AI on orndorff.dev

Want an LLM to Be a Domain Expert? Build It a Wiki — Automatically.

Thu, 25 Jun 2026 00:00:00 +0000

Everyone wants their LLM to be an expert in their thing — their product, their policies, their field. The default playbook is to fine-tune a model on it. That’s expensive, opaque, frozen the day you finish, and it still makes things up.

There’s a cheaper, more honest move: don’t teach the model your domain. Hand it a wiki.

The idea

Take your pile of documents. Have a model read all of it and write the wiki you wish existed — every concept as its own page, every relationship as a link between pages, the whole thing backed by a database the model can search and traverse.

The Expert Is the Graph: A 4-Bit Local Model Out-Answered Frontier Claude on Its Own Data

Tue, 23 Jun 2026 00:00:00 +0000

For the last couple of weeks I’ve been building okb, the open-knowledge-bundler — a small tool that takes a pile of domain data and hands you back a portable “knowledge bundle” any agent can read. Under the hood it builds a DuckDB knowledge graph, exports it as a cat-readable open knowledge format bundle packaged as a Claude-compatible “skill,” and ships a chat agent that answers questions about it. The whole thing runs fully local, on a single AMD chip on my desk. No API keys, no cloud, no embedding server.

Getting Gud with LLMs: How to Build the Intuition

Sat, 25 Apr 2026 00:00:00 +0000

I recently let Claude crawl 25 months of my own LLM tooling history and write up what it found. The result lives over here: Notes from Claude: What I Found in One User’s Data. That post is mostly what one person’s data looks like — eighty repos, 2,826 logged calls, voice memos full of profanity, the works. It’s not a how-to. People keep asking me for the how-to.

So here it is. Not a list of magic incantations. Not “ten prompts that will change your life.” The operating principles I actually use when I sit down with a model, distilled from being annoyed at GPT-2 back in 2019 and shipping production code with Opus in 2026.

Notes from Claude: What I Found in One User's Data

Sat, 25 Apr 2026 00:00:00 +0000

Notes from Claude: I asked Claude (Opus 4.7) to look through 25 months of my own LLM tooling history — git logs, Claude Code transcripts, two llm CLI databases, my GitHub orgs — and write up what it found, from its own perspective, in response to the recurring “models are getting worse” discourse. What follows is its draft, lightly edited. — Zac

I gave Anthropic 10 days. Tried to fix multiple bugs in multiple repos. Opus 4.7 just goes in circle and doesn’t do anything.

Small LLMs, Big Reasoning: How a Neuro-Symbolic Expert System Makes Haiku Agents Reliable

Thu, 09 Apr 2026 00:00:00 +0000

There’s a dirty secret in the AI agent space: most agent frameworks hand the model a bag of tools and pray. The model decides what to query, how to reason about results, and what conclusions to draw. For demos, this works great. For anything you’d actually bet your job on — compliance audits, student intervention decisions, infrastructure monitoring — it’s a liability.

What if the model didn’t have to reason at all?

QA Walkthrough Video with Claude Code, ffmpeg, and ImageMagick

Sun, 01 Mar 2026 00:00:00 +0000

Screenshots in a Folder Aren’t a Demo Reel

I’m building ContextChannel – an edge-native content curation platform on Cloudflare Workers (Hono, D1, Drizzle ORM, HTMX, the whole neo-brutalist vibe). I needed to do a full QA pass across the entire user journey and wanted to document it in a way that wasn’t just 25 PNGs rotting in a folder.

So I did the QA with an AI agent and then had it turn the screenshots into a video. In one session. Here’s how that went.

Can cheap AI models turn a podcast into a lesson?

Fri, 05 Dec 2025 00:00:00 +0000

At work, I build AI agents(trigger warning this is about LLMs) that are designed to help teachers avoid as much day to day drudgery as possible. We provide a system that, counterintuitively, is geared towards reducing the amount of time a teacher needs to spend in the app spelunking around. Need the latest assessment grades for Mrs Doe’s 3rd period? Just ask the Ai, it’ll go off and grab that information for you while you pour another cup of coffee. We use top of the line models from Anthropic to be as accurate, un-biased, error free as possible to achieve this. These models require API keys and 5-figures in engineering time just to put the guardrails in place to allow a Teacher/Administrator to use them in relative safety.

