Blog | Ralph Loops Are Cool

Mar 24, 2026

Comment-driven development

Spec-driven development is a real workflow with a product attached. Comment-driven development is the older, simpler version — and it works even better with agents.

Mar 20, 2026

The night shift: how to stop babysitting your AI agents

Human attention is scarce. Agent compute is cheap. The best workflows treat these as different shifts, not the same job.

Feb 21, 2026

The frozen mind: how LLM training actually works (and what it can never do)

Going deeper than next-word prediction — what embeddings, attention, and RLHF actually mean, and why understanding the training process makes you dramatically more effective with these tools.

Feb 19, 2026

Claude Code, Codex, Copilot Agent, and Cursor compared

Four autonomous coding agents, one honest look at the 2026 data — 1.7x more bugs, masked failures, infinite loops, and what actually separates the tools that work from the ones that don't.

Feb 18, 2026

Sure, let the AI run your codebase unsupervised

AI tool companies want you to believe agents can replace developers. They can't. Here's why the human-in-the-loop isn't a limitation — it's the entire point.

Feb 17, 2026

How generative AI actually works (explained like you're five)

LLMs aren't magic and they aren't thinking. Once you understand what's really happening under the hood, you stop fighting them and start getting results.

Feb 16, 2026

OpenClaw turned my Mac Mini into a 24/7 AI employee

OpenClaw (formerly Clawdbot) is a self-hosted AI agent that connects Claude, GPT, and Gemini to your messaging apps and desktop. Here's how to set it up on a Mac Mini.

Feb 15, 2026

AGENTS.md outperforms skills in agent evals

Vercel found that passive context in AGENTS.md beat active skill retrieval — hitting 100% pass rates where skills stalled at 53%. Here is why that matters.

Feb 15, 2026

Your Mac Mini is the best agent server you're not using

A dedicated Mac Mini running your AI agents keeps experiments off your daily driver, gives you persistent context, and costs less than you'd think.

Feb 14, 2026

I switched to Claude Opus 4.6 — here are 5 reasons it's my default

Opus 4.6 isn't flashy. It doesn't generate images or browse the web. But for actual productivity work, nothing else comes close.

Feb 14, 2026

Spec-driven development and why it matters

Vibe coding is fun until your agent forgets what it is building. Specs fix that. Here are the best practices for making them work.

Feb 13, 2026

Behavior-driven prompting: PRD to BDD to living spec

The most effective AI coding workflow starts with a PRD, decomposes into behaviors, breaks those into tasks, and maintains a living spec of what the system actually does.