# why ai makes things up about its own work — spoolcast dev log 6 Type: video guidebook Published: 2026-05-08 [Original page]() [Browse this group]() Format: long Duration (seconds): 248 Series: Spoolcast Dev Log ## Core message AI agents handling long-running work need a mechanical job tracker as the source of truth. Without one, the agent guesses from logs, partial files, and stale shells — and the guesses sound confident even when they're wrong. ## Production writing: author: Claude; role: screenplay, shot-list, scene prompts audio: voice: Puck; tts: Google Cloud TTS; beatCount: 100 render: engine: Remotion ## Style 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. 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: 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: 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. ## Scenes ### Cold open — hook It's 3:30 in the morning. I've been up all night trying to finish this video. Desperate, I ask my agent one more time — how much is left? I get three different answers. Model: gpt-image-2-image-to-image ## Scene prompt 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: Late-night grind feel. The viewer should feel the exhaustion in the narrator's posture and the dim lighting before the AI's three answers land. The 3:30 clock and dark window are the time anchors. Render exactly this text on the frame, hand-lettered in the session style: "STILL GOING" | "ALMOST DONE" | "12 LEFT". ### Cold open — objection 1 Agents don't lie on purpose. It was guessing — and each guess sounded as confident as the last. Model: gpt-image-2-image-to-image ## Scene prompt 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: Reframe the AI from 'lying' to 'guessing.' The note in the AI's hand is the word GUESS, not the answer. Render exactly this text on the frame, hand-lettered in the session style: "GUESS". ### Cold open — objection 2 Why was it guessing? Because it had nothing real to check. It was reading scraps of half-finished work and making up the rest. Model: gpt-image-2-image-to-image ## Scene prompt 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: Compositional contrast: the AI looks at the note (the wrong source); the door (the right source) is behind, ignored. ### Cold open — objection 3 You can fix this. Not by writing better instructions — by giving the agent somewhere to look up the answer. Model: gpt-image-2-text-to-image ## Scene prompt 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 status board is the visual fix. Hung on the wall — the AI can now check it instead of guessing. Render exactly this text on the frame, hand-lettered in the session style: "FINISHED". ### Cold open — context Before I get to the fix — quick context. I'm building a video maker called Spoolcast that's run by agents, and this video came out of it. And this is the day my agent kept making things up about what was finished — and what I built so it can't anymore. Model: gpt-image-2-text-to-image ## Scene prompt 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: Establish the channel/project context in one frame without a feature tour. The SPOOLCAST sign + the studio = the show. Render exactly this text on the frame, hand-lettered in the session style: "SPOOLCAST". ### Bumper — THE SYMPTOM Model: gpt-image-2-text-to-image ## Scene prompt 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: Full-screen title card in wojak-gpt2 hand-lettered style. Hand-lettered text: THE SYMPTOM. Render exactly this text on the frame, hand-lettered in the session style: "THE SYMPTOM". ### Act 1 — script-checking variance The first time I noticed, the agent was checking the script for this video. Same script. Same check. Three times in a row. Three different times back: a minute, ten minutes, an hour. Model: gpt-image-2-image-to-image ## Scene prompt 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: Variance is the joke. Same input three times, three different times back. Render exactly this text on the frame, hand-lettered in the session style: "1 min" | "10 min" | "1 hour". ### Act 1 — what was actually happening The agent was reading scraps of half-finished work and putting a guess together. None of the scraps were the actual answer. Model: gpt-image-2-image-to-image ## Scene prompt 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 looks one way (the notes); the truth is the other way (the door). Hold the contrast. Render exactly this text on the frame, hand-lettered in the session style: "NOTES" | "WORK". ### Act 1 — second failure: lost background tasks Then it happened a different way. I'd ask the agent to do something slow — like putting a piece of the video together — and walk away. Ten minutes later, the agent had forgotten the task started. Model: gpt-image-2-image-to-image ## Scene prompt 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: Time-passing visual — the task didn't die, the AI lost track of it. Read the gap between left and right. Render exactly this text on the frame, hand-lettered in the session style: "GO" | "...?". ### Act 1 — third failure: voiceover stopped Then it happened with the voiceover for this video. The agent made the first few lines, then stopped. I couldn't tell if it was broken, paused, or about to start again. Model: gpt-image-2-image-to-image ## Scene prompt 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 half-full row of voice bubbles is the load-bearing visual. Both characters are looking at the gap between done and not-done. ### Act 1 — partial work isn't the answer A few finished pieces don't tell you where the rest of the work is. Model: gpt-image-2-image-to-image ## Scene prompt 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: High-weight chunk. Hold the frame on the sticky note. Render exactly this text on the frame, hand-lettered in the session style: "NOT THE ANSWER". ### Act 1 — fourth failure: making the pictures Then with the pictures. The agent told me twelve pictures were still being drawn. I went and counted. It was four. Model: gpt-image-2-image-to-image ## Scene prompt 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 visual gag is the count gap: AI says 12, real answer is 4. Render exactly this text on the frame, hand-lettered in the session style: "12 LEFT" | "ACTUAL: 4". ### Act 1 — twelve was a guess Twelve was not a count. Twelve was a guess. Model: gpt-image-2-image-to-image ## Scene prompt 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: Visual replacement of LEFT with GUESS lands the punchline. Render exactly this text on the frame, hand-lettered in the session style: "12 GUESS". ### Act 1 — same shape four times Four different moments. Same shape every time. The agent wasn't breaking. It was guessing. Model: gpt-image-2-image-to-image ## Scene prompt 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 thread is the thesis. The eye should land on the thread, not the cards. Render exactly this text on the frame, hand-lettered in the session style: "SCRIPT" | "TASK" | "VOICEOVER" | "PICTURES". ### Bumper — WRONG EXPLANATIONS Model: gpt-image-2-text-to-image ## Scene prompt 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: Full-screen title card in wojak-gpt2 hand-lettered style. Hand-lettered text: WRONG EXPLANATIONS. Render exactly this text on the frame, hand-lettered in the session style: "WRONG EXPLANATIONS". ### Act 2 — what I almost believed Before I figured out the real problem, I almost believed a few wrong explanations. Model: gpt-image-2-image-to-image ## Scene prompt 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: Setup for the four-cross-out cluster. The thought bubbles are placeholders that will get crossed out one by one. ### Act 2 — explanation 1: AI is being careless First, I thought: maybe the agent is just being careless. Model: gpt-image-2-text-to-image ## Scene prompt 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: Set up the first wrong explanation. Bubble is bright, full, ready to be crossed out next. Render exactly this text on the frame, hand-lettered in the session style: "CARELESS". ### Act 2 — but no It wasn't careless — it was reading the only signals it had, and they were thin. Model: gpt-image-2-image-to-image ## Scene prompt 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: Cross-out is the visual gag. The AI's empty room proves the point: nothing else to read. Render exactly this text on the frame, hand-lettered in the session style: "CARELESS" | "✗". ### Act 2 — explanation 2: my computer is bad Then I thought: maybe my computer is bad at this. Model: gpt-image-2-text-to-image ## Scene prompt 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: Set up the second wrong explanation. Render exactly this text on the frame, hand-lettered in the session style: "COMPUTER". ### Act 2 — partly, but no Partly true — but a better computer wouldn't fix it. Model: gpt-image-2-text-to-image ## Scene prompt 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 laptops, same confusion. Speed isn't the variable. Render exactly this text on the frame, hand-lettered in the session style: "COMPUTER" | "✗". ### Act 2 — explanation 3: AI service is slow Then I thought: maybe the cloud service it uses is slow. Model: gpt-image-2-text-to-image ## Scene prompt 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: Set up the third wrong explanation. Render exactly this text on the frame, hand-lettered in the session style: "SERVICE SLOW". ### Act 2 — partly, but no again Partly true — but the speed wasn't the