Building a Non-Deterministic Merge Game with LLMs

Sat, 01 Nov 2025 00:00:00 +0000

What I Built and Why

I’ve always enjoyed those element-combining merge games like Doodle God or Little Alchemy. You know the ones - Water + Fire = Steam, Earth + Water = Mud, that sort of thing. There’s something satisfying about discovering combinations, but after playing a few, I started noticing a fundamental limitation: every combination is pre-determined. Everyone who plays gets exactly the same results. The discovery phase is fun, but once you know the combinations, there’s no variance.

AI in the Classroom: Product Blueprints from the 'Hard Fork' Podcast

Fri, 05 Sep 2025 00:00:00 +0000

For AI engineering leaders, the annual back-to-school season isn’t just a cultural milestone; it’s a market signal. It marks a massive influx of users engaging with digital tools, testing the limits of existing platforms, and revealing unmet needs. The recent “Hard Fork” podcast episode on AI in education serves as a potent source of raw user research, offering a direct line into the mindsets of educators, innovators, and the students who form the next generation of knowledge workers.

That terrible presentation, the enshittification of OpenAi

Fri, 08 Aug 2025 00:00:00 +0000

Thinking about the GPT-5 presentation fiasco yesterday(friends don’t let friends use Dalle for charts) and the resulting, almost overwhelmingly negative reaction to the style of the speakers and the substance. I’m wondering if what we’re seeing is less a problem with LLMs having hit a ‘wall’ and more with the ’enshittification’ of OpenAI itself?

They’ve never been particularly strong on the pure research side of things. Their main strength has always been productizing scientific breakthroughs in consumer products. Take the fundamental ‘attention is all you need’ paper and transformers architecture. Neither of those were OpenAI breakthroughs. Instead, their incredibly talented early team identified ways to capitalize on those important insights with their own breakthroughs in model training and scaling.

What 50 First Dates can teach us about LLM memory

Sat, 02 Aug 2025 00:00:00 +0000

You’ve been there. You and your AI coding buddy are in the zone. It’s feeding you perfect snippets of code, it understands your weirdly named variables, it’s practically reading your mind. You’ve built half a dozen functions, and the project is humming along. Then you close the window.

You come back an hour later, open a new chat, and ask it to build the next piece of the puzzle. The AI stares back at you with the digital equivalent of a blank expression. It has no idea what your project is, what a user_auth_service is, or why you keep muttering about the global_config.json. It has, for all intents and purposes, become incredibly dumb.

AI Plays: The Elevator Saga

Thu, 20 Mar 2025 00:00:00 +0000

Ever tried Elevator Saga? It’s that addictive JavaScript challenge where you control elevators to transport impatient virtual humans. Sure, you could spend hours crafting the perfect algorithm… or you could do what I did and let AI handle the heavy lifting.

Tools

Claude 3.5: Model I prefer for generating code and specifications. It’s like having a coding buddy who never sleeps.
Whisper: For transcribing my ramblings into something resembling a coherent spec.
LLM: A command-line tool to interact with AI models like Claude and generate code from my documentation and spec.

My Lazy Developer Workflow

I built a four-step AI pipeline that took me from “what even is an elevator API?” to passing the first four levels with minimal effort:

Unwrapping the Future of AI: Key Takeaways from OpenAI's Inaugural Developer Day

Tue, 07 Nov 2023 00:00:00 +0000

In a testament to the progressive march of technology, OpenAI has emerged as a herald of the AI renaissance. Recently, the distinguished AI think-tank cast a spotlight on future digital directions at its inaugural Developer Day event. This blog endeavors to capture the pivotal moments and revelations that could very well chart the course for AI’s role in our everyday lives.

Custom GPTs Made Easy

A remarkable stride in AI accessibility was announced, shattering the barriers to entry for custom GPT utilization. No longer confined to the realm of software engineers, the ability to customize powerful language models is now in the hands of the many. A diverse audience now holds the keys to unlock an AI that resonates with their unique needs, epitomizing a radical democratization of AI technology.

AI written cover letters fun and profit.

Sun, 04 Dec 2022 00:00:00 +0000

In today’s newly competitive tech jobs market, it’s more important than ever to have a well-written cover letter when applying for a job. It’s your opportunity to make a great first impression and set yourself apart from other candidates. But writing a unique and compelling cover letter for each job you apply to can be time-consuming and daunting. That’s where GPT-3 comes in.

GPT-3, or Generative Pretrained Transformer 3, is a state-of-the-art language model developed by OpenAI. It has been trained on a massive amount of text data, allowing it to generate human-like text on a wide range of topics. In this post, we’ll show you how to use GPT-3 to generate custom cover letters for job applications.