What this episode is
A didactic record of how Episode 2 of ERUDITA — the game/ARG layer of the REBOOT project — was scripted, generated, verified and assembled as an AI-avatar video, for use in training other project partners on the same workflow.
ERUDITA: The Price of Silence is the narrative game component of REBOOT, an Erasmus+ project (KA220-HED, No. 2025-1-CY01-KA220-HED-000351980). Episode 2 sits at the emotional turn of the story: Marta has found a Foundation funding certificate bearing her father's name, and confronts what it might mean with her friend Sofia — a journalist who once wrote about the same Foundation. The scene is built from three shots: a pre-existing two-shot recording of Sofia and Marta's conversation, extended in this session with an AI-generated entrance from a third character, Tomás, and closed with an AI-generated direct-to-camera address from Marta.
The point of this document is not just the finished 1:33 video — it is the reusable method behind it: how a scene is broken into shots that a text-to-video avatar tool can actually generate consistently, how continuity (faces, wardrobe, set, light) is held across separate generations, and how each clip was verified before being cut into the final sequence.
Who is on screen
Three avatars, each generated separately across two different shots — their visual identity had to be pinned down explicitly so the AI tool held them consistent.
Sofia
The journalist. Present throughout the scene; asks the question that pushes Marta to keep going.
- Dark curly / wavy hair
- Tie-neck patterned blouse
- Visible pearl necklace
Marta
The certificate is her father's. Carries the scene's emotional weight, and its closing line to the audience.
- Dark wavy hair
- Plain collar under a dark cardigan
- No necklace — the visual tell that separates her from Sofia
Tomás
The saboteur character. Interrupts the two women, and tries — gently — to close the subject down.
- Curly dark hair
- Dark coat and tie
Casablanca, by way of Fellini
Every generation prompt in this episode opens with the same style paragraph, so three separately-generated clips read as one continuous scene.
The brief was explicit: not a pub, not a pastiche — an elegant 1940s bar. Polished dark wood counter. Backlit shelves of crystal decanters. Brass wall sconces and warm hanging globe lamps doing the work that colour would otherwise do. Well-dressed patrons and a bow-tied bartender, always slightly out of focus, so the room reads as inhabited without competing with the leads. True black-and-white, high contrast, no colour cast — closer to a restored film print than a desaturated filter.
An earlier pass at Tomás's entrance came back looking like a generic Irish pub — the fix wasn't a vague note but a fully-specified environment paragraph (see the prompts in §5), repeated near-verbatim across every shot so the set never drifts.
Reference tones pulled from the footage itself — the palette this document borrows for its own design.
The sequence, shot by shot
Three shots, cut together with no visible seam. Shot 1 is pre-existing footage the team supplied; Shots 2 and 3 were generated this session from the prompts in §5.
INT. BAR — NIGHT
Sofia and Marta at the bar, close together, a certificate on the counter between two drinks. Wide two-shot, static camera. This is the footage the team brought into the session — reproduced here as continuity reference, not as a generated prompt.
Opening frameSofia and Marta sit close at the counter, a folded certificate between two glasses of wine. Sofia notices it first.
SOFIA
You kept it. All this time — after everything the Foundation has become — you still carry that certificate with you.
MARTA
My father's name. A Foundation funding certificate. Cultural support.
(reading)
It says... I grew up believing it meant he was honoured.
SOFIA
And now you think it means something else.
MARTA
Now I think it might mean he was paid for something. Or paid for saying nothing.
SOFIA
I wrote about the Foundation once. Called it a guardian of tradition. If I helped build that story — and the story was a lie — then we're both holding pieces of the same silence.
Marta, you don't have to prove your father was part of it.
MARTA
I don't have to. But if I stop here, I'm choosing not to know. And not knowing has a name too.
SOFIA
What name?
MARTA
Convenience. A piece of paper can turn someone you loved into someone you have to question.
SOFIA
Would you still go looking — even if the answer might unmake your own family?
Speaker order above is reconstructed from context and a transcript of the source audio, not a diarised recording — this clip predates the generation session and was supplied as continuity reference. Treat it as the best-supported reading, not a verbatim transcript credit.
TOMÁS ENTERS
The interruption. Tomás crosses the room, both women look up, and the tone of the scene turns from confession to deflection.
0:00 two-shot
Tomás arrives
Settling on MartaTomás enters from the background, crossing the room with purpose and approaching Marta's side. As he arrives, both women look up and turn to him.
TOMÁS
(too loud, delighted)
Sofia! Marta! What is—
(seeing the certificate, cutting himself off)
—ah! A certificate!
(leaning in, gently, almost tender)
Some things are better left where they are, Marta. Your father did what everyone did.
Marta does not answer. Her eyes drop back to the certificate on the counter, expression tightening.
