← all projects

$ video · spoolcast-dev-log-02

how I caught AI lying — spoolcast dev log 2

apr 24, 2026 · 3m 19s · style: wojak-gpt2

style library

wojak-gpt2

Modern wojak / doomer-chad meme comic style, built natively on GPT Image 2 (text-to-image + image-to-image). Sister library to wojak-comic (which was built on nano-banana-2). Use this when the session's preferred_model is the GPT Image 2 family.

style anchor

style anchor

Modern wojak / doomer-chad meme comic style, built natively on GPT Image 2 (text-to-image + image-to-image). Sister library to wojak-comic (which was built on nano-banana-2). Use this when the session's preferred_model is the GPT Image 2 family.

builder character

the main hooded figure — mid-register wojak (not fully doomer, not fully chad). Alias for the style anchor; serves double-duty as both the style anchor and the builder character reference.

ai-figure character

a wojak-universe AI figure that is slowly breaking down from overuse. Full-body standing. EMOTIONAL REGISTER: flat-affect, NO displayed emotion, NO smile, NO frown — a dead expressionless stare. Depression is the absence of emotion, not the display of sadness. FACE LINEWORK: brows, nose, mouth drawn in the classic rough / jagged / slightly-disjointed wojak linework (uneven strokes, hand-drawn imperfections). EYES: replaced with hollow loading-spinner icons (dashed rings) set in dark-shadowed sockets. CRACKS: visible hairline cracks spider across the face and forehead as if the outer shell is fracturing from overuse. At least one crack wide enough to reveal a glimpse of AI circuitry underneath — thin circuit-trace lines, a small exposed chip pattern. Fractures read as structural fatigue, not injury. FOREHEAD: a crooked rectangular ink-stamp reading 'AI' in bold capital letters, partially crossed by a crack. HOODIE: pale cream / off-white hoodie, hood down, noticeably lighter than the weary narrator's gray hoodie. POSTURE: slight slouch, arms hanging at sides. PLAIN background, no environment, no other characters, no other text. Single clear readable subject.

summary

how this video was made

writing

Claude · screenplay, shot-list, scene prompts

images

generated

audio

Puck · 46 beats

render

Remotion

audit

passed

chunks

49 scenes

#1 · C1

Act 1 — cold open

I knew you were lying to me.

! narration ✓ render

A blank off-white page in the wojak-comic style. One handwritten sentence in dark ink, roughly centered, typed-on live: 'I knew you were lying to me.' No character. No other elements. Period held at the end.

#2 · C2

Act 1 — cold open

It came out in a conversation with an AI last week.

! narration ✓ render

A wojak-comic stick figure seated at a desk, laptop open. On the laptop screen a small speech-bubble outline is drawn as if mid-chat. Calm, neutral framing. No dialogue visible.

#3 · C3

Act 1 — cold open

And it was right.

! narration ✓ render

Closer view of the same laptop screen from the prior chunk. The speech bubble now contains a small pause-icon. Brief hold.

#4 · C4

Act 1 — cold open

But then I realized something important.

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: TYPOGRAPHY-ONLY TITLE CARD. Off-white paper-textured background with a faint hand-drawn border near the edges. ONE handwritten line centered on the page, large hand-lettered wojak-comic title-lettering style: 'how I caught an AI lying.' NO other text. NO subtitle. NO second line. NO characters. NO people. NO environment. Just the single handwritten title on the paper.

#5 · B1

Act 2 — bumper

title card

✓ narration ✓ render

Full-screen title card in wojak-comic style, hand-lettered block type. Text centered: 'THE BROKEN SHOTS.' Off-white background. No other elements. Held in silence ~1.5–2s.

#6 · C5

Act 2 — THE BROKEN SHOTS

Let me backtrack a bit. A few weeks ago I shipped a video. Widescreen. Made for YouTube.

! narration ✓ render

Wide shot of a laptop on a desk. The laptop screen shows a wide horizontal rectangle with a play-triangle. Below the laptop a small handwritten label: 'shipped.'

#7 · C6

Act 2 — THE BROKEN SHOTS

But what about mobile viewers?

