The Wiki Warden: Automating the Boring Stuff with a Closet-Bound 3090

I love shipping features. I absolutely hate the soul-crushing work of maintaining documentation.

As a solo dev, the "admin debt" is real. You spend your limited PTO or late-night hours building something like The Hangout Spot, but then you're faced with the internal Wiki, the public change notes, and the README. It’s the drudgery that kills momentum.

So, I decided to kill the drudgery instead. I built the Wiki Warden; an automated documentation pipeline running entirely on my local server in my closet (the 3090 beast I recently rebuilt).

The Infrastructure: Zero Cloud, Zero Latency

I’m not sending my proprietary code to a cloud LLM. My production server has a secure VPN tunnel directly to my local AI instance. I’m running Qwen3-Coder:27b for technical reasoning and Gemma3:30b for the high-level wiki drafting.

It’s self-hosted, it’s private, and it feels like I have a junior dev living in my closet who never complains about writing docs.

The Evolution of the Warden

Building this wasn't as simple as "ask AI to write docs." It took a few phases of "accidental engineering" to get it right.

Phase 1: The Prompting War

At first, the AI would talk back to me or hallucinate instructions into the wiki pages. I had to get aggressive with my prompt engineering:

• The Header: Switched from generic "Instructions" to "Task Instructions" (which models seem to respect more).
• The Constraints: Added hard rules like "Don't talk to me" and "Only acceptable answers follow this format."

Phase 2: "Poor Man’s RAG"

When the context window started getting tight, I moved to a logic-gate approach. Instead of feeding the AI everything, the Warden now:

1. Summarizes the change notes and README.
2. Compares that summary against a flat list of every Wiki page path.
3. Requests only the specific pages it thinks are relevant for review.
4. Drafts the edits.

Phase 3: The Second Pass

I realized I had a gap: New features need new pages. Now, the Warden makes a second pass after the updates are drafted to determine if an entirely new page needs to be spawned to cover the new architecture.

The Workflow: Human-in-the-Loop

To be clear: I’m not offloading my brain to the AI. Everything the Warden drafts drops into a review queue for me.

I still QC every word. But now, instead of staring at a blank screen trying to remember what I changed three weeks ago, I’m just an editor. I spend a fraction of the time on the stuff I like the least, and 100% of the time on the projects I actually care about.

Why This is the Best Use Case for AI

People are out here trying to make AI write their whole app. I think that’s a trap. The absolute best use for local LLMs is killing the admin tasks that slow down humans. It’ll take a while for the server to "pay for itself" compared to cloud API costs, but that’s not the point. The point is that I own the pipeline, I own the data, and the Wiki Warden doesn't take vacation days.