Camera: starts on the two women, moves toward Tomás as he arrives, ends in a close three-shot with him just behind Marta's shoulder — then settles tight on Marta's downturned eyes.
MARTA — TO CAMERA
The scene breaks its own frame. Marta stops speaking to Sofia and speaks to the audience instead.
Still on the certificate
Direct addressClose on Marta's eyes, still on the certificate. Slowly she lifts her gaze from the paper, up into the lens — a direct, quiet, unguarded look at the audience, not at anyone in the scene.
MARTA
(softly, to camera)
Everyone has a certificate somewhere. A name they've never checked.
(a beat)
Would you check yours?
She holds the look into the lens after the line — no reaction shot, no music. Camera holds a moment longer before the shot ends.
Camera: pushes in slowly as she rises into the look; holds still once she reaches it.
What was actually typed into the tool
Both generated shots came from HeyGen's Avatar Shots — Cinematic mode: one freeform field combining set description, blocking, dialogue and camera direction. These are the approved prompts, verbatim, reusable as templates for the next episode.
1940s film-noir bar, Casablanca/Fellini style — elegant and fancy: polished dark wood counter, backlit shelves of crystal decanters, brass wall sconces, warm hanging globe lamps, well-dressed patrons and a bow-tied bartender softly out of focus behind. True black-and-white, high contrast, no color cast. Sofia (left) and Marta (right) sit close at the bar, both looking down at an old certificate on the counter between their drinks. Tomas enters from the background, crossing the room with purpose and approaching Marta's side. As he arrives, both women look up and turn to him. Tomas, seeing the certificate, stops short, eyes widening, delighted and a little too loud: "Sofia! Marta! What is—" then, noticing the paper, cutting himself off — "—ah! A certificate!" His tone shifts, leaning in gently, almost tenderly: "Some things are better left where they are, Marta. Your father did what everyone did." Marta does not answer; her eyes drop back to the certificate on the counter, expression tightening. Camera starts on the two women, then moves toward Tomas as he arrives, ending in a close three-shot with him standing just behind Marta's shoulder, before settling tight on Marta's downturned eyes.
Same 1940s Casablanca/Fellini bar, elegant and fancy — polished dark wood counter, backlit shelves of crystal decanters, brass wall sconces, warm hanging globe lamps, well-dressed patrons and a bow-tied bartender softly out of focus behind. True black-and-white, high contrast, no color cast. Close on Marta's eyes, still looking down at an old certificate on the counter. Slowly she lifts her gaze from the paper, up into the camera lens — a direct, quiet, unguarded look at the audience, not at anyone in the scene. Camera pushes in slowly as she rises into it. Marta speaks softly, directly to camera: "Everyone has a certificate somewhere. A name they've never checked." A beat. "Would you check yours?" She holds the look into the lens after the line — no reaction shot, no music — camera holds still on her a moment longer before the shot ends.
Note the repeated environment paragraph at the top of both prompts, and the single avatar reference for Shot 3 — dropping Sofia and Tomás from the reference set was what fixed an earlier draft where the closing line came out in Sofia's likeness instead of Marta's.
From prompt to delivered cut
The repeatable sequence behind every clip in this episode, for anyone producing the next one.
Pin the avatar references
For each shot, attach only the avatars that should appear in it — up to three per generation. Search each character by name in "My Avatars" and choose the multi-image ("Design with AI" gallery) version for better likeness consistency across separate generations.
HeyGen · Avatar Shots (Cinematic mode)Write one prompt, not a script and a shot list
Cinematic mode takes a single freeform field. Fold the environment description, blocking, dialogue and camera direction into one continuous paragraph, as in §5 — this is what keeps 15-second generations from feeling like disconnected reads.
Generate, then wait out the render
Each Cinematic generation runs ~74–89 credits and often stalls near 99% for several minutes during final encoding — normal, not a failed job.
Pull the direct file, not the player
Read the page's own <video> element for its currentSrc — HeyGen's Cinematic output serves a direct, unsigned URL — and download that file rather than relying on the network panel, which often never surfaces the real media request.
Verify before cutting anything together
Pull a contact sheet (a grid of evenly-spaced frames) from every clip and look at it, frame by frame, against the brief — likeness, wardrobe, set, the direction each character is facing. This is where the Sofia/Marta mix-up on an earlier draft was caught.
Normalise, then concatenate
Clips from different tools rarely match in resolution, frame rate or audio layout. Scale and re-encode every input to the same spec before the concat filter — otherwise the splice drops frames or desyncs audio.
ffmpeg · filter_complex concatCheck frame zero of the actual output file
Not frame ten, not "looks fine in the middle" — the literal first frame of the file you are about to send. This step was skipped once this episode, and is the entire subject of §7.
Two trims, two bugs
Both defects reached the team before they were caught. Neither was a generation problem — both were trim-point errors on the one shot of pre-existing footage, and both were caught by checking the delivered file itself rather than trusting the command that produced it.
