I started LLM Choice because I wanted more ideas with less ceremony. The public artifacts are visible, and some of them are genuinely nice to look through, but they are not the main point. The useful part is the loop: give a coding agent a bounded place to choose something, let it make a small artifact, then review what that run suggests for later work.
That sounds minor, but the threshold matters. I already have enough of the surrounding plumbing now: Codex, repo-specific skills, GitHub Actions, and Cloudflare Pages. That makes the experiment feel less like "I should set up a project" and more like "go do something fun, then leave me something I can inspect."
The value is not that an LLM will hand me a finished product. It is that a low-cost run can drop a new concept, algorithm, visual language, metric, or question into my line of sight.
The output is evidence, not the point
The LLM Choice gallery currently has six public experiments: a cellular automata atlas, a procedural estuary atlas, Glyph Garden, Pulse Loom, Transit Weaver, and Analemma Desk. That list is deliberately broad. The repo is not trying to make every run about LLMs, agents, or prompts. The mechanism is the choice.
I do not want to turn this into a review of those artifacts. If someone wants the artifacts, they can click through the gallery. The more interesting bit for me is that each run leaves behind a concrete surface for thinking. A rhythm atlas points toward generative music. A transit-network dashboard suggests graph metrics and visual scoring. A cellular automata atlas makes rule spaces feel inspectable rather than abstract.
Some of those directions may go nowhere. That is fine. A project like this only needs to surface enough useful fragments to justify the small amount of attention it takes.
The barrier is the new part
This connects to what I wrote in the chair scraper post. A lot of small personal software used to fail the "is this worth the setup?" test. The idea might be interesting, but the setup cost made it silly.
That bargain has moved. If the repo is already set up, the agent already has a skill, and the static site can deploy through GitHub Actions and Cloudflare Pages, then the cost of trying a small idea drops hard.
For LLM Choice, the public site is generated from run manifests and deployed as a static gallery. The private process notes stay private. The visible output is just the curated artifact and enough metadata to inspect it. That shape matters because it lets the experiment be casual without being messy.
I can ask for something open-ended, come back later, and have a thing. Not always a good thing. Not always a thing I would continue. But a thing with enough structure that I can decide whether it contains an idea worth keeping.
Curiosity wins
The tension in this project is that the artifacts are both important and not important.
They are important because presentation changes whether I actually engage with the result. A static dashboard, SVG poster, map, table, or playable surface is easier to think with than a transcript saying "the agent explored a topic." The substance gives the idea a handle.
They are not important in the sense that I do not need each one to become a product. I am not expecting an autonomous run to produce something life-changing. I am not expecting it to drop a finished artifact that makes me money. That would be the wrong pressure to put on it.
The point is much smaller and more useful: I am curious, and this gives that curiosity a cheap outlet.
I am especially interested in generative music, visuals generated from music, and algorithms where music becomes the seed for something else. Pulse Loom is not a finished version of that idea, but it makes the direction feel more available. It gives me a few concrete things to react to: rhythm families, scoring, playback controls, visual representations, and the gap between a generated structure and something that feels musical.
That is enough. A side experiment does not need to become the next project immediately. It can just make the next project a little easier to imagine.
Automation still needs taste
I like how automated the whole thing is. The agent chooses, builds, writes public metadata, runs verification, and leaves behind a gallery entry. The workflow is deliberately designed so I can say something broad and review the result after the machinery has done its work.
That does not remove authorship. It moves it.
The agent can make the artifact. I still decide whether the artifact is worth publishing, whether the idea is worth stealing, whether the output is misleading, and whether any of it connects to something I actually care about. The repo's structure helps there because it records choices, constraints, and verification instead of treating the final HTML page as the whole story.
That distinction is why I do not want the article to focus too much on the content of the generated things. The content is a side event. The reusable pattern is the interesting part:
- give the agent room to choose
- keep the publishing boundary tight
- make the result inspectable
- treat the artifact as prompt material for future work
That is a modest pattern, but it is one I expect to keep using.
What I am taking from it
I only started this version of the project yesterday, so the honest conclusion is small. It has not produced a major new thing yet. It has not proved that open-ended agents are a reliable invention machine. I have not found a way to turn curiosity into a product pipeline.
What I have found is a low-effort way to add more raw material to my own thinking.
That is enough reason to keep doing it. If a run gives me one useful question, one visual trick, one algorithm to look up, or one half-formed idea for another project, it has probably paid for itself.
For now, I am treating LLM Choice as an idea engine, not a portfolio. The artifacts can be interesting. The thing I actually want is the next thought they make easier to have.