Posts

Giving Claude a Brush: What Tools Actually Are

By the end, Claude repaints a Monet, signature and all. The whole trick is tools, granted one at a time: write tools let a model act, read tools let it see. Built from scratch on a drawing canvas, every session recorded and replayed.

Stop Being the One Who Clicks Around

After every change I was the one opening the preview, clicking the flow, and finding the bug the agent could not see. Now the builder shells out a second agent that drives the browser, files a report, and gets a fix back. Here is the skill that does it.

Your Coding Agents Should Come From Different Labs

My daily driver is the most expensive model on the market, and I still route half my work to a competitor. Why mixing models from different labs beats loyalty to one, and the skills that make it effortless.

Find and Kill the Agent That's Eating All Your RAM

Coding agents balloon to 10 GB when an MCP leaks or a test runner leaves workers behind. I built a monitor that names the session, blames the subprocess, and hands you a kill command that doesn't lose your work.

Use Worktrees. Your Agents Already Understand Them.

Running multiple coding agents in one checkout means collisions, red test suites, and burned tokens. Worktrees give every agent its own desk. You don't need to learn them; you need to know why to ask for them.

Claude Code Has a Configuration System. You Should Use It.

I used Claude Code for a year before I started customizing it. Skills, hooks, MCP servers, project instructions. Here's the mental model and my actual config.

Is That Improvement Real? A visual guide to eval statistics

You changed the prompt and the score went up. Should you ship it? An interactive companion to Anthropic's eval guide that walks through variance, standard error, and paired comparisons on a concrete example — with less assumed stats background.

Tree of Thoughts: What happens when you let a model explore before committing

How treating reasoning as search — generating, evaluating, and pruning candidate thoughts — unlocked planning capabilities that CoT and ReAct couldn't reach.

Self-Refine: What happens when you let an LLM critique its own work

How a single model can generate, critique, and revise its own outputs — improving 5–40% across seven tasks without any training.

Replicating Reflexion: What happens when you actually run the code

A reimplementation of the Reflexion framework with modern models reveals that stronger models don't improve reflection — they make it worse.

Reflexion: What happens when an agent can learn from its own mistakes

The Reflexion paper added memory across attempts — letting agents reflect on failures in natural language and improve without retraining.

ReAct: How giving LLMs the ability to think and act changed everything

How the ReAct paper established the think-act-observe loop that powers every modern agent system.

Chain-of-Thought: The prompting trick that unlocked reasoning in language models

How a simple prompting technique unlocked multi-step reasoning in LLMs — and why it only works above a critical scale threshold.