Deep Dive Development

The complete episode build, start to finish. You fill in one form, seat the panel, generate both rooms, composite the hosts, then generate the intro right on the page. Four steps in, you walk out with every asset an episode needs.

The short-form kit for watch parties and quick reactions. Same flow as the full build but trimmed to the watch-party room only — room, panel composite and the generated intro. In and out.

Two rooms in one. The creator builds new characters from a sentence or a reference. The editor works on the family you have: AI seasonal outfits and the open-mouth speech frames that make them talk on stream — generate, redo with a note, or upload your own. Every result files itself into the Codex.

The live stage. Cut between intro, rundown, set, watch party and outro (Q/W/E/R/T); the rings light over whoever's talking (their real mic in the room, or 1–5 to pulse). Drag a ring to move it, scroll to size it, press the little avatar on a chip to pick its colour and deepen or soften it. Gameplay captures straight into the watch-party screen; nameplates carry the names. The dock's trio: green Hand (hold to raise it — the room sees it, never the tape), yellow Mute, and the red CUT block — a five-second countdown numeral on the tape with every mic silent and the time logged for the edit. Go on air rolls the tape (host only) — the intro, rundown and outro record no microphones. The ≡ on the stage gives Full screen, Theater mode and Main; in the first two nothing on the picture can be moved. Share this window into StreamYard.

Your gameplay or a room fills the frame, your character sits bottom-left and their mouth moves with your real voice — place it once per character and it's remembered. The destinations vault keeps your Twitch, YouTube and Kick servers and keys behind one click, and the whole window records to .webm or feeds OBS by window capture.

The character repository. Every host card carries a portrait, who plays them, and a bio you can click into and edit — changes save on their own, no button. Custom characters land here the moment they touch a canvas.

Drop in a recording of the show. It reads the whole talk (long tapes at 3× speed), finds the twenty strongest self-contained moments and scores them 100 → 1 — only the single best clip may wear 100. Open one in the editor: only the SOUND of the recording is used; the picture is a blank canvas until you add B - Roll (for its seconds) or a Background (an image, or an mp4 that loops under the whole short). Captions pop as you talk in seventeen styles (the looks the big channels use — Hormozi, Beast, highlight tiles, karaoke, comic, mono, glow…), three sizes, drag them anywhere and scroll to size; the title card sits over the open for 5–10 seconds in fifteen styles (the anime bubble by default — Impact, sticker, gradient, chrome, hollow, 3D, glitch and more). Fix any word Whisper misheard in the words box; click a word above the cut and that is the start, one below and that is the end. The Raw Talk emblem is on every short — drag it, scroll to size it. Preview plays and pauses; Export writes a 1080×1920 file. The photo creator at the top makes 9:16 posters of whoever is on for laying text over.

One voice chain the whole program uses — the studio tape and the room both run your mic through it the moment it opens. Enable the mic and the scope shows your raw voice under the treated one; the meter's green band is where a podcast voice sits. Pick a preset (Podcast warm by default) or set it by hand — low cut, body, presence, air, de-ess; compressor, noise gate, trim, a limiter so nothing clips. Record a five-second test and A/B it raw against the equalizer; Hear myself monitors on headphones; Bypass hears the raw mic. It saves itself on this browser. Run the mic sequence from here too — three reads (low, middle, high) that put your voice at the room's loudness and light your green check for the night.

Say what footage you need — "Halo Infinite cutscenes", "true crime courtroom footage" — and it writes three pinpoint YouTube searches, footage only, three angles. Each card finds its own video and puts it in (with a YouTube Data API Key in Keys & models it comes straight from YouTube every time; otherwise it tries free search mirrors, the model's web search, then the Internet Archive). Next video walks the results; Open the search on YouTube and paste a link if you want a specific one. The video plays right in the card; if YouTube refuses to start it there (the app opened as a plain file), the card shows the poster and one big Play that opens the player in its own small window — or run the app with the Deep Dive Development launcher and it plays in the card every time. Record video takes the picture only, no sound; each take lands on the shelf below with a Download.