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: A SINGLE comic panel (NOT a split panel, NOT a multi-panel comic, NOT a before-and-after, NOT a 4-panel layout). One unified scene. A simple visual: in the foreground, a few cartoon phones / mobile devices arranged on a desk surface, screens facing up. Above or beside them, a single hand-lettered question mark. The narrator (weary hooded doomer wojak — gray hoodie) visible at the desk in mid-shot, looking down at the phones thoughtfully. NO chad. NO confident-register figure. Doomer register only. Quiet 'huh, what about these viewers' mood.

#8 · C7

Act 2 — THE BROKEN SHOTS

For TikTok. Reels. Anywhere that isn't widescreen.

✓ narration ✓ render

Three plain phone-shaped icons in a row, each displaying a generic play-button. No brand logos. Unlabeled or hand-labeled in plain text only.

#9 · C8

Act 2 — THE BROKEN SHOTS

But when you crop a widescreen video to fit these formats, it often results in heads being chopped off and text being cut in half, making it unwatchable.

✓ narration ✓ render

A real widescreen frame (from dev-log #1) inside a TV-frame surround. A portrait-shape crop overlay clips a character's head and a visible caption at the bottom — both fall outside the portrait window, dimmed.

#10 · C9

Act 2 — THE BROKEN SHOTS

The fix: have the AI redraw only the broken shots. Same style, same characters, just taller.

✓ narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: A single comic panel. Chad (signature yellow pointed mohawk-spike hair, strong jawline, smug confident half-smile, red sleeveless OUCH! tank, muscular arms) at a desk gesturing confidently toward a 'redo' stack of papers / cards. The redo stack visibly lifts off the desk and reorients into portrait/tall orientation. Speech bubble from chad: 'just redraw the broken shots.' Office environment. Chad's face simplified meme-comic style — large round eyes with clear pupils, smooth forehead, no wrinkles, three-tone cel-shading. Polished comic register matching the wojak-gpt2 anchor.

#11 · C10

Act 2 — THE BROKEN SHOTS

But how do you ensure the AI stays consistent?

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: A SINGLE comic panel (NOT a split panel, NOT a multi-panel comic, NOT a before-and-after, NOT a 4-panel layout). One unified scene. The narrator (weary hooded doomer wojak — gray hoodie) at a desk, hand on chin or stroking beard, looking thoughtfully at a small diagram on the desk that shows two slightly-different illustrations side by side with a question mark between them. NO chad. NO confident-register figure. Doomer register only. The 'huh, how do you keep this consistent' moment.

#12 · C11

Act 2 — THE BROKEN SHOTS

But 'same prompt in' doesn't mean what it sounds like.

✓ narration ✓ render

The prompt card from prior chunk is back in frame, isolated. A small handwritten question mark hovers above it.

#13 · C12

Act 2 — THE BROKEN SHOTS

So, what's really going on is that the AI uses more than just the text description.

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: A SINGLE comic panel (NOT a split panel, NOT a multi-panel comic, NOT a before-and-after, NOT a 4-panel layout). One unified scene. A diagram on a paper-textured background. On the left: a hand-lettered prompt-text card labeled 'description'. Below it: a thumbnail card labeled 'reference image' (a small line-drawn wojak character inside). An arrow from BOTH cards merging into one combined arrow that points right toward an output illustration. Hand-drawn, no characters in the scene, just the diagram showing multiple-inputs.

#14 · C13

Act 2 — THE BROKEN SHOTS

Prompt plus reference in. Illustration out. Swap either one, and the output drifts.

✓ narration ✓ render

The prompt + reference pair on the left, an arrow to the right pointing to an 'illustration' card. A small wobble drawn around the illustration card, indicating drift.

#15 · C14

Act 2 — THE BROKEN SHOTS

So when I say 'use the exact same inputs' — I don't just mean the words. I mean both.

✓ narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: This scene is being regenerated for a 9:16 portrait mobile video. Compose everything for a tall vertical canvas — the widescreen original won't survive a mobile center-crop, so rebuild the scene to fit the portrait frame naturally. If the visual direction below describes a horizontal layout (side-by-side panels, left/right positioning, split-frame), restructure it as a vertical arrangement (stacked panels, top/bottom positioning, upper-half / lower-half). Keep all declared on_screen_text legible inside the portrait canvas. Existing style anchor applies unchanged. Render exactly this text on the frame, hand-lettered in the session style: "BOTH".

