Most Tech Writing Explains What. This Is About Why.

I write about what actually happens inside engineering orgs. Not the polished version.

March 2026Engineering14 min read
Cut Claude Code Token Usage by 60–90%: 16 Proven Techniques

Cut Claude Code Token Usage by 60–90%: 16 Proven Techniques

Battle-tested techniques for .claudeignore, context compression, and multi-agent task splitting that cut token usage by 60–90% without losing Sonnet's predictive accuracy.

Read article →
July 2026Engineering8 min read
90% of What Your AI Agent Scrapes Is Wasted Tokens

90% of What Your AI Agent Scrapes Is Wasted Tokens

A typical web page runs 80,000 tokens once a model reads the raw HTML. Over 90% of that is CSS, JavaScript, and markup an agent will never quote. Here's how to stop paying for it.

Read article →
July 2026Engineering10 min read
BeautifulSoup Scrapes Pages. Crawl4AI Assumes No One Reads Them.

BeautifulSoup Scrapes Pages. Crawl4AI Assumes No One Reads Them.

Crawl4AI hit 75,000 GitHub stars in three years, faster than Scrapy did in eighteen. The reason isn't a better scraper. It's that access and legibility are different problems.

Read article →
July 2026Engineering9 min read
Prefect Bought Dagster. Airflow Still Has More Stars Than Both Combined.

Prefect Bought Dagster. Airflow Still Has More Stars Than Both Combined.

Prefect acquired Dagster this week. Combined GitHub stars and PyPI downloads for both still trail Apache Airflow alone. Here's what the deal actually changes.

Read article →
June 2026Engineering10 min read
Why AI-Generated Plans Need Visual Review, Not More Text

Why AI-Generated Plans Need Visual Review, Not More Text

AI solved the generation problem and created the comprehension problem. The bottleneck is no longer output volume. It is review surface area.

Read article →
June 2026Engineering17 min read
On-Device AI Comes of Age, but Hardware Still Sets the Limits

On-Device AI Comes of Age, but Hardware Still Sets the Limits

Compact models run offline on phones, laptops, and edge devices. The bottleneck is no longer model capability. It is RAM, memory bandwidth, and thermal headroom under sustained load.

Read article →
June 2026Engineering11 min read
MD vs MDX: The Decision Most AI-Agent Repos Get Wrong

MD vs MDX: The Decision Most AI-Agent Repos Get Wrong

Markdown is a storage format. MDX is an execution format. Every AI coding tool chose .md for agent instructions. The distinction explains why, and when MDX earns its place.

Read article →
June 2026Engineering18 min read
Claude Code for Designers: Why Most Still Aren't Using AI (and the Repo That Changes That)

Claude Code for Designers: Why Most Still Aren't Using AI (and the Repo That Changes That)

74% of AI coding tool users report meaningful productivity gains. Most designers are not in that number. The blocker is not skills or fear. It is project structure. A complete designer-ready Claude Code repo with DESIGN.md, design tokens, skills, agents, and a weekend setup guide.

Read article →

Showing 8 of 54 posts

Browse by Topic

Explore all topics to discover insights tailored to your interests.

View All Tags

Contact

Let's Talk