Behind the scenes

How AI Is Used to Make These

These interviews are produced by one person. That's only possible because AI handles the slow, repetitive parts of the workflow — so the time goes into the stories, not the busywork. Here's exactly how it's done, which tools are used, where AI helps (and where it deliberately doesn't), and roughly what it costs.

Last updated: July 19, 2026 · We'll update this page as the workflow evolves.

01 The short version

One brother, a microphone, and a stack of AI tools that do the work a small production team would normally do.

Recording a good conversation is the easy part. Turning it into something people will actually listen to — cleaned-up audio, several lengths for different attention spans, chapters to jump around, restored photos, and a members-only website to hold it all — is where the hours pile up. Most of that simply wouldn't get done by one person without help.

So AI is used as a force multiplier: it transcribes, cleans, condenses, organizes, restores, formats, and even helps write and build. The result is faster and more thorough than one person could manage alone — not perfect, but good, and real.

The one rule

The stories, the words, the voices, and the faces are real. AI is used to clean up and re-package real material — never to invent it. No fake quotes, no made-up people, no fabricated events. Where AI touches a photo, it's to restore or enhance a real photo of a real person, not to conjure one.

02 What's real vs. where AI helps

Always 100% real

  • The interviews themselves — every word is the brother's own, spoken in a real conversation.
  • Every fact, name, quote, and story.
  • The people in every photo and video — their real faces, in real places.
  • The voices you hear — no voice cloning, no synthetic narration.

Where AI does the work

  • Transcribing and cleaning the audio (removing "um"s, dead air).
  • Condensing the full interview into shorter versions.
  • Finding chapter breaks and writing chapter titles.
  • Restoring and upscaling old, faded, or low-resolution photos.
  • Building and running the website itself.

Generative AI — making new imagery — is used as little as possible. When it is used, it's on real source photos (restoring an old scan, or making a clean portrait from a real snapshot while keeping the person's real likeness and setting). It is never used to fabricate a moment that didn't happen.

03 The audio: from a raw recording to a finished episode

The heart of the project. Here's the pipeline every interview goes through.

  1. RecordGary records the conversation — often across several sessions and a couple of hours.
  2. Transcribe & mergeThe recordings are imported into Descript, automatically transcribed, and stitched into one continuous interview.
  3. Clean upDescript's Studio Sound evens out the audio; an AI editing pass removes filler words, false starts, and long dead-air pauses — with smooth transitions, not harsh cuts.
  4. Condense into multiple lengthsAI edits the full interview down into a 1-hour, 30-, 15-, and 5-minute version (more on this below).
  5. Chapter & organizeAI reads the transcript, finds where the subject changes, and writes chapter titles and timestamps for each length.
  6. PublishThe finished audio is stored privately and streamed to logged-in members through the website's player.
Example — Mike Siegel: three separate recordings (~2h45m of raw tape) became one clean, chaptered 2h40m episode, plus four shorter cuts — produced in an afternoon rather than a week.

Wait — who's actually sitting in the editor?

Nobody. Descript is the editing tool, but a person isn't in it dragging clips around a timeline. Claude operates Descript directly — it connects to the app and runs the work itself: the merge, the Studio Sound pass, the filler-word removal, the condensing into shorter versions, and the chaptering, all from plain-English instructions. The human role is essentially to keep the account and publish the finished files. There's almost no manual editing — the AI is the editor; Descript is just the workshop it works in.

04 The multiple lengths — the part AI makes possible

Full, 60, 30, 15, or 5 minutes. Same story, your amount of time.

People explore stories differently. Some want the whole two-and-a-half-hour conversation; others have five minutes. So every interview is offered at several lengths, and AI figures out how to shrink it.

It works like a set of Russian dolls: the AI takes the full interview and produces a tighter one-hour cut, then condenses that into thirty minutes, then fifteen, then five — each shorter version living inside the longer one. It's told to keep the narrative arc and protect the signature moments, and to cut repetition, tangents, and setup. A person then reviews the result.

Is it perfect? No. An automated edit occasionally trims something it shouldn't, or a transition is a touch abrupt. But it turns a job that would take days of manual editing per episode — and therefore wouldn't happen — into something one person can actually ship. That's the whole point: it's not about replacing careful human work, it's about making the impossible-for-one-person possible.

Chapters & topic cues

Because a listener can lose track of where a subject changes, each version is chaptered (jump to any topic), and a short musical sting bleeds in when the subject shifts. The player photo even changes through the chapters — moving from younger to present-day as the life story unfolds. All of it is generated from the real recording.

05 The words: transcripts, summaries & hooks

The same transcript that powers the audio also feeds the written side. AI is used to:

Every one of these is a different on-ramp into the same real story — for people with different time, or who'd rather read than listen, or who want to skim.

06 The photos — restoration, not invention

Old scans made usable. Real faces kept real.

A lot of the imagery is decades old — faded album pages, dark snapshots, tiny low-resolution scans. AI is used to rescue these:

On generative AI: it's used sparingly — for example, turning a real casual snapshot into a cleaner "portrait," always keeping the person's real likeness and their real surroundings. The line that isn't crossed: AI never invents a person, a place, or a moment. If it's in a photo here, it happened.