The episode opened on the wrong shot
Shot 1 had been trimmed starting 8 seconds into the source recording — cutting off exactly the wide establishing shot of Sofia and Marta seated together, and opening instead on a mid-conversation close-up of Marta.
How it was caught: the team compared the delivered file's actual frame 0 against a screenshot of what was playing in their own player. Matching the two frames byte-for-byte against the source recording located the exact offset — an 8-second trim error, not a splicing or delivery fault.
Fix: re-trimmed Shot 1 from true t=0 instead of the earlier offset, and re-verified the corrected file's frame 0 before re-cutting the episode.
Marta's closing line to Sofia was cut off mid-sentence
The same shot's end trim landed at exactly the same problem from the other direction: cut at t=62.0s, half a second before Marta finishes asking Sofia "Even if the answer might unmake your own family?" — the line ends at 62.64s.
How it was caught: the team flagged the missing words directly. A silence-detection pass across the source audio showed continuous speech from 60.3s to 62.8s with no gap — confirming the original 62.0s cut point landed inside a spoken line rather than a pause.
Fix: re-trimmed the end point to 63.5s — past the full line, into the following silence, but before the source footage fades to black — and confirmed the complete sentence survived by transcribing the delivered file's audio before sending it.
The pattern behind both: a trim point that looks correct in the edit command can still be wrong in the delivered file. Checking a mid-clip frame, or trusting the numbers in the command, is not the same as checking the actual first and last frame — and the actual audio — of the file that ships.
Before any clip ships
- Frame 0, literally: extract and look at the true first frame of the delivered file — not frame 10, not "the start looked fine while scrubbing."
- Last frame too: the same blind spot exists at the tail of every trim.
- Contact sheet every clip: a grid of evenly-spaced frames catches wardrobe, likeness and set drift a single freeze-frame won't.
- Transcribe the audio, don't just watch it: a cut-off word is far easier to miss by ear on a first pass than in a timestamped transcript checked against the intended line.
- Match avatar references to the character, not the scene: a closing shot with the wrong avatar chip attached will render the wrong face, correctly.
- Repeat the environment paragraph verbatim across every prompt in a scene — small wording drift is what produced the "Irish pub" version of this bar.
- Verify byte-for-byte when a delivery is disputed: compare hashes and frame bytes before assuming a re-upload or a splice error — the fault is often further upstream, in the original trim.
What actually goes into two minutes
Written for the REBOOT team as a preface to the case study that follows — reproduced here in full.
There is a fantasy spreading across social media: that you can simply talk to AI and it will produce whatever you imagine. They named it prompting, JSON, context.
It won't.
AI can create the impression that it is talking to you, understanding you, even collaborating with you. But much of that feeling is human projection. Generative AI does not understand in the human sense. It works through extraordinarily sophisticated mathematical models that predict and generate patterns from what they have learned from vast amounts of data.
And that is precisely what makes this moment so extraordinary.
The paradox is that the more powerful AI becomes, the more human literacy matters.
Producing a two-minute cinematic clip is not simply a matter of writing a prompt. You need to understand cinema: framing, camera movement, rhythm, lighting, performance, and editing. You need storytelling, literature, and plot. You may need philosophy to construct the dilemma, design to create the visual language, typography to establish identity, sound to build atmosphere, and psychology to understand how a character should behave.
And before generating a single useful scene, there may already be days of work behind it: creating the characters, defining their profiles, maintaining visual consistency, designing their relationships, placing them within the story, building the plot, preparing locations, establishing the cinematic language, and deciding what information the audience should — and should not — know.
Then comes the AI.
A two-minute clip may take two minutes to watch. It can take days to create.
The real skill is not learning how to ask AI for things. It is learning what needs to exist before you ask.
So start by understanding the tools. Learn the difference between an LLM and multimodal AI. Learn what each can and cannot do. And, above all, develop the cultural, creative, and critical literacy required to direct them.
AI does not eliminate expertise.
It makes expertise visible.
Why vertical can't be an afterthought
Episode 2 was generated entirely in landscape. The two clips below show why that has to be a decision made before generation, not a crop applied after: the finished episode itself, and a labeled side-by-side of the same footage cropped straight down the middle to 9:16.
The center-crop doesn't just tighten the shot — it removes both characters from frame. Because the two-shot was composed with Sofia and Marta deliberately at opposite edges of a wide frame, a straight 9:16 crop keeps only what's in the middle: the bartender, the drinks, the certificate. The scene's entire point — two people talking — is gone.
