png for B-roll of "real marked work"

Editing tools

•

Tella — primary tool if you record there (handles the camera + screen overlay automatically)

•

CapCut for adding music, advanced transitions, kinetic text overlays the script calls for

•

DaVinci Resolve if you want pro-level color grading

Music suggestion

Low-key, instrumental, ~110-120 BPM. Avoid lyrics. Some free options:

•

Epidemic Sound (paid trial, then ~$15/mo)

•

YouTube Audio Library (free, royalty-free, decent selection)

•

Pixabay Music (free)

Search terms: "indie startup", "tech upbeat", "minimal corporate", "lo-fi build"

Thumbnail

Cover image options (in web/public/marketing/):

•

marked-cover.png — the IB Chemistry paper with "82/90 = 91.1%" stamp

•

A custom thumbnail with: your face on left + the URL on right + "8 HOURS" big yellow text

For YouTube thumbs, bigger text wins. "8 HOURS" or "$3/MO SAAS" works better than your face alone.

Title options for A/B testing

"I built and deployed a paid SaaS in 8 hours with Claude Code"

"From terminal tool to $19/mo SaaS in one Saturday"

"I almost gave up at hour 7. Here's how I built a SaaS in a day"

"I built a real paid SaaS in 8 hours (and the deploy almost broke me)"

Number 3 or 4 are the strongest hooks — the conflict + resolution narrative outperforms straight feature claims.

Description (for the YouTube post)

In one Saturday I took a command-line marking tool and turned it into a paid
multi-tenant SaaS. Real users can sign up, pay via Stripe, mark IB exam papers,
and get teacher-quality red-pen annotations back. I didn't write a single
line of code by hand.

The unedited story — including the part where the deploy nearly broke me at
hour 7.

→ Try it: <https://ai-marker-rouge.vercel.app>
→ Read the full write-up: [Medium URL when published]
→ Twitter/X: [@yourhandle]

00:00 The hook
00:15 What I'm building
00:50 Hour 1-2: Multi-user auth
02:00 Hour 3-5: Library, batch upload, review
04:00 Hour 5-6: Stripe + plan tiers
05:30 Hour 7-8: Deploy hell
07:30 The reveal
08:15 What it costs to run
09:00 What I learned
10:00 Outro

#SaaS #ClaudeCode #IndieHacker #AI #BuildInPublic
Plain Text
복사