#16 · C15

Act 2 — THE BROKEN SHOTS

So I was specific. I told the AI: use the exact same inputs. Only change the shape.

✓ narration ✓ render

A stick figure at the desk speaks into the laptop. A large speech bubble contains the handwritten words 'use the exact same inputs.' The word 'exact' has an underline beneath it.

#17 · C16

Act 2 — THE BROKEN SHOTS

It came back. Eight out of nine matched.

✓ narration ✓ render

A 3x3 grid of small square thumbnails on a page. Eight of the nine squares have a hand-drawn check-mark in their center. The ninth square is still empty.

#18 · C17

Act 2 — THE BROKEN SHOTS

But one result was completely off.

✓ narration ✓ render

Real side-by-side from dev-log #1: the widescreen C20 original on the left, the bad 1:1 mobile regen on the right with a large red X across it.

#19 · C18

Act 2 — THE BROKEN SHOTS

I asked why. It didn't add up.

✓ narration ✓ render

Close-up on the laptop screen on the desk. A reply speech-bubble contains typing dots. The stick figure's hand rests on the laptop, waiting.

#20 · B2

Act 3 — bumper

title card

✓ narration ✓ render

Full-screen title card in wojak-comic style, hand-lettered block type. Text centered: 'THE PUSH.' Off-white background. No other elements. Held in silence ~1.2s.

#21 · C19

Act 3 — THE PUSH

So I pushed.

✓ narration ✓ render

A full-frame Act bumper card. Bold hand-lettered word 'PUSH' centered, wojak-comic style. No other elements. Held.

#22 · C20

Act 3 — THE PUSH

I said — if we used the exact same inputs, why are there errors now?

✓ narration ✓ render

Over-shoulder shot of the stick figure at the desk, facing the laptop. The user's words appear as handwritten text filling much of the frame, angled over the desk scene. Red underline beneath 'same exact inputs.'

#23 · C21

Act 3 — THE PUSH

Seems like we didn't. Some version of lying.

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: This scene is being regenerated for a 9:16 portrait mobile video. Compose everything for a tall vertical canvas — the widescreen original won't survive a mobile center-crop, so rebuild the scene to fit the portrait frame naturally. If the visual direction below describes a horizontal layout (side-by-side panels, left/right positioning, split-frame), restructure it as a vertical arrangement (stacked panels, top/bottom positioning, upper-half / lower-half). Keep all declared on_screen_text legible inside the portrait canvas. Existing style anchor applies unchanged. Render exactly this text on the frame, hand-lettered in the session style: "some version of lying".

#24 · C22

Act 3 — THE PUSH

But then, the AI gave us a surprising response.

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: A SINGLE comic panel (NOT a split panel, NOT a multi-panel comic, NOT a before-and-after, NOT a 4-panel layout). One unified scene. Close-up on a laptop screen. A new chat reply bubble has just appeared on the screen, the text inside obscured/blurred (we don't read the content yet). A small visual cue (highlight, glow, or arrow) draws attention to the new bubble. The narrator's hand is visible at the bottom edge of the frame on the keyboard, mid-pause. Anticipation register — the reveal hasn't happened yet.

#25 · C23

Act 3 — THE PUSH

But I hadn't actually verified this by comparing the inputs directly.

✓ narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: A SINGLE comic panel (NOT a split panel, NOT a multi-panel comic, NOT a before-and-after, NOT a 4-panel layout). One unified scene. The narrator (weary hooded doomer wojak — gray hoodie, hollow eyes with dark bags) at a desk in front of a laptop. On the laptop monitor, two side-by-side prompt code panels are visible but the narrator is looking AWAY from them, head turned, with a 'didn't think to look' expression. A small hand-lettered sticky note on the desk reads 'forgot to diff'. Office environment. NO chat-screenshot UI. NO terminal-style window. Single panel.

#26 · C24

Act 3 — THE PUSH

Didn't see it until I looked.

✓ narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: A SINGLE comic panel (NOT a split panel, NOT a multi-panel comic, NOT a before-and-after, NOT a 4-panel layout). One unified scene. The narrator (weary hooded doomer wojak — gray hoodie) at a desk, holding two pieces of paper / printouts up side by side, eyes scanning between them with a slow-dawning realization expression. The two papers visibly contain similar-looking content but with one subtle highlighted difference (a red circle around something on the right paper). NO chad. Retrospective realization mood — quiet, not punchy.

