Last Week

Last week, I was trying to stop the Hermes setup from becoming its own moving target. The next priority was supposed to be Shokken marketing. This week, that brought me to a different problem: how do I produce enough useful videos without personally polishing every step?

I want a video assembly line. I do not have one yet.

Shokken is available on Android and iOS in the United States. Building the product is only part of the job, though. I also need to reach the people I built it for. AI-assisted coding has made creating software more accessible, but it has not removed the work of helping an audience discover it.

Three Videos, Three Different Jobs

I am thinking about three kinds of videos, and treating them all as the same production problem would be a mistake.

These Saturday updates are still mostly manual. Their original purpose was to keep me accountable: talk through what I tried, what happened, and what I want to do next. A fairly direct conversation with the camera suits that purpose. I want to keep the genuine version of me in this series.

The Thursday Shokken tutorials have a different job. They need to show a customer how the product works, and the presentation is part of the signal that somebody cares about its quality.

The latest tutorial pilot was a significant improvement over my first manually produced videos. I am pleased with how much more polished it looked. I am less pleased with the amount of back-and-forth needed to get there. A good finished video is progress, but it is not proof that I have a repeatable assembly line.

The third format would explain the systems and ideas I have been learning about. A Saturday update can mention agentic memory; a useful educational video needs to help someone understand what it is, why they might need it, and how the pieces fit together. My face alone is not much of a visual explanation.

Moving the Workflow, Not Just the Model

I made the tutorial pilot through Codex and am trying to transfer the knowledge and conventions from that work into Hermes.

Both are agent harnesses: the surrounding software that lets a model work with tools and instructions. Codex is proprietary, while Hermes is open source. I want less dependence on any single proprietary setup and more freedom to change the model behind it.

Using the same model does not make two harnesses behave the same way. The migration has not been seamless. At the time of this recording, I had not even reached a usable production result for the next Thursday tutorial in Hermes.

I still think the direction is right. I want to explain the way I work once and have the agent carry those conventions forward. Right now, there is still a lot of hand-holding between that goal and the result.

From a Voice Memo to Something Worth Watching

For the educational-video experiment, I recorded a voice memo while driving to get groceries. I talked through OpenViking, the agentic-memory subsystem I have been using: how I found it, why I wanted it, the problems I encountered, and the compromises that come with it.

The recording was a little over twenty minutes. I gave the transcript to an agent and asked it to turn that rambling explanation into a script a viewer would want to follow.

Even before adding animation, the script exposed a gap. It assumed the audience already knew things that the video was meant to explain. For example, it did not clearly distinguish the AI application from the language model behind it. It also treated components such as the planner, embedder, reranker, and vision-language model as familiar concepts.

The script made sense to me because I already knew the subject. That is not the same as making sense to the person I am trying to teach.

AI can help bridge my lack of proficiency with editing software. It has not yet given me a reliable path from twenty minutes of talking into an engaging finished video.

The Tools Are There; the Workflow Is Not

Programmatic animation is promising. Tools such as Remotion, and projects that bring production tools together such as OpenMontage, let an agent describe and build visuals in code instead of requiring me to become proficient in a motion-graphics editor first.

The tradeoff is that those tools are scattered and move quickly. Learning a stable editing application gives you a skill you can reuse. With this approach, I am also chasing changes in the tools and trying to make the pieces work together.

That is why I do not find the claim that AI makes everything effortless very convincing. Poor, low-effort AI videos exist. Carefully produced AI-assisted content can also be useful and difficult to distinguish at first glance. I do not know how much of another channel’s workflow is automated just from watching its output.

What matters to me is whether the content teaches me something or gives me an experience worth watching. The production method alone does not settle that question.

Next Week

My software-development workflow is much further along than the video workflow. There is still manual judgment, but I have a clearer pipeline from an idea to implementation and deployment. With video, I am still working out the middle.

Eventually, I also need to tackle short-form discovery. I personally prefer long-form videos, but my preference does not remove the need to reach people who discover products through shorter content.

For now, the next concrete goal is to move the Thursday product-video workflow into Hermes successfully. I hope to have a more substantive update next week. That is a hope, not a claim that the migration is finished.

The assembly line is the goal. This week is an account of the work still needed to build it.

If you are interested in the restaurant waitlist tool itself, visit Shokken. It is available on Google Play and the iOS App Store.