I used AI to moderate AI content — here's what I learned building AIHallucination

#webdev #ai #vibecoding #showdev

Most of my team got laid off because "AI can do their jobs now."

I'm probably the last one standing. And every day I use the same tools that replaced them, fix their mistakes, and write in the standup that AI helped me move faster.

Nobody was being honest about this. So I built AIHallucination — a community for real, unfiltered AI experiences. The fails, the wins, the absurd outputs, the expectation vs reality moments we all quietly live through.

This is the technical story of how I built it.

The Stack

TanStack Start — React 19 with SSR and file-based routing. First time using it in production. The mental model takes getting used to but the DX is excellent once it clicks.
PostgreSQL + Drizzle ORM — type-safe queries, no magic, exactly what I wanted
Better Auth — email/password + Google OAuth, username plugin. Surprisingly painless to set up.
Railway — deployment, Postgres hosting
Tailwind v4 + shadcn — dark editorial design, electric lime accent

The Interesting Part — Using AI to Moderate AI Content

Every post on AIHallucination goes through GPT-4o mini before going live. Yes, I used AI to moderate a community about AI experiences. The irony is not lost on me.

Here's why it made sense:

Moderation needs to be fast and async — users shouldn't wait
The rules are nuanced — reject spam but allow angry, frustrated, sarcastic posts that are genuine
GPT-4o mini is cheap enough that even at scale it costs pennies

The system prompt took the most iteration by far. The hard part wasn't rejecting obvious spam — it was getting the model to understand platform tone.

AIHallucination is blunt and honest. Strong criticism of AI is welcome. Frustration is welcome. Sarcasm is welcome. But pure harassment or content with no real AI experience behind it gets rejected.

Early versions over-moderated. Sarcastic posts got flagged as "low quality." Rants got flagged as "purely negative." I had to explicitly tell the model:

"Sarcasm, frustration, and informal language are part of the platform voice — don't reject for tone alone"

The moderation response includes a reason and a suggestion — so if your post gets rejected, you know exactly why and how to fix it.

Rejection with no explanation is frustrating. Rejection with actionable feedback is useful.

The Flow

User submits post
Post saved to DB with pending status — not visible on public feed
moderatePost() fires async (fire and forget)
GPT-4o mini evaluates against 12 platform rules
Post status updates to approved, rejected, or failed
User can check status in My Posts, see rejection reason, edit and resubmit

Cost: GPT-4o mini at this scale costs almost nothing.

The Data Model

Posts have a moderationStatus column for fast filtering (pending, approved, rejected, failed).

Every moderation check writes to a separate moderation_event table as an audit log — so you can see the full history of checks on a post, not just the latest status.

moderation_event (
  id, post_id, user_id, status, reason, suggestion, created_at
)

This separation matters.

The status column is for performance (indexed, used in every feed query)
The event table is for transparency (show the user their rejection history, allow appeals)

What I'd Do Differently

Seeding content is harder than building the platform. A community with 5 posts looks dead. Getting real people to post first is the actual hard problem — the tech is the easy part.
Moderation prompt engineering is underrated. I spent more time on the system prompt than on any single feature. Small wording changes had big effects on what got approved vs rejected.