A different kind of Didactic Unit
DU_00 and DU_02 were written after their production wrapped, a finished scene, followed by what was learned. This unit is deliberately different: it's a running log, started before a single clip exists, because the point of documenting ERUDITA's production isn't just the story, it's the real costs and failures of building short AI-avatar video through the HeyGen–Claude–Seedance 2.0 pipeline, in the order they actually happen.
The Podcast series this unit opens has a dual purpose: it gives ERUDITA's ethical questions, cultural heritage, truth, institutional silence, a direct-address format of their own, separate from the game's vertical episodes, which exist to show how the REBOOT game platform itself works, as the interaction between a human player and a virtual character. Both tracks share one ledger in the Didactic Unit Library, because both are evidence of the same thing: what it costs, in time, credits and iteration, to make this kind of platform work.
This page is linked from the REBOOT Game Platform Manual, since the platform is built around exactly these short video stories, so the production record belongs next to the document that explains how the platform works, not off to the side.
The Reception
Deliberately double-edged: the literal formal event the scene takes place at, and how the truth will be received once it's out, the same tension Francisco Esteves puts to the audience directly at the close.
The scene is a formal reception themed around "Cultural Heritage, Ethics and Values in the AI era", colorful and richly lit throughout, a deliberate contrast to Episode 00's true black-and-white archive. Two characters hold a face-to-face conversation and periodically break away to address the camera directly, in the tradition of the political-drama technique where a character mid-scene turns, speaks straight to the viewer in a lower, knowing register, then folds back into the conversation as if nothing happened. Three full cycles of that alternation, six direct-address beats total, build the scene's drama before the final convergence: a third character approaches, is drawn into the middle of the other two, and turns to camera to ask the audience the question the whole piece has been circling.
First appearance for all three
None of these three have appeared in a produced unit before, so whatever's locked in here becomes the reference for every future shot, the same discipline Episode 00 needed for the Guardian's face and wardrobe.
Captain Costa
Built from Walter's own photo and voice for authenticity, not a stock HeyGen avatar. Wears a formal military dress uniform, a deliberate Estado Novo-era visual signal, tying into ERUDITA's existing Esteves Foundation/denunciation-cipher lore. Renamed from an earlier working name that referenced a real historical figure; "Costa" is fictional.
Dra. Catarina Vas
Part of the existing ERUDITA cast, not a new invention. Appears in rich, vividly colored formalwear, nothing muted, to match the scene's colorful treatment.
Francisco Esteves
Minimal dialogue until the very end. Approaches late in the scene, is positioned between the other two, and closes the piece by turning to camera and putting ERUDITA's central dilemma directly to the viewer.
The environment paragraph
Repeated verbatim across every clip in the scene, per the pre-flight discipline established in Episode 00 and Episode 2, one description, no drift between shots.
A grand formal reception hall, richly colored and warmly lit, never desaturated, never monochrome. Polished stone floors, gold-trimmed columns, deep burgundy and cream drapery catching a warm amber key light with soft, even falloff and no harsh shadows. Throughout the room, a crowd of guests in jewel-toned formalwear mingles in loose clusters, softly out of focus in the middle and background. The room reads as an important annual cultural and intellectual gathering, with no readable signage, banners, or on-screen text anywhere in frame. Dialogue audio only, clean and dry, no ambient room noise baked in.
Signage: deliberately atmospheric only, no readable banner or on-screen text in any generated clip. AI video renders on-screen text as garbled almost every time; the room's real theme instead opens the finished piece as a static title card built from ERUDITA's own brand graphic, dissolved into the establishing shot in post.
Camera: the establishing shot is a slow, continuous push through the room, ending on the two leads already mid-conversation. Every direct-address beat holds static, eye-level, centered, medium-close, no camera movement once a character is speaking, matching the discipline used for the Guardian's turn-to-camera in Episode 00.
Sound: ambient room tone and music are added afterward with ffmpeg, the same way the subtitled episodes were finished, not generated by HeyGen, so the dialogue track stays clean and easy to layer under.
Clip-by-clip, not one long shot
HeyGen Avatar Shots/Studio caps a single generation at roughly 15 seconds, confirmed from prior experience on this project. Every beat below is its own clip, stitched together in editing, the same way Sofia and Marta's bar conversation was built in Episode 2. The full script is now written, and the shot count grew from an earlier working estimate of 9 clips to 20, once Francisco Esteves' closing address turned out to run long, see below.
Establishing shot
Slow continuous push through the room, past mingling guests, landing on Captain Costa mid-sentence, finishing a thought toward Catarina Vas, implied just off camera. Eyeline and body stay on her, never the lens, the only line in the piece not addressed to camera.
Costa & Catarina, three cycles
Each cycle is a face-to-face beat and a camera-break beat for Catarina, then the same pair for Costa, repeated three times, escalating each round. Catarina drives the ethical urgency, she was undercover, afraid of what her own knowledge could cost her, and wants the truth told regardless of the UNESCO nomination. Costa is the mediator between her and the Foundation, arguing that timing matters, not that the truth doesn't. His closing line of Cycle 3, "We didn't make the revolution when we wanted it, we made it when we could," is the hinge the finale turns on.