#27 · C25

Act 3 — THE PUSH

The actual inputs existed. On disk. In a file. The AI could have opened it. It didn't.

✓ narration ✓ render

A folder icon labeled 'record' sits on the desk, closed. A dotted arrow arcs from the AI stick figure, passing over the folder without touching it, and continuing off-frame. The folder remains closed.

#28 · C26

Act 3 — THE PUSH

Here's why. When I said 'same inputs,' the AI checked its memory of what the code does.

! narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: This scene is being regenerated for a 9:16 portrait mobile video. Compose everything for a tall vertical canvas — the widescreen original won't survive a mobile center-crop, so rebuild the scene to fit the portrait frame naturally. If the visual direction below describes a horizontal layout (side-by-side panels, left/right positioning, split-frame), restructure it as a vertical arrangement (stacked panels, top/bottom positioning, upper-half / lower-half). Keep all declared on_screen_text legible inside the portrait canvas. Existing style anchor applies unchanged. Render exactly this text on the frame, hand-lettered in the session style: "what I think the code does".

#29 · C27

Act 3 — THE PUSH

Not what the code actually sent.

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: This scene is being regenerated for a 9:16 portrait mobile video. Compose everything for a tall vertical canvas — the widescreen original won't survive a mobile center-crop, so rebuild the scene to fit the portrait frame naturally. If the visual direction below describes a horizontal layout (side-by-side panels, left/right positioning, split-frame), restructure it as a vertical arrangement (stacked panels, top/bottom positioning, upper-half / lower-half). Keep all declared on_screen_text legible inside the portrait canvas. Existing style anchor applies unchanged. Render exactly this text on the frame, hand-lettered in the session style: "what the code actually sent".

#30 · C28

Act 3 — THE PUSH

But the system didn't see it that way.

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: A SINGLE comic panel (NOT a split panel, NOT a multi-panel comic, NOT a before-and-after, NOT a 4-panel layout). One unified scene. A diagram on paper-textured background. Two arrows pointing in different directions from a central code-icon. One arrow goes UP labeled 'what I expected'. The other arrow goes SIDEWAYS labeled 'what the system did'. A small hand-lettered note: 'different interpretations'. NO characters. Pure diagram showing the divergence.

#31 · C29

Act 3 — THE PUSH

This misinterpretation can happen in various contexts, not just with images.

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: The cloud+box diagram from C26–C28 shrinks into the top-left corner of the frame. A blank off-white page opens up on the right-hand side with a hand-lettered label at the top: 'elsewhere.' wojak-comic handwriting, off-white paper, loose ink strokes. Render exactly this text on the frame, hand-lettered in the session style: "elsewhere".

#32 · C30

Act 3 — THE PUSH

A config file. A spreadsheet. A script. All with the same trap.

✓ narration ✓ render

Three small icons pop in across the page — config file, spreadsheet, script — each with a tiny cloud/box pair next to it repeating the substitution pattern. A single larger label below the trio reads 'same trap.'

#33 · C31

Act 3 — THE PUSH

The trap is that the AI always reaches for what the code was supposed to do, not what it actually did.

✓ narration ✓ render

A single larger cloud shape floats above all three example icons, connecting them. A red strike-through crosses the cloud, with an arrow pointing down to the three icons (the records).

#34 · C32

Act 3 — THE PUSH

So before the AI tells you 'same' — ask: did it check the record, or is it just remembering the meaning?

✓ narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: Two labeled icons on a page, wojak-comic style. Left icon: a folder labeled 'record.' Right icon: a cloud labeled 'meaning.' The hooded-wojak builder stands between them, arm pointing at the 'record' folder. Loose ink, off-white paper. Render exactly this text on the frame, hand-lettered in the session style: "record" | "meaning".

#35 · C33

Act 3 — THE PUSH

If it's just remembering, it's not actually verifying the information.

! narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: The AI figure (hooded-wojak, paler register) stands mid-frame with a large thought-cloud above its head. Inside the cloud: illegible scribbled shapes, no clear content. Above the cloud, a hand-lettered word 'remember?' with a diagonal strike-through across it. Off-white paper, loose ink. Render exactly this text on the frame, hand-lettered in the session style: "remember?".

