Recent articles
Insights on modern web development, performance, and building better digital products.
Your AGENTS.md Is Your Moat Now: What Goes Inside Production Files
Next.js 16.2 made AGENTS.md default scaffolding on Tuesday. Augment's empirical study and my production template show what useful agent instructions actually contain.9 min readOpenAI Privacy Filter: The Open-Weight PII Model That Runs in Your Browser (And Where It Belongs in an Agent Stack)
OpenAI's new open-weight PII classifier runs in your browser and fills a governance layer that many agent stacks underweight. Here's how the model works, where it fits, and what it breaks.10 min readWhen Clients Ask For RAG in 2026: My Decision Tree (And Why I Rarely Start There)
I still build RAG pipelines when clients insist, but I rarely recommend them first in 2026. Most of the 2024 LLM tooling stack got obsoleted between January and April. Here is the decision tree I walk every prospect through, and what I ship when the tree points away from RAG.9 min readKimi K2.6: An Open-Weight Frontier Coding Model at One-Tenth the Cost
Moonshot's Kimi K2.6 is the second open-weight Chinese coding model to land at frontier level in four months. For agencies shipping AI agents to clients, the stack decision changed overnight.10 min readHermes Agent in Production: the Day-30 Problem
Most Hermes Agent 4-profile teams work on day one and show signs of voice convergence by day 30 in production deployments. The operator layer that prevents it: handoff contracts, memory-KPI audits, and policy gates per role.9 min readWhy Most AI Agent Setups Run Far Slower Than They Could
The architecture around the model explains the gap between modest and substantial AI productivity outcomes. Five builders published the same thesis in one week.9 min readThe AI Knowledge Layer: 127 Pages, Zero Vector Databases, and What I Got Wrong
Karpathy's LLM wiki gist hit 99,000 bookmarks in a week. It resonated because it named what every AI user feels: your agents have no memory. I run a knowledge layer in production. Here is what works, what does not, and how to build one in 20 minutes.8 min readFrom Rules to Results: What 22K Stars on a Single CLAUDE.md Reveal About AI-Assisted Development
The karpathy-skills repo proves that AI coding bottlenecks aren't about model capability. They're about the behavioral contract between human and LLM.7 min readMost Business Knowledge Bases Do Not Need RAG
I run my internal wiki on five shell commands and a hand-maintained index file, no vector database. For a 200-document knowledge base, that setup is cheaper, faster to build, and more accurate than a RAG pipeline. Here is why I skipped RAG and when you actually need it.8 min read