# TAKO and AI agent production conversation These excerpts show the real exchanges that changed the acceptance criteria for the Aha videos. User statements, technical findings, and AI interpretations remain separate, and no reason is invented when the user did not provide one. ## 1. The project began with one sentence ### TAKO > I want to make Aha videos featuring Japanese landmarks, cute cats, and cute dogs. ### AI agent The agent used this request to propose subjects and changes, create images, make 22-second challenges and 26-second answer versions, inspect intermediate states, and perform machine checks. The creator was not asked to prepare before-and-after images or internal production data at the start. ## 2. One rejected dog video did not ban every dog subject ### TAKO > That dog one does not work. ### AI agent The dog video in which a ball changed into an apple was recorded as rejected. The comment was not generalized into “all dog subjects are prohibited.” Later dog videos were judged separately for naturalness, change design, and interest. ## 3. Accepted examples supplied criteria for later trials ### TAKO > They are all OK. ### AI agent The torii gate, hunter cat, and pug videos were recorded as accepted. Keeping the main subject stable, limiting the work to one surrounding change, and making the answer understandable afterward became checks for later candidates. Three accepted videos were not treated as proof of a universal success formula. The same ideas still had to be tested in later batches. ## 4. A technically valid video could be far too easy ### TAKO > Four of them do not work. The last two are especially obvious—you would have to try not to notice them. ### AI agent Four of the six videos were recorded as rejected and two as accepted. The last two were technically valid but failed as challenges because their changes were too obvious. Large object movement and replacement were added to the list of conditions to avoid in the next proposals. ## 5. “Usable but duplicated” remained separate from technical failure ### TAKO > It is close, but the idea is too similar to an earlier one, so exclude it. ### AI agent When the technical quality was usable but the change duplicated previous work, the result was stored as “excluded for idea duplication,” not NG. Production complete, technically valid, human accepted, and duplicate exclusion remained separate states. ## 6. Hard to notice did not automatically mean good ### TAKO > Reject the jellyfish and contrail. Their changes are too small. ### AI agent Those two videos were rejected from the ten works in Batch 9, while the other eight were accepted. A challenge should resist immediate detection, but the difference must still become understandable after the answer. A lower limit for change size was added to later trial criteria. ## 7. The approved special rule omitted the storyboard ### TAKO Wife, confirming the process > You did not make a storyboard this time? ### TAKO > Correct. This time, I asked the AI agent to make direct video trials without a storyboard. ### Production note The standard workflow approves a storyboard before video production. This episode used a special rule because it needed many short trials sharing one specification. It demonstrates adapting the workflow to the work instead of following one fixed template unconditionally. ## 8. What repeated evaluation achieved The finished making-of summarizes the change as follows: - Initially, only two or three videos out of ten were accepted. - After roughly ten rounds, eight out of ten were fully accepted. - The remaining two were usable with minor adjustments. The detailed production record contains 57 evaluated works: 32 accepted, 22 rejected, and three excluded because their ideas duplicated previous works. The batches used different subjects and conditions, so these figures are not presented as a controlled performance comparison. They document how recorded human feedback influenced later production criteria. ## What this conversation taught us An AI agent can produce many trials quickly while continuing technical checks and evaluation records. It must still preserve the user’s actual wording, avoid overgeneralizing one comment, and carry only confirmed reasons into later work. A person makes the final judgment about whether a video is natural, challenging, and interesting.