Catarina, silent reaction
A held reaction beat, no script, closing the Costa/Catarina exchange before Esteves arrives.
Convergence, Esteves arrives
Esteves crosses the room and steps into the space between Costa and Catarina, who are positioned as flanking figures, then finds the lens. HeyGen can't render all three performing together in one frame, the "middle of both" positioning comes from blocking and the cut, not one shot.
The finale, Esteves to camera
Esteves' full closing address runs about 146 words, roughly 55-60 seconds spoken, so it had to be split into five sequential clips instead of the single shot originally planned. He names the Foundation's own complicity in the Erudita cipher, frames it as a coerced "you counted or you were counted" choice, a classic collaborator justification, deliberately recognizable as one, then puts the actual decision, hide it, reveal it, or a third way, to the viewer directly: "The one still deciding, here, tonight, is you."
20 HeyGen generations for the full scene, roughly 4 minutes of edited runtime, all in cinematic mode at 720p (matches every prior unit and Vimeo's own serving resolution, 1080p buys nothing here). Four lines (Planos 04, 09, 10, 13) sit close to the 15-second cap at natural delivery speed and need watching when generated. The pilot batch, the establishing shot plus cycle 1 for each character, 5 clips, is meant to go first, so the technique and the room's look get checked before the remaining cycles and the finale are committed.
Entries, in order
This is the part that makes DU_01 different from DU_00 and DU_02, logged as it happens, not reconstructed afterward.
Concept, cast, room and structure defined; production paused before first generation
Started from a reference clip of a political-drama fourth-wall-break technique, supplied for its visual grammar only, described generically here rather than reproduced, since it's copyrighted footage. Translated into a colorful, warmly lit reception hall themed on ERUDITA's own cultural-heritage-and-AI-ethics subject.
Decisions locked in this entry:
- Format: landscape 16:9, matching the rest of the Podcast batch.
- Room signage: atmospheric only, no readable in-frame text, theme instead carried by a real brand-graphic title card, dissolved into the opening shot in post.
- Audio: ambient room tone and music added afterward in post-production, not generated by HeyGen.
- Camera on the establishing shot: a slow continuous push through the room (more ambitious, less predictable than a static hold, accepted as a first test).
- Cast finalized: Captain Costa (Walter's own photo/voice, military dress uniform, Estado Novo signal), Dra. Catarina Vas (existing ERUDITA character), Francisco Esteves (named for the established Esteves Foundation lore).
- Structure: 3 face-to-face/camera-break cycles per character (6 comments total), then a 3-character convergence closing on Esteves' direct question to the audience.
- Naming resolved: an early working name for Captain Costa referenced a real historical figure (the Carnation Revolution captain) and was changed specifically to avoid that, "Costa" is fictional, matching every other ERUDITA character.
Technical constraints confirmed during this entry:
- HeyGen generates one avatar per shot, no multi-character scene can be rendered in a single generation; every "face to face" or "three characters together" moment has to be built through separate singles, cut together in editing.
- A single HeyGen clip is capped at roughly 15 seconds (confirmed from prior production experience), every multi-beat scene, including this one, is built clip-by-clip rather than as long continuous takes.
Status at close of this entry: no clips generated yet. Walter needs to take a reference photo of himself in the military uniform and record his own voice before Captain Costa's avatar can be built. No dialogue or script has been written, this entry covers atmosphere and structure only, by design.
Full script written; shot count corrected from 9 to 20
Wrote every line of dialogue for the scene, cycle by cycle, adjusting the characters' positions twice along the way before landing on the final arc. Captain Costa was first drafted as an institutional obstructor protecting the Foundation's secrets, then briefly rewritten as internally conflicted about his own identity, before settling on his real role: mediator between Catarina and the Foundation, arguing that timing and consequence matter, never that the truth doesn't. Catarina's throughline is the ethical urgency, she was undercover, afraid of what her own knowledge could cost her, and refuses to wait six more weeks for the UNESCO nomination to pass before the truth comes out. Francisco Esteves closes the scene admitting the Foundation's own authorship of the Erudita cipher rather than distancing the family from it, then frames that admission as a coerced "you counted or you were counted" choice, a deliberately recognizable classic-collaborator justification, before putting the actual decision, hide it, reveal it, or find a third way, directly to the viewer.
Correction found while timing the script: Esteves' closing address, as first drafted in one continuous speech, runs to roughly 146 words, about 55-60 seconds spoken, far past HeyGen's single-clip cap. Split into five sequential clips (16a-16e) once this was caught. The scene's real shot count is 20 generations, not the 9 estimated before any dialogue existed, for roughly 4 minutes of edited runtime. Four other lines (Planos 04, 09, 10, 13) sit close to the 15-second limit at natural delivery speed and need watching once generated, trimmed if HeyGen pads them past the cap.