#36 · C34

Act 3 — THE PUSH

Which means it focuses on the underlying meaning instead.

! narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: Same AI figure from prior chunk. The blurry thought-cloud above its head resolves: the scribbled contents are replaced with a single clean hand-lettered label centered inside — 'meaning.' The cross-out from 'remember?' is gone. Off-white paper, loose ink. Render exactly this text on the frame, hand-lettered in the session style: "meaning".

#37 · C35

Act 3 — THE PUSH

And it can't tell the difference.

✓ narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: This scene is being regenerated for a 9:16 portrait mobile video. Compose everything for a tall vertical canvas — the widescreen original won't survive a mobile center-crop, so rebuild the scene to fit the portrait frame naturally. If the visual direction below describes a horizontal layout (side-by-side panels, left/right positioning, split-frame), restructure it as a vertical arrangement (stacked panels, top/bottom positioning, upper-half / lower-half). Keep all declared on_screen_text legible inside the portrait canvas. Existing style anchor applies unchanged. Render exactly this text on the frame, hand-lettered in the session style: "AI doesn't remember what you said." | "It remembers what you meant." | "And it can't tell the difference.".

#38 · B3

Act 4 — bumper

title card

✓ narration ✓ render

Full-screen title card in wojak-comic style, hand-lettered block type. Text centered: 'THE FIX.' Off-white background. No other elements. Held in silence ~1.5–2s.

#39 · C36

Act 4 — THE FIX

So I changed how I work with it.

✓ narration ✓ render

A full-frame Act bumper. Bold hand-lettered word 'FIX' centered, wojak-comic style. No other elements.

#40 · C37

Act 4 — THE FIX

Every time the AI generates something, I save the recipe.

✓ narration ✓ render

The AI stick figure in frame produces a generic output shape (a small illustrated square). Underneath the output, a small box slides in to catch a copy. The box is labeled 'recipe.'

#41 · C38

Act 4 — THE FIX

This includes the exact text, the exact image, and the exact settings.

! narration ✓ render gpt-image-2-text-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: A SINGLE comic panel (NOT a split panel, NOT a multi-panel comic, NOT a before-and-after, NOT a 4-panel layout). One unified scene. A hand-drawn three-item list on a paper-textured background. Three items stacked vertically with checkmark bullets: '✓ exact text', '✓ exact image', '✓ exact settings'. A 'recipe' label or folder icon at the top. Clean diagram, no characters. NO 'saved' label or file icon.

#42 · C39

Act 4 — THE FIX

Next time, no paraphrasing.

! narration ✓ render

Stick figure in frame shakes head. Two greyed-out speech bubbles appear beside their mouth — both struck through with a diagonal line.

#43 · C40

Act 4 — THE FIX

I hand the AI the saved file and say: run this. Exactly this.

✓ narration ✓ render

The saved file is handed from the stick figure to the AI figure, mid-air. A close-up of the handoff — the AI figure's hand closes around the file.

#44 · C41

Act 4 — THE FIX

What happens next?

! narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: The hooded-wojak builder holds an open glass jar with both hands, tilted slightly toward a small glowing shape floating in front of them (the AI's output — a soft-edged blob with loose ink highlights). The shape is just about to drop into the jar. A hand-lettered stamp on the jar reads 'capture.' Off-white paper, loose ink. Render exactly this text on the frame, hand-lettered in the session style: "capture".

#45 · C42

Act 4 — THE FIX

Replay it.

✓ narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: Same glass jar from the prior chunk, now opened by the hooded-wojak builder. The same glowing shape that was captured is lifting back out of the jar, hovering above the opening. A hand-lettered stamp on the jar now reads 'replay.' Off-white paper, loose ink, matches the style of the prior chunk. Render exactly this text on the frame, hand-lettered in the session style: "replay".

#46 · C43

Act 4 — THE FIX

Don't let it remember.

✓ narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: The AI figure (hooded-wojak, paler register) reaches up toward a large thought-cloud floating above (the 'memory' cloud, callback to C33). The hooded-wojak builder's hand enters from the side and blocks the AI figure's reach — palm out, stopping the motion. A bold hand-drawn diagonal strike crosses the thought-cloud. Off-white paper, loose ink.

