Notes fromproduction.
What I learn building and fixing production systems, written down as I go: AI agents, back-office automation, and what it takes to make an AI-built app safe to ship. Plain words, from the engineer who does the work.
Is ChatGPT Pro 500 Worth $500 a Month? A Three-Step Check Before You Upgrade
A three-step check before paying $300 more a month for ChatGPT Pro 500: try the cheaper model, ask what speed buys you, do the hourly math.
ReadA Week With a Personal AI Agent: It Sent the Email Without Asking, and Waited for My Click to Spend Money
A week of errands with Meta's Muse: it emailed support for me unasked, but spending money still took my click. Where to draw that line.
ReadYour AI Plan Got Smaller. Five Steps to Stop Depending on One Model
AI plans now buy less of the best models. Five steps I took to make my context and tools portable, so switching models does not mean starting over.
ReadYour old AI agent conversations are worth more to the next agent than to you
Most people never reopen an AI conversation. Six ways I put old agent transcripts to work, and the one rule to follow before trusting one.
ReadWhat I learned trying to use Jev to judge AI-generated photos
Jev can't see images, so its photo scores are only as good as the model describing each photo. What six describers taught me.
ReadWhat I changed in my AI workflows after trying Claude Opus 5.5
What changed when I audited my coding-agent skills with Opus 5.5: fewer procedural rules, clearer ownership, and stronger verification.
ReadHow I Automated Myself Out of a Job (In a Good Way)
I built an AI webmaster so my client could stop waiting on me. Here's how it works, what broke, and what I'd do again.
ReadWhy So Much AI Writing Sounds the Same, and What I Am Doing About Mine
A model can copy how I sound. It cannot decide what I think, which is why I'm spending more time writing, not less.
ReadAn $80-a-Month Coding Tool Matched What I Was Paying $200 a Month For
An $80/month AI coding tool matched a $200/month subscription on one task, used a fraction of the tokens, and defaulted to better privacy. One gap: no ZDR.
ReadWhy I Ordered a $10,000 Mac Studio for Local AI
A practical look at the rate limits, data constraints, and hardware tradeoffs behind one agency's move toward local AI.
ReadWhy Nextdoor Suspended My AI Agent for Doing Exactly What I Approved
An AI agent got suspended for replying to job applicants too fast - a preview of what happens as agents become the majority of web traffic.
ReadWhy We Checked an AI-Visibility Sales Pitch Instead of Ignoring It
A cold AI-visibility sales pitch got the symptom right and the fix wrong - and pointed straight at bugs we had shipped on our own site.
ReadThe danger isn't that AI turns evil. It's that we're racing.
The people who understand AI best are the most scared of it. The danger is not evil AI, it is the arms race - and three rules I run every agent by.
ReadThe scream test hears only what still points at it
Pruning 325 agent skills by waiting for errors missed the quiet failures; three counts from the logs found them.
ReadThe error message is the most honest bill you will get for your AI tooling
A local model refused a two-letter prompt because my coding tool sent 1,500 skills first. Two ways to see what your AI tooling sends before you pay for it.
ReadIntelligence Is Now Cheap. The Signature Is Not.
AI made first drafts cost a few dollars. What businesses still pay for is the person who checks the draft and signs it. How to find that gap in your own costs.
ReadFable 5.1 and GPT Astra Changed How I Prompt. The Effort Dial Was the Wrong Place to Reach First.
Cranking effort to max burns a week of tokens in a night. Sort the task first: walk-away or sit-at-the-desk, then pick the model.
ReadCheck Whether You Are Allowed to Use Your AI Tools Before You Build on Them
A compliance contract left exactly two legal ways to use AI on client data. Checking which tools qualified turned into a lesson about vendor dependency.
ReadBy Hand or By AI Is the Wrong Question
A designer with twenty years of depth quit AI because the output looked like a template. He was right about the output, wrong about the conclusion.
ReadAI Is Not Erasing Careers. It Is Erasing the Bottom Rung.
Whole careers are not disappearing. The entry-level work that turned graduates into trusted professionals is, and nobody has replaced it.
ReadWhat Claude's Text Watermark Actually Does, and What It Cannot Do
Crawlers cannot read Claude's text watermark, and retyping does not remove it. What it is, who can read it, and why it changes nothing.
ReadBuilt-In AI or an Outside Tool? Start Inside, Almost Always
The AI built into software you already pay for is usually worse. Start with it anyway, because every integration is one more thing that breaks.
ReadThe One-Page AI Policy Your Business Actually Needs
An AI policy written by AI gives you four pages of principles and changes nothing. The version that works fits on one page.
ReadStop Waiting for the Green Light. Build the Alarm Instead.
Nobody is coming to say AI is allowed. What owners actually want is to find out their automation broke before a customer does.
ReadWho Wrote the Task? A Cheap Test for Any Agentic AI Product
Before you sign anything with the word agentic on it, ask who decided the task needed doing. Usually the answer is a person.
ReadA Gateway Abstracts Your Bill, Not Your Integration
Swapping AI models looks like one config line. A gateway abstracts your bill, not your integration, and the difference cost me a day.
ReadWhere to Start With AI: Automate the Job You Would Rather Not Do
Where to begin with AI: pick the task you avoid but still have to do well. What building a voice agent to screen candidates taught me.
ReadFable 5 Is the Real Thing, and You Have Until July 7 to Test It Cheaply
Fable 5 is a genuine step change, not hype. Why early adopters compound their advantage, and how to test it on Claude Max before the July 7 window closes.
ReadThe Agent Spend Playbook: 5 Rules to Set Before You Let AI Agents Run
Autonomous AI agents are cheap when they replace coordination and expensive when they replace thinking. Five rules — with real numbers — that decide which one you get.
ReadHit one of these in your own system?
Bring the app or the process to a free 15-minute call. I will tell you what I would look at first, and whether I am the right person for it.