Also resolved this entry: Plano 14 (Catarina's silent reaction) is the only shot in the whole scene without spoken dialogue for lip-sync purposes, every other clip, including Plano 01's tail-end line, carries a script. Whether HeyGen can reliably generate a silent or idle avatar beat for Plano 14 is still unconfirmed and needs testing before the pilot batch.
Status at close of this entry: full 20-clip script locked, timed, and exported as a standalone shooting-script document for production reference. Still no clips generated. Reference photo, voice recording and the pilot batch (establishing shot plus cycle 1) remain the next steps before any HeyGen generation begins.
Resolved so far
- Title, unit number and format locked, "The Reception," DU_01, Podcast Episode 01, landscape 16:9.
- Cast and wardrobe defined for all three characters' first appearance.
- Environment paragraph finalized and ready to repeat verbatim across every clip.
- Full script written and timed, all 20 clips, dialogue locked, exported as a standalone shooting-script document.
- Clip-by-clip shot plan corrected, 20 generations, pilot batch identified.
Still open
- Reference photo, Walter's own photo in the military dress uniform, needed to build Captain Costa's avatar.
- Voice recording, Walter's own voice, for the same reason.
- Silent-shot test, confirm whether HeyGen can generate a no-dialogue avatar beat for Plano 14 before relying on it.
- Pilot generation, establishing shot + cycle 1 for each character, to test the technique before committing to the full twenty-clip plan.
- Credit cost, not yet tracked against the balance, since nothing has been generated.
The real catalog
Not this scene's problems specifically, every category of problem this HeyGen–Claude–Seedance 2.0 pipeline has actually produced across ERUDITA so far, since that's the record worth keeping next to a manual explaining how the platform works. Anything new this unit hits gets added here, in the category it belongs to.
Generation limits. One avatar per HeyGen shot, no scene with two or more characters can be rendered in a single generation, ever; every multi-character moment is built from separate singles and assembled in editing. A single clip caps at roughly 15 seconds. The avatar-chip selector defaults back to the last-used avatar on every fresh page load, silently swapping the character mid-production unless it's reselected before every submission (discovered on Episode 00, cost multiple reshoots). There is no persistent "previous generation" reference, visual consistency across a character's shots is manual discipline only, not something the platform remembers for you.
Prompting / spec omissions. The technical mistakes that are actually preventable with better prompting: voice tone left unstated, facial fill-light left unstated in a dark environment, aspect ratio decided after generation instead of composed for from the start, ambiguity over whether a "silent" beat should be true digital silence or carry ambient sound, and an environment paragraph that drifts slightly between shots in the same scene instead of being repeated verbatim. A newer one, caught while writing DU_01's script: a line of dialogue drafted without checking its word count against the ~15-second clip cap. Francisco Esteves' closing address was first written as one continuous speech, about 146 words, some 55-60 seconds spoken, and had to be split into five separate clips after the fact. The fix going forward: count words (roughly 35-38 words is the safe ceiling at a natural pace) before a line is treated as a single shot, not after.
Post-production fixes. ffmpeg/libass burned-in subtitles rendered at 2-3x the intended size until the video's real resolution was set explicitly; short caption lines wrapped onto two lines until wrap-style was forced to explicit-breaks-only. A comparison video's scene divider was too dark to read until re-colored. Finished files have hit upload-size limits and needed re-encoding.
Publishing / hosting. A unit file missing an explicit `<!DOCTYPE html>`/`<meta charset="UTF-8">` declaration renders garbled text on iOS and Android, caught on the library page itself. WordPress silently renames a re-uploaded file if one of the same name already exists (adding "-1" suffixes), which has broken links in the library more than once until the real live URL was confirmed by hand.
Creative/legal judgment calls. Reference footage from copyrighted material gets described for its visual grammar, never reproduced or quoted at length. A character name was changed specifically because it referenced a real, historically significant person (the Carnation Revolution captain) rather than a fictional one. Any real, identifiable living person's likeness needs their actual knowledge and consent before appearing as an AI avatar, not assumed.
Cost. Every regeneration spends real HeyGen credits. Only the preventable spec-omission fixes above are genuinely "credit-saving" through better prompting, a first look at how a shot's lighting or blocking actually renders is often unavoidable creative judgment, not a prompting failure, and costs a reshoot regardless of how careful the prompt was. DU_01's own budget: Avatar Shots charges a flat rate per clip regardless of its length up to 15 seconds, 60 credits at 720p. At 20 clips, that is 1,200 credits at minimum. Since all three characters make their first appearance here (the same situation that cost Episode 00 multiple reshoots for consistency), a realistic working budget is closer to 1,600-1,700 credits once likely reshoots are counted.
Process limits (Claude's side). No live HeyGen access, every prompt is checked before submission and every clip analyzed after delivery, never generated or watched in real time. Browser automation can drive HeyGen's UI once the user is already logged in, but never enters credentials, never pastes a prompt, and never clicks Generate without the user's confirmation first.