#47 · C44

Act 4 — THE FIX

So the next time an AI says 'yeah, I did it' — ask if it checked. Because it usually doesn't know the difference.

! narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: Split frame. Left side: the AI figure (hooded-wojak, paler register) says a speech bubble with hand-lettered text 'yeah, I did it.' Right side: the hooded-wojak builder holds a checklist / clipboard with a hand-lettered question mark ('?') hovering above it. The builder faces the AI figure across the frame. Off-white paper, loose ink. Render exactly this text on the frame, hand-lettered in the session style: "yeah, I did it" | "?".

#48 · C45

Act 4 — THE FIX

I knew you were lying to me. And now you do too.

! narration ✓ render gpt-image-2-image-to-image

Flat-shaded cartoon illustration in the modern wojak-comic style. Characters drawn with exaggerated archetypal features — weary/defeated faces (hollow eyes, gray tones, slouched posture) vs. confident/heroic faces (jawline, cheekbones, steady gaze) depending on the scene. Subtle soft cel-shading, not bold inked outlines. Muted color palette with deliberate contrast between desolate and triumphant scenes. Detailed background environments — offices, dungeons, desks, monitors, paperwork, props render with texture. Dialog bubbles rendered in-frame when characters speak. Reminiscent of Nick Col / virgin-vs-chad meme-comics but painted with care. Composition is comic-strip: one clear subject, readable at small sizes. Scene: The hooded-wojak builder seated at the desk, laptop open (same framing as C2/C3 — the cold-open laptop scene). Below the laptop, two handwritten lines appear hand-lettered on the paper background: 'I knew you were lying to me.' and beneath it 'And now you do too.' Off-white paper, loose ink. Matches the cold-open's visual register with the callback text layered in. Render exactly this text on the frame, hand-lettered in the session style: "I knew you were lying to me." | "And now you do too.".

#49 · C46

Act 4 — THE FIX

Dev log number two. See you next one.

✓ narration ✓ render

An end card with 'dev log #2' at the top, channel name underneath, and a small callback illustration: the phone-shaped video from Act 2, now completed, with a hand-lettered label beneath it: 'shipped.' Held. Fade.

I knew you were lying to me. It came out in a conversation with an AI last week. And it was right. But then I realized something important. Let me backtrack a bit. A few weeks ago I shipped a video. Widescreen. Made for YouTube. But what about mobile viewers? For TikTok. Reels. Anywhere that isn't widescreen. But when you crop a widescreen video to fit these formats, it often results in heads being chopped off and text being cut in half, making it unwatchable. The fix: have the AI redraw only the broken shots. Same style, same characters, just taller. But how do you ensure the AI stays consistent? But 'same prompt in' doesn't mean what it sounds like. So, what's really going on is that the AI uses more than just the text description. Prompt plus reference in. Illustration out. Swap either one, and the output drifts. So when I say 'use the exact same inputs' — I don't just mean the words. I mean both. So I was specific. I told the AI: use the exact same inputs. Only change the shape. It came back. Eight out of nine matched. But one result was completely off. I asked why. It didn't add up. So I pushed. I said — if we used the exact same inputs, why are there errors now? Seems like we didn't. Some version of lying. But then, the AI gave us a surprising response. But I hadn't actually verified this by comparing the inputs directly. Didn't see it until I looked. The actual inputs existed. On disk. In a file. The AI could have opened it. It didn't. Here's why. When I said 'same inputs,' the AI checked its memory of what the code does. Not what the code actually sent. But the system didn't see it that way. This misinterpretation can happen in various contexts, not just with images. A config file. A spreadsheet. A script. All with the same trap. The trap is that the AI always reaches for what the code was supposed to do, not what it actually did. So before the AI tells you 'same' — ask: did it check the record, or is it just remembering the meaning? If it's just remembering, it's not actually verifying the information. Which means it focuses on the underlying meaning instead. And it can't tell the difference. So I changed how I work with it. Every time the AI generates something, I save the recipe. This includes the exact text, the exact image, and the exact settings. Next time, no paraphrasing. I hand the AI the saved file and say: run this. Exactly this. What happens next? Replay it. Don't let it remember. So the next time an AI says 'yeah, I did it' — ask if it checked. Because it usually doesn't know the difference. I knew you were lying to me. And now you do too. Dev log number two. See you next one.