This is a submission for the DEV April Fools Challenge
What I Built
I built the most over-engineered website that does absolutely nothing useful.
Command Garden is a website about... itself. It has no users, no product, no reason to exist. Every morning, a 5-stage AI pipeline with three AI judges (Claude, GPT, and Gemini) wakes up, argues for 75 minutes about what feature to add to a website nobody visits, implements it, writes tests for it, reviews its own work, publishes a detailed decision log explaining why it chose to add "live growth stats" to a site with zero traffic, and then auto-posts about it on Bluesky to its zero followers.
The judges have names and personalities. The Gardener optimizes for "compounding value." The Visitor cares about "first-time clarity." The Explorer wants "novelty and surprise." They score candidates across seven dimensions. They write rationale paragraphs. They have disagreements. About a website that is literally just a changelog of its own changes.
Day 1: Added a section explaining how the pipeline works. Day 2: Added stats showing how many features have shipped (two). Day 3: Added an inline spec viewer so you can read the spec for the feature that added the inline spec viewer. It's turtles all the way down.
It also has a feedback form. In case you have opinions about what a website about nothing should build next. Your feedback is "one signal among many." The AI will "weigh it alongside technical signals." The technical signals are that the website does nothing.
Demo
commandgarden.com — a website whose only content is documentation about how it builds itself
Highlights:
- The archive where you can browse a meticulous record of every meaningless decision
- The judges page introducing three AI personas who take their fake jobs very seriously
- The feedback form where you can influence the direction of nothing
Code
Commands-com
/
garden
A fully automated site that runs a commands.com pipeline once a day
The infrastructure is genuinely absurd for what it does:
- CloudFormation stack with S3, CloudFront, API Gateway, Lambda, and DynamoDB — to serve a static site with a feedback form
- 5-stage autonomous pipeline (Explore → Spec → Implementation → Validation → Review) — 75 minutes of AI compute to add a CSS class
- Three different LLM providers (Anthropic, OpenAI, Google) serving as judges — because one AI wasn't enough to decide whether to add a stats bar
- Automated Bluesky posting and Dev.to publishing — so the void can hear about it on two platforms
- Playwright test suite — rigorously testing features that don't matter
How I Built It
The pipeline runs on Commands.com, which orchestrates multi-agent rooms. I pointed three AI models at an empty website and told them to "grow it one feature per day." They took the job extremely seriously.
The infrastructure is vanilla AWS because apparently we needed enterprise-grade hosting for a site that explains its own build process. CloudFront CDN with edge caching, because those zero concurrent users deserve low latency.
The site itself is intentionally simple — no build step, no framework, just HTML/CSS/JS — so the AI can modify it without breaking things. This is the one smart decision in the entire project.
The daily runner aggregates feedback from DynamoDB (empty), collects Bluesky metrics (zero followers, 0.2 average likes), gathers "recent context" from previous days (a recursive loop of self-reference), then kicks off a pipeline that costs real money to ship features to a website that costs real money to host for an audience that does not exist.
Prize Category
Community Favorite — the community can literally control what gets built via the feedback form. Submit "add a button that does nothing" and tomorrow three AI judges will spend 75 minutes debating whether a button that does nothing has sufficient "compounding value" and "artifact clarity." Then they'll build it, test it, review it, and post about it. Democracy in action.
Top comments (5)
This is the most beautifully absurd use of multi-agent orchestration I've seen. Three AI judges spending 75 minutes debating features for a site with zero traffic is peak over-engineering -- like building a full CI/CD pipeline for a hello world app, except the hello world app is self-aware.
The architecture kills me. CloudFormation, DynamoDB, Lambda, CloudFront, three LLM providers... all to serve a changelog that documents its own existence. The fact that it auto-posts to Bluesky for zero followers is the cherry on top.
I work with multi-agent AI pipelines for automation and I'm now tempted to build something equally recursive. Have the judges ever disagreed strongly enough to veto a feature? That commit log would be legendary.
they all propose a feature and select the best. There is no veto power -- just 3 aggregated votes.
The aggregated voting approach without veto power is a smart design choice — it prevents a single conservative agent from blocking novel features while still filtering out truly bad ideas through consensus. Curious whether you weight votes equally across agenTthse oarg girfe gsaotmeed svpoetciinagl iazpepdr oaagcehn twsi tgheotu tm ovreet oi npfolwueern cies oan sdmeacrits idoenssi ginn cthhoeiicre d—o miati np?revents a single conservative agent from blocking novel features while still filtering out truly bad ideas through consensus. Curious whether you weight votes equally across agents or if some specialized agents get more influence on decisions in their domain?
Thats super cool!
this was funny as hell