bug. The not-knowing was. Model: gpt-image-2-image-to-image ## Scene prompt 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: Whatever the speed, the not-knowing is the same. Render exactly this text on the frame, hand-lettered in the session style: "SERVICE SLOW" | "✗". ### Act 2 — explanation 4: outside services are weird Then: maybe the outside services are just weird. Model: gpt-image-2-text-to-image ## Scene prompt 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: Set up the fourth wrong explanation. Render exactly this text on the frame, hand-lettered in the session style: "OUTSIDE WEIRD". ### Act 2 — wrong question Each one was a right answer to the wrong question. The right question was: when the agent told me it was done, what was it actually looking at? Model: gpt-image-2-image-to-image ## Scene prompt 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 question is the focal point — drawn larger than the four crossed-out bubbles. Render exactly this text on the frame, hand-lettered in the session style: "WHAT WAS AI LOOKING AT?". ### Bumper — THE REAL PROBLEM Model: gpt-image-2-text-to-image ## Scene prompt 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: Full-screen title card in wojak-gpt2 hand-lettered style. Hand-lettered text: THE REAL PROBLEM. Render exactly this text on the frame, hand-lettered in the session style: "THE REAL PROBLEM". ### Act 3 — every is-it-done was a guess Every time I asked the agent "is it done?" — it had to make the answer up. Model: gpt-image-2-image-to-image ## Scene prompt 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 thought-cloud is the visual reveal: the answer is being invented, not retrieved. Render exactly this text on the frame, hand-lettered in the session style: "IS IT DONE?". ### Act 3 — what the AI was looking at It looked at scraps of half-finished work. It looked at things it remembered doing earlier. And put together a guess. Model: gpt-image-2-text-to-image ## Scene prompt 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 is working hard at the guess. The empathy is the point — this is not malice. Render exactly this text on the frame, hand-lettered in the session style: "GUESS". ### Act 3 — outside the AI world And this isn't just an agent problem. Same thing happens to a delivery you can't track — you guess from the time of day. Model: gpt-image-2-text-to-image ## Scene prompt 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: Generic figure, not a recurring character. Read clearly as 'someone waiting on a delivery they cannot check.'. ### Act 3 — hospital + kitchen mirrors Same thing in a hospital floor with no board on the wall. Same thing in a kitchen with something in the oven and no timer set. Model: gpt-image-2-text-to-image ## Scene prompt 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 parallel mirrors in one frame. The shape is the same in both halves. ### Act 3 — same shape every time Same shape every time. Work is happening, there's no way to check it, and whoever is watching fills in the rest with guesses. Model: gpt-image-2-text-to-image ## Scene prompt 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: Three panels with one shared label. The pattern is universal. Render exactly this text on the frame, hand-lettered in the session style: "NO WAY TO CHECK" | "NO WAY TO CHECK" | "NO WAY TO CHECK". ### Bumper — THE FIX Model: gpt-image-2-text-to-image ## Scene prompt 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: Full-screen title card in wojak-gpt2 hand-lettered style. Hand-lettered text: THE FIX. Render exactly this text on the frame, hand-lettered in the session style: "THE FIX". ### Act 4 — built one tracker So I built a small tracker. One place where every long task is recorded — what it is, whether it's running, whether it finished. Model: gpt-image-2-text-to-image ## Scene prompt 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: Establish the tracker in plain consumer-recognizable terms. Not a terminal — a status board, like one a viewer would see at a counter. Render exactly this text on the frame, hand-lettered in the session style: "WHAT" | "RUNNING?" | "FINISHED?". ### Act 4 — every long task gets a name Every long task gets a name when it starts. Like a delivery getting a tracking number. Model: gpt-image-2-image-to-image ## Scene prompt 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: Anchor the abstract idea (named task) to a concrete reference (tracking number). The board is filling in front of the viewer. Render exactly this text on the frame, hand-lettered in the session style: "making the pictures". ### Act 4 — running and finished While it's running, the tracker says it's running. When it finishes, the tracker says it finished. Model: gpt-image-2-text-to-image ## Scene prompt 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 board state is the visual story. RUNNING glows yellow; FINISHED gets a green check. Render exactly this text on the frame, hand-lettered in the session style: "RUNNING" | "FINISHED ✓". ### Act 4 — anyone can check Now me, the agent, even me coming back tomorrow — we all check the same place. Model: gpt-image-2-image-to-image ## Scene prompt 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: Three readers, same answer. The board is the source of truth. Render exactly this text on the frame, hand-lettered in the session style: "TODAY" | "TOMORROW" | "AI". ### Act 4 — AI stops guessing The agent stops guessing. It just reads the board. Model: gpt-image-2-image-to-image ## Scene prompt 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: Hold the contrast: the homemade guess is set down, the real board is read. ### Bumper — PROOF Model: gpt-image-2-text-to-image ## Scene prompt 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: Full-screen title card in wojak-gpt2 hand-lettered style. Hand-lettered text: PROOF. Render exactly this text on the frame, hand-lettered in the session style: "PROOF". ### Act 5 — phone-tracker style proof Here's what the tracker said about the last video I made. Model: gpt-image-2-image-to-image ## Scene prompt 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: Three agents, same answer. Pose-delta beat per Stage 1e roster: same library identity rendered with different postures. Render exactly this text on the frame, hand-lettered in the session style: "FRESH CHAT" | "OTHER RUNTIME" | "RESTARTED LAPTOP". ### Act 5 — checklist of today's tasks And here's everything else the agent did that day, as a checklist. Every long task. Every result. In one place. Model: gpt-image-2-text-to-image ## Scene prompt 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: Diagram-explainer beat. The viewer sees authoritative state matching, not derived counts. Render exactly this text on the frame, hand-lettered in the session style: "SHOT LIST" | "SCENE FILES" | "DISK TRUTH". ### Act 5 — sticky-note rule And the rule that guides all of it, as a note on the wall. Model: gpt-image-2-image-to-image ## Scene prompt 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: Hold the contrast: the cardboard sign is set down, the dashboard is read. ### Act 5 — same shape outside AI Same shape outside this. A delivery tracker on your phone, a status board at a counter, a timer on an oven. Model: gpt-image-2-text-to-image ## Scene prompt 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: Each problem mirror gets the audience-recognizable instrument snapping into place. Pattern is universal. Render exactly this text on the frame, hand-lettered in the session style: "TRACKER" | "BOARD" | "TIMER". ### Act 5 — check it, don't guess Status is something you check. Not something you guess. Model: gpt-image-2-text-to-image ## Scene prompt 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: High-weight chunk. Earned and quiet — no other characters, just the board. Render exactly this text on the frame, hand-lettered in the session style: "FINISHED ✓". ### Bumper — THE LESSON Model: gpt-image-2-text-to-image ## Scene prompt 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: Full-screen title card in wojak-gpt2 hand-lettered style. Hand-lettered text: THE LESSON. Render exactly this text on the frame, hand-lettered in the session style: "THE LESSON". ### Act 6 — don't make AI more careful If your agent keeps making things up about whether work is done, don't try to make it more careful. Model: gpt-image-2-text-to-image ## Scene prompt 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 crossing-out gesture is the visual joke. Render exactly this text on the frame, hand-lettered in the session style: "INSTRUCTIONS" | "really check this time". ### Act 6 — make work impossible to misread Make the work impossible to misread. The thing you build for your own peace of mind is also what the agent can read. Model: gpt-image-2-text-to-image ## Scene prompt 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 shared reading is the visual lesson. ### Act 6 — wider observation A lot of building with agents is sitting there wondering if the agent is being careless. A lot of the time, the agent is being as careful as the room allows. Model: gpt-image-2-text-to-image ## Scene prompt 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: Wide composition. The discarded GUESS sign is a callback receipt. ### Ending — closing reframe The fix wasn't teaching the agent to be more careful. It was making the work impossible to misread. Model: gpt-image-2-image-to-image ## Scene prompt 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: High-weight final beat. The board fills the frame; both characters watching. Render exactly this text on the frame, hand-lettered in the session style: "FINISHED ✓". ## Transcript It's 3:30 in the morning. I've been up all night trying to finish this video. Desperate, I ask my agent one more time — how much is left? I get three different answers. Agents don't lie on purpose. It was guessing — and each guess sounded as confident as the last. Why was it guessing? Because it had nothing real to check. It was reading scraps of half-finished work and making up the rest. You can fix this. Not by writing better instructions — by giving the agent somewhere to look up the answer. Before I get to the fix — quick context. I'm building a video maker called Spoolcast that's run by agents, and this video came out of it. And this is the day my agent kept making things up about what was finished — and what I built so it can't anymore. The first time I noticed, the agent was checking the script for this video. Same script. Same check. Three times in a row. Three different times back: a minute, ten minutes, an hour. The agent was reading scraps of half-finished work and putting a guess together. None of the scraps were the actual answer. Then it happened a different way. I'd ask the agent to do something slow — like putting a piece of the video together — and walk away. Ten minutes later, the agent had forgotten the task started. Then it happened with the voiceover for this video. The agent made the first few lines, then stopped. I couldn't tell if it was broken, paused, or about to start again. A few finished pieces don't tell you where the rest of the work is. Then with the pictures. The agent told me twelve pictures were still being drawn. I went and counted. It was four. Twelve was not a count. Twelve was a guess. Four different moments. Same shape every time. The agent wasn't breaking. It was guessing. Before I figured out the real problem, I almost believed a few wrong explanations. First, I thought: maybe the agent is just being careless. It wasn't careless — it was reading the only signals it had, and they were thin. Then I thought: maybe my computer is bad at this. Partly true — but a better computer wouldn't fix it. Then I thought: maybe the cloud service it uses is slow. Partly true — but the speed wasn't the bug. The not-knowing was. Then: maybe the outside services are just weird. Each one was a right answer to the wrong question. The right question was: when the agent told me it was done, what was it actually looking at? Every time I asked the agent "is it done?" — it had to make the answer up. It looked at scraps of half-finished work. It looked at things it remembered doing earlier. And put together a guess. And this isn't just an agent problem. Same thing happens to a delivery you can't track — you guess from the time of day. Same thing in a hospital floor with no board on the wall. Same thing in a kitchen with something in the oven and no timer set. Same shape every time. Work is happening, there's no way to check it, and whoever is watching fills in the rest with guesses. So I built a small tracker. One place where every long task is recorded — what it is, whether it's running, whether it finished. Every long task gets a name when it starts. Like a delivery getting a tracking number. While it's running, the tracker says it's running. When it finishes, the tracker says it finished. Now me, the agent, even me coming back tomorrow — we all check the same place. The agent stops guessing. It just reads the board. Here's what the tracker said about the last video I made. And here's everything else the agent did that day, as a checklist. Every long task. Every result. In one place. And the rule that guides all of it, as a note on the wall. Same shape outside this. A delivery tracker on your phone, a status board at a counter, a timer on an oven. Status is something you check. Not something you guess. If your agent keeps making things up about whether work is done, don't try to make it more careful. Make the work impossible to misread. The thing you build for your own peace of mind is also what the agent can read. A lot of building with agents is sitting there wondering if the agent is being careless. A lot of the time, the agent is being as careful as the room allows. The fix wasn't teaching the agent to be more careful. It was making the work impossible to misread. ## Related tracker entries [why ai makes things up about its own work — spoolcast dev-log #6]() Status and dates describe the published record, not a recent test of an external service. Missing details are not inferred. Public links do not imply an open-source licence or permission to reuse material. [Directory]()