Example: the 88 restored shots in Mike's photo history were pulled from scanned album pages, straightened, exposure-corrected, and AI-upscaled — the same people, the same 1980s, just legible again.

07 The videos — assembled by AI, frame by frame

The trip videos aren't just uploaded phone clips — AI cuts them together.

Take the Day 2 video on Mike's page: paddleboarding out of the marina, then into the surf. It started as a pile of raw 360-camera clips. Turning that into a single watchable reel is real video editing — and here AI does the editing itself, not just the planning.

Directed in plain English — "string these clips in order, start at the paddle house, then the ocean, put this song underneath but keep the natural sound, and drop the fade-in" — the AI writes and runs the actual editing commands. Step by step, it:

The heavy lifting runs on FFmpeg — a free, open-source, command-line video engine most people never touch because it's daunting to script by hand. The AI writes those commands, so a non-editor gets studio-style assembly at essentially no cost.

Example — Mike's Day 2 reel: twelve 360-camera clips became one two-minute montage with the natural audio plus a music bed — then re-cut on request (fade-in removed, clip audio kept) in minutes. No timeline, no dragging, no manual editing.

08 The website itself — also built with AI

The site you're reading was built by describing it to an AI coding assistant.

Rather than hand-coding everything or hiring a developer, the entire site — pages, the audio player, the member login, the editable photo galleries, this page — was built by working with an AI coding assistant (Anthropic's Claude, via Claude Code). A person directs it in plain English; it writes and wires up the code, processes the media, and deploys. A few of the pieces it built:

Documented together, by the members

The photo captions aren't locked in. Any logged-in member can edit a photo's title or add a description right on the page — and it's saved for everyone. That turns the archive into a shared project: the person being interviewed can correct a name or a date, and any brother who was actually there can fill in what the AI could only guess at. The more people who pitch in, the truer and richer the record gets — collaboration is the point, not an afterthought.

09 The tools & what they cost

Most of the stack is free or a few dollars a month. Figures below are rough estimates as of mid-2026.

ToolWhat it does hereRough cost
Claude / Claude Code
(Anthropic)
The AI assistant that builds the site, processes media, condenses and edits, and drafts copy. Also powers Descript's editing.~$20–200/mo
(subscription tier)
DescriptTranscription, Studio Sound, filler removal, and the AI editing that condenses each interview into shorter versions + chapters.~$16–24/mo
CloudinaryImage & video hosting/CDN, automatic optimization, face-aware crops, AI photo restore/upscale, video transcoding.Free tier → ~$0–100/mo
(with heavier AI use)
FFmpeg
(open source)
The free video engine the AI scripts to stitch clips, mix in the music, add fades, and compress the trip videos.Free
SupabaseMember accounts + private audio storage (secure links) + the photo-caption database.Free
NetlifyWebsite hosting, contact/signup forms, and small serverless functions.Free (Starter)
ResendBranded signup, approval, and password-reset emails.Free (within limits)
CloudflareThe domain and DNS (deltasigsniu.com).~$10/yr (domain)
GitHubStores and version-controls the site's code.Free
Topaz Photo AIRestoring and upscaling the old scanned album photos.~$99 one-time
Gemini / OpenAI / fal
(image tools)
Occasional photo enhancement — used sparingly, on real photos.Pennies per image
(a few $ total)

Notably, Supabase — which handles all the member accounts and the private audio — is free, as are hosting (Netlify), email (Resend), and code storage (GitHub) at this scale. The only real monthly subscriptions are the AI assistant and Descript.

All in

Depending on the AI-assistant plan and how heavily the media tools are used, the whole operation runs on the order of ~$40–250 a month, plus a domain (~$10/yr) and a one-time photo tool. That's a small fraction of what it would cost to hire out the editing, restoration, and web work — which, realistically, means it would simply never get made.

10 The gear

The physical kit that captures the raw material. Prices are approximate, as of mid-2026 — check the links for current pricing.

Hollyland LARK M2S

~$149

The wireless clip-on microphones that capture clean interview audio — tiny, nearly invisible lavalier mics with noise cancelling and long range, so a conversation sounds good even outdoors or across a room. View on Amazon →

iPhone 16 · Voice Memos

Free app

The recorder itself. The mics feed straight into Apple's built-in Voice Memos app — no special recording rig or software needed. (Uses a phone you likely already carry.)

Insta360 X5

~$480

The 360° action camera behind the immersive video and the "little planet" shots — the paddleboarding, the ocean footage, the overhead backyard view. Waterproof, replaceable lenses, and strong built-in stabilization, so one person can film hands-free and reframe later. View on Amazon →

11 Where AI stops

To keep it plain: AI here is a workflow tool, not a storyteller. It doesn't decide what's true, invent quotes, fabricate people, or replace the real conversations. It removes the friction that would otherwise keep these stories from being captured and shared at all.

If any of this changes — new tools, new techniques, a different approach for a particular episode — we'll update this page. Questions about the process are welcome; that's the whole point of documenting it here.

← Back to the interviews