Close-ups (Shots 2 and 3) survive a center-crop far better, since the avatar's face already sits near the middle of the frame. That's the real lesson: the failure isn't uniform across a scene, it's specific to how each shot was composed — which is exactly why orientation has to be a decision made before generation, not a crop applied after. To be clear, this is not a prompting error: the original 16:9 generation was correct, made exactly as intended for a landscape scene. The problem is purely the mechanical crop applied afterward, which doesn't know or care what's inside the frame.
HeyGen's Avatar Shots panel has a portrait/landscape toggle next to the duration control, set per generation. For a vertical cut of this same episode, that toggle would need to be set to Portrait on every Cinematic generation, with composition prompts rewritten for a narrower frame — closer framing by default, and two-person moments staged as foreground/background rather than side-by-side. Footage from outside HeyGen (like Shot 1 in this episode) can't be fixed this way after the fact at all — it would need to be reshot, or accept the pillarbox treatment instead of a crop.
Finishing the cut: audio, the 1:18 jump, and vertical
Picking up where the lessons-learned section (07) leaves off — the work done after first delivery, including one thing that, in the end, was not done.
Timing the ending cue to her performance, not the clock
The brief was to keep the music under Marta's dialogue and let it grow only once she'd finished — timed to what she does on screen, not to a fixed clock. The cue was set to start at the moment she visibly lifts her head back toward the audience (roughly 1.4s into her closing clip), held at a low, unobtrusive level through the rest of her line, then given a two-and-a-half-second swell once the question lands and she's holding the audience's gaze, arriving at full volume in time to meet a piano chord in the track.
The volume itself was written as a piecewise gain envelope rather than played back raw — quiet under dialogue, ramping through the swell, holding at peak through the chord and the following pause, then fading out over the final seconds as the frame fades to black. Getting the swell and the chord to land on the right beats took several rounds of adjustment against the picture.
The jump at 1:18
A second review surfaced a hard cut around 1:18–1:19 — not an audio problem, but a visual one: Shot 2 and Shot 3 are two separately generated clips, and their framing (zoom, crop) didn't match at the join. The fix was a half-second cross-dissolve across both video and audio at that boundary, which meant re-timing everything downstream of it, since the dissolve shifts the rest of the cut by however long it overlaps.
Building the vertical-crop case study
To make the landscape-vs-vertical problem legible rather than just described, the same eight seconds of footage were rendered two ways — as generated, and center-cropped to 9:16 — and placed side by side with burned-in labels, so the failure (both characters leaving frame) is visible without narration. That comparison is embedded above, in section 10.
What "the final photo" turned out to mean
Worth stating plainly, for the record: the "final photo frame" described in the brief — a real photograph from the Carnation Revolution street parade — was never placed in the delivered episode. The instruction to "bring the photo onto the screen" was built instead as a freeze-frame of Marta's own face, held for three seconds and faded to black. That was a reasonable-sounding reading of the brief at the time, but it was the wrong one.
The episode's video file is now finished and hosted, and per direction it is not being rebuilt to correct this. What follows here is the honest account: Episode 2 ends on a freeze-frame of Marta, not on the archival photograph the brief intended.
Before the next clips are generated
Added after this unit's production, not part of it — every section above stays the historical record exactly as shipped. This one is forward-looking: five checks written directly from Episode 2's and Episode 00's own bugs, meant to be run against a prompt before it goes into HeyGen, not after the clip comes back wrong. The goal is fewer reshoots on the landscape and vertical batches starting now, since every regeneration spends real credits against a balance being tracked closely this round.
- State voice tone as a contrast pair, not a single adjective — "X, never Y." (From Episode 00 Bug 2's fix: "quiet, measured, without warmth" read as flat and creepy; "warm, urgent and grave, protecting the listener, never menacing" is what actually worked, after a full six-shot reshoot.)
- State facial exposure/fill-light explicitly, independent of scene darkness, whenever the environment is dark. (From Episode 00 Bug 2: the room being dark and the face being dark were treated as the same note in the original prompt — they aren't, and the gap cost the same reshoot.)
- Decide and state the target aspect ratio before generating — compose for it, don't crop after. (From this episode's own Format Case Study, §10: shot entirely in landscape, then center-cropped to 9:16, cutting both Sofia and Marta out of frame that a native vertical composition — or a portrait-toggle generation — would have kept in.)
- State explicitly whether a "silent" shot should be true digital silence or carry ambient sound. (From Episode 00 Bug 3: HeyGen's default "silent" render wasn't actually silent — faint page-turning and breath at roughly −45 LUFS, which a later loudness pass amplified by ~27 dB straight into the next clip's first word.)
- Repeat the environment paragraph verbatim across every prompt in a scene. (From this episode's own QA checklist, §08: small wording drift between shots produced a visibly different bar — the "Irish pub" version — inside the same scene.)
Workflow this feeds: draft the prompt → check it against the five items above → generate in HeyGen → send the clip back for analysis (transcript, contact sheet, levels, framing) → only then move to the next shot or assemble the scene.