The show's own editor, for the podcast and the mini podcast. When a take comes off air, Edit this clip takes it straight there with every CUT it carries; any recording can also be loaded from the card. It reads the tape once — the sound becomes the waveform on the bar, so pauses and dead air are plain to see, and each CUT countdown (from the show's log, or found in the picture) becomes a marked stretch starting two seconds before the countdown, where the flub lives. Click the bar to jump anywhere; tick the stretches to lose, drag their ends, add one at the playhead by hand, Play one to check it. Remove asks first — a warning that video sections are about to be erased — then the pieces come out and the show closes up; the preview jumps straight over the seams, and you can keep cutting on the edited show, with Undo for the last removal. Save edited video writes the whole edited show as a new file (real time — it plays it through once). It works on a phone: the timeline takes a finger, the buttons fill the width.

The quote card maker — a 1080 × 1920 poster of the words themselves. The shorts photo creator docks at the top (the same one the clip editor uses, photos and all): generate 9:16 pictures of the cast and press Use it for the quote, or bring any image of your own — no picture at all gets a starry night of its own. Type the words below: a blank line starts a new verse, *stars* make a word large and gold, _underscores_ set a line in italics. Seven looks from the Storybook serif of the poster to Impact meme caps, three sizes plus scroll-to-size, drag the words anywhere on the picture; the signature line with its little gold divider is a checkbox and a name. Export writes the PNG. The whole card — words, style, picture — saves to this browser and is waiting next visit.

One character, any world around them. Pick a face, type the scene — sitting on a throne in a castle with gloomy lights, a neon gamer den, deep space — choose where they look and the mood, and generate. Their design is locked from their portrait; the scene is painted around them, cinematic, with no text anywhere. Three shapes: 1920 × 1080 widescreen, 1080 × 1080 square, 1080 × 1920 tall. Refine any picture, paint it a new scene, or lift the background away for a transparent PNG — then download at full size for Steam, Discord, YouTube, anywhere a face goes.

The pre-flight review, in Extra Tools beside the equalizer and the B-Roll Buster. Load any cut before it uploads, answer the what-only-you-know row — whose voice, whose words, whose footage, whose music — and press Audit. With the shared OpenAI key it listens to the whole tape and looks at frames right across it; without one it grades your answers alone. Every rule comes back broken, close or clear with the fix on the same line: the Partner Program's reused- and inauthentic-content policies, the advertiser rules (graphic imagery, profanity, sensitive events and the rest), the AI-disclosure toggle, copyright reality, and the legal tripwires around real cases and real names. Download the report or copy it. It reads the public rulebooks hard, but it is not a lawyer and it is not YouTube — the last word is always the reviewer's.

A song in, a lyric video out — in Creator Tools. Load the track, give it a title and artist, and get the words timed one of three ways: Hear the lyrics has Whisper listen with the shared key and time every word; Tap them in plays the song while you press Space at the start of each pasted line; Spread evenly is the rough road with no key and no tapping. The words pop up as they're sung — karaoke gold, neon or bold white — over an old-school visualizer that moves with the music: Winamp bars, an oscilloscope, a Milkdrop tunnel or a lava lamp. Preview it, then Render writes the whole song as a 1080p video onto the shelf. Make thumbnail paints art from the song's title, mood and lyrics (or takes the frame on screen without a key) and burns the title in — 1920 × 1080, ready for the upload. Title, artist, lyrics, timings and the look save to this browser.

Two arts for the whole brand, one press. The banners are pure scenery wearing the RAW TALK GAMING title — no characters — and the round profile pictures carry whoever's ticked, blended into the same world with the scene's own light. The kit cuts to every platform's exact size and folds into three coloured dropdowns — YouTube in red, Twitch in purple, Kick in green: eight matching JPEGs, every one inside its platform's upload limit. Download them one by one or as a single zip, then walk the four dashboard doors and change them everywhere in one sitting. Bring-your-own works too — one image for the banners, one for the profiles, no key needed.

