Knowledge archive / 12

Applied AI notes for systems that have to work.

Long-form field notes on controlled AI workflows, agent architecture, evaluation, tooling, and the operating systems around them.

First-party knowledge hub

Articles and practical guides

01 / On this site

The Deterministic Layer: Where AI Workflows Need Certainty

Why models should interpret ambiguous inputs while deterministic workflow controls govern validation, state, permissions, approvals, recovery, and consequential actions.

02 / On this site

Inside my AI agent setup: LoopEngineering, Harness

A look inside a role-based AI agent environment connecting Discord, Obsidian, Hermes, OpenCode/Superset, persistent memory, and human review.

03 / Medium article

Why AI Coding Results Depend 80% on the Harness

A field report on how routing, memory, permissions, review loops, verification, and smaller agent responsibilities shape AI coding results.

04 / Medium article

Why AI Agent Harnesses Matter More Than Model Choice

A practical argument for focusing on harness design, feedback loops, tools, and evaluation instead of only swapping models.

05 / Medium article

How I Cut Token Usage in Half by Redesigning My Agent Architecture

A breakdown of how architecture choices, context management, and cleaner agent boundaries reduced token consumption.

06 / Medium article

How I Built a Second Brain for Claude Code

How I designed a knowledge workflow to make Claude Code more useful with persistent project memory and reusable context.

07 / Medium article

AutoResearch Beyond ML: How I Adapted Karpathy’s Loop to Trade Prediction Markets

An experiment applying iterative research loops beyond classic ML to prediction-market research and trading workflows.

08 / Medium article

The Impostor Loop: Why Vibecoding Made Me Feel Like a Fraud

A personal reflection on AI-assisted coding, confidence, skill development, and the tension between speed and ownership.

09 / Medium article

How I Keep Claude Code Fast With Zero Plugins

A lightweight workflow for keeping Claude Code responsive using repository structure, focused context, and simple operating rules.

10 / Medium article

How I use OpenClaw to ship MVPs

A practical article on using AI agents, a Python script, and a low-cost VPS to move faster from idea to working product.

11 / Medium article

How to Use Kimi K2.5 Free on OpenClaw

A hands-on guide to connecting Kimi K2.5 with OpenClaw and experimenting with low-cost AI agent workflows.

12 / Medium article

How I Turned Obsidian Into an AI-Powered Second Brain With Claude Code

A workflow for combining Obsidian and Claude Code into a practical AI-assisted second brain for notes and project memory.