I'm Mark J. Salisbury. Operator first. I build real tools with AI, test them until I trust them, and teach people the parts that hold up. This site is the manual.
Anybody can make a model do a trick. The job starts when the trick has to survive a Tuesday. I work in three moves, in this order, every time.
Start with a real job, not a model. Find the step that eats the afternoon, build the smallest tool that removes it, and hand it to the person who owns the work.
Near-correct is the dangerous kind of wrong. I run the ugly cases, bring in a second model only when the risk earns it, and keep the receipts. No evidence, no credit.
A tool only one person can run is a liability. I teach the people who own the work how it runs, how to check it, and when to stop trusting it.
This is how my open-source toolkit, TeamLoop, thinks about a job. Pick one and press run. Watch who gets called in, who gets left out, and what it takes for a claim to count.
A family dashboard, a civic lookout, a karaoke studio, a pattern studio, an open-source toolkit and a lot of websites. All built on my own time and documented in the open. No logo wall. No mystery metrics.
A spare screen became BuryBoard: shared calendars, weather, lists and a whole wardrobe of holiday styles. A repurposed phone became Berry, the voice-and-picture companion. Here is the build history, the designs we chose between, and the checks behind the cute parts.
It reads Killingly’s board pages, meeting packets and volunteer listings every hour, then tells me where to show up: one countdown, one phone call, a short list of shifts. AI built it. No AI runs it.
A home karaoke studio for the emo and post-hardcore the catalogs skip. Vocals stripped on a laptop, timed lyrics, and a TV screen styled like a show flyer. The build log fills in as it ships.
It started as a research library with a checked source behind every claim. It spiraled into a free pattern studio: hit Generate, fix it one shape at a time, then take home a tile, a wallpaper or a Cricut cut file. Here is how it got there, dead ends included.
One conversation. A deliberately small team. Evidence before agreement. I already paid for Codex, Gemini, and Claude. I wanted them to work as a team without me becoming the human API between three chat windows. So I built the toolkit that does it, then published the tests that shaped it.
Want to watch the team work? The Dream Machine puts every AI session on my machine in an 8-bit arcade room.
Twenty-five single-file sites built in three sittings. Liquid chrome, a playable arcade, a synth you patch by hand, a trail race flown over a real mountain. Design is the part I do for fun.
A book of 65 AI recipes, a pattern studio for 80s arcade carpet, a gradient studio, a shader shelf, a service-area map maker. No signup on any of them.
Two fighting games you can play in this tab, an arcade shooter, an infomercial for a glowing box. Nobody assigned these.
Full equipment list →The range, on one sheet →Everything you can click →
Sessions start with the jobs people already do and the friction they already feel. People leave with a working example, a way to check the output, and a process that still works on Friday.
Four 90-minute sessions for small-business owners with the Northeastern Connecticut Chamber of Commerce. Mixed-skill room. Every session builds one guided artifact and one the participant owns. Open the sessions and handouts →
Fall 2026Guest talk for the Thompson Business Association, built from a survey of their members. Three live demos and one prompt run side by side in ChatGPT, Gemini, and Claude. No vendor pitches.
May 2026Workshops and multi-session programs for organizations run through my company, StarForge Dynamics.
StarForgeI built this site the way I build everything now. I wrote the brief, directed AI agents through the build, and checked the work. No template. No page builder. If you are wondering what AI can do in the hands of someone who knows the job, you are looking at it.