Writing

Essays & ideas

$ ls /writing — 8 entries, newest first

• 01

  AI agents get tired too

  14 August 2026

  An AI agent will never be 100% accurate — and neither is any person you've ever worked with. Why perfection is mathematically off the table, why agents have their own version of fatigue, and what to do about it instead of waiting.

• 02

  Summaries that don't lose facts

  1 August 2026

  AI summaries of medical records go wrong mostly by leaving things out, not by making things up. Notes from the research on omissions, and on how insurance companies catch them.

• 03

  Fixing intent drift: a spec + governance layer for AI-assisted development

  6 June 2026

  Three hours into a session, the agent confidently implements the thing you ruled out an hour earlier. I stopped treating that as a prompting problem and built a governance layer instead. Then I ran a real production feature through it.

• 04

  I didn't invent AppMaker. I curated it.

  23 May 2026

  AppMaker borrows every one of its ideas, on purpose. I keep a references file recording what I took from each of five systems, what I rejected and why, and the two failed projects that shaped the design more than any of them.

• 05

  The agent is the new integration layer

  9 May 2026

  The user has always been the real integration layer, copying values between windows to answer one routine question. I counted the tools behind the agent we run in production — 46 of them — and they replace exactly that job.

• 06

  Building a grounded AI agent for a consultancy website

  25 April 2026

  What it took to put a chat agent on our consultancy site without creating a liability: build-time grounding, budgets that fail closed, and rate limits that get stricter when identification breaks.

• 07

  The four kinds of AI agents (and what's actually behind Claude Code)

  11 April 2026

  A taxonomy I keep coming back to: sort agents by what tools they hold, what they may write to, and how long they run unsupervised. Four kinds cover everything I've built, including what sits behind a coding agent.

• 08

  Talk to your AI agent like a caveman

  28 March 2026

  Every polite word you type at an agent competes with your actual instructions for the model's attention. On the research behind terse prompting, and a practice called caveman mode that I ended up adopting.