In Creator Tools and Extra Tools both. Write the master prompt — what the logo is — drop a reference if you have one, pick a shape and how many (up to ten) and generate. Click the ones worth keeping: they turn green and land in the gallery, where each keeper takes a note ("the eyes look wrong, make them rounder") and Refine paints a new version beside the old one. Every keeper shows the colours it is built from; the palette strip at the bottom pools them all and downloads as a card. Download all gives the whole page as one numbered sheet, or every logo as its own clean PNG.

For the picture that is already right except for one thing. Put it in — drop, click or paste — write the one change (the eyes a touch rounder, the light a little warmer, the scratch on the left gone) and press Adjust. The instruction is wrapped so the model touches only what you named and leaves the rest exactly as it was. Each adjustment builds on the last: Undo and Redo walk the versions, Hold to compare shows the one before, and the original is always underneath. Download PNG saves whichever version is showing.

One step, one shot: type the game, your angle and the set piece, then tap hosts to seat them — the number on each card is their seat, left to right. Hit Build the episode and every prompt below is written; the seat count even sets the number of empty chairs in the rooms.

Always before any thumbnail. Both prompts produce an empty set — chairs, table, atmosphere, and a hole where the panel goes. Generate two versions and they come back with your panel already seated, so you judge the finished shot; show-empty reveals the raw room. Pick the keeper and it drops into the studio. The ChatGPT buttons are the backup route.

Load the room and your seated panel drops into the chairs on its own, in seat order. Dial anyone in — size, rim light, shadow, flip, layer order — type the two title lines, burn them in, and export at 1920×1080 or flip to Shorts for 1080×1920. Generate 2 AI thumbnails sends the finished shot through your OpenAI key and returns two polished takes — pick one, burn the title, export. Arrow keys nudge, Delete removes, and the safe-zone guides never print.

The same three beats every episode — the eyes ignite, the lion forms out of the episode's world and roars, the skull and the title land — and the title is always EXACTLY "RAW TALK GAMING". Generate it right on the page with Sora: Draft at 720p to try skins, Final at 1080p for the keeper, 8 or 12 seconds, an optional locked opening frame, and a hard anatomy law — one lion, one head, on the neck — so the shift stage never glitches. The finished tape lands on a shelf with a Download. Two folded dropdowns at the top carry the paper roads: the intro sheet, and Show master prompt — the full master prompt — the exact prompt the Generate button sends, following the episode and the length pick — with a Copy button for pasting into Gemini/Veo, Kling, fal or any other video maker if Sora's take isn't hitting.

Your OpenAI key powers images, transcripts, shorts, the Video Auditor and the generated intro; the YouTube Data API Key lets the B-Roll Buster place videos straight from YouTube. By default both live only in this tab — but tick Remember the keys on this computer and they are kept in this browser profile and auto-filled every visit. They are never written into the app file and never sent anywhere but their own service. Check what keys work checks whatever is pasted — and a restricted-but-working key now reads green, not red; Forget the keys wipes them everywhere, the remembered copy included. The drawer's two circles tell the truth at a glance — red with no key or a dead one, green when a key works — and the image model and quality are locked to the best, no picking. The spend counter is an estimate.

Open the room in Pod Forge and Deep Dive Development hands you a five-character code. Anyone with that code types it at the front door — "Join a room" on the login screen — and lands straight in the show: your stage picture, the show's sound, their own mic, and nothing else from the tool. Their voice drives their ring on the set and the watch party. They can leave whenever they like; you can take anyone off the panel with the ✕ on their row, or close the room and send the lot home.

Codex bios and played-by names save to this browser the moment you type them and survive a refresh. Everything else — rooms, composites, custom characters — lives for the session, so export what you want to keep.