I am an operator who builds with AI agents. Getting a model to produce something is the easy part. I care whether the thing works and how I know. Every sheet below answers both.
Four things I wanted badly enough to build with AI agents. Each one asked for a different set of skills. Each has a full build record: the choices, the dead ends and the checks behind it.
Five design finalists became Pillow Fort by day and Lantern Lane by night, with eight holiday styles. Berry, on a spare phone, takes voice requests and never saves a change without a yes.
Killingly’s board pages, meeting packets and volunteer listings, checked every hour and boiled down to one countdown, one phone call and a short list of shifts. No AI model runs when it checks.
Strips the vocals on my laptop, finds timed lyrics and puts them on the TV like a show flyer. Built in public: the build log grows as it ships.
Four research reports, a pivot on day two and four interface prototypes later: a free tool that makes original patterns inspired by 80s arcade carpet, as tiles, wallpapers and Cricut cut files.
Different jobs, different skills. This is where each one showed up. I direct AI agents through the builds and check the work myself, so read it as what each job demanded and what I had to get right, not as a claim that I typed every line.
| The skill | BuryBoard & Berry | Town Familiar | Basement Karaoke | Memphis Pattern Builder | TeamLoop | 25 websites |
|---|---|---|---|---|---|---|
| Interface and visual design | Five finalists, eight holiday styles | Rebuilt around four goals | Three singing screens, one pick | Four interface prototypes, then my own layout ideas | Not needed | Shaders, glass, 3D, motion |
| Reading messy sources | Calendar and weather feeds | Town pages, PDFs, OCR | Synced lyrics, matched by length | Museum records, four research reports, each source checked | Not needed | Not needed |
| Audio and voice | Voice requests, read back aloud | Meeting recordings, to the minute | Vocal separation on a laptop | Not needed | Not needed | A synth you patch by hand |
| Privacy and safe boundaries | Nothing saves without a yes | Passphrase, private notes, offline wipe | Local only, nothing uploaded | Draws its own shapes; nothing traced from a real pattern | Bounded runs, no paid fallback | Not needed |
| Hardware and always-on services | Mini PC kiosk, spare Pixel | Scheduled tasks on a home PC | One laptop, one job at a time | Cut files for a home cutting machine | Not needed | Not needed |
| Tests you can rerun | Edge and Firefox checks each release | 163 browser checks, two browsers | 80 interface checks | 12 cut-file tests, data checks | 67 tests at release | Screenshot and frame-rate checks |
| Directing AI agents | Designs in Claude, builds in Codex | A design bake-off, an outside review | Pipeline and interface split | Gemini and Claude research, Codex checks it | The toolkit for exactly this | Parallel builders, one kit |
| Shipped | On the kitchen table | Hosted, installs on a phone | Prototype, log growing | Live, free, no signup | Public on GitHub | Live, linked from this site |
A hatched cell (on a phone, a missing name) means that build did not need it. The column headings link to each record.
A local toolkit for running several AI subscriptions as one small, accountable team from a single conversation. The lead does the judgment. The runner does the plumbing. Nothing runs in the background.
The starting question was practical. How do I use the subscriptions I already pay for without becoming the human API between three chat windows?
TeamLoop helps the lead model decide when another model is worth involving, sends it a focused brief, preserves what happened, and verifies the useful parts before anything changes in the project. Some jobs need a different eye. Some need one local test and no additional model at all. The useful measure is whether the team helped finish the job better, not whether more models showed up in the transcript.
It is a public preview pulled out of my own working setup. It is not a universal model leaderboard, a new chat app, or a promise of token savings. The guardrails bound the runner. They are not a security sandbox, and the README says so in bold.
Before I let a model build, review, or investigate anything on my behalf, I wanted evidence about which one to trust with which job. So I ran the experiment and published all of it.
Synthetic tasks, isolated runs, browser checks, blind scoring, and every failure I could preserve. It is one practitioner’s structured experiment. It is not a scientific benchmark, an employer evaluation, or a universal ranking, and the first screen of the site says that before anything else.
The AI training and advisory company I founded. It helps teams use AI well and helps their technical people work out what comes next. It runs outside my day job, on its own hours, in its own market.
The StarForge site does not describe a workflow. It lets you run one, with sample data, in the page. The fine print under each demo says exactly what is and is not happening.
People in my workshops ask whether using AI is wrecking the planet. They deserve a sourced answer instead of a shrug, so I built an evidence brief with the research behind it.
The StarForge Cook Book, 65 AI recipes, is on the tools shelf →
Each one is a single self-contained HTML file. No accounts, no connections to business systems, nothing saved. They exist so a room full of people can poke the idea instead of nodding at a slide. All three run right here.
Write down one recurring task before choosing any tool. What starts it, who checks it, what information is allowed, and what happens when something is missing.
Twelve fictional records. Nine included, three exceptions. The arithmetic says PASS and the release decision says HOLD. Change the numbers and find out why both are right.
Answer four questions and get a build brief for a small custom tool. It forces the decisions that matter before anyone connects it to anything.
Nobody assigns the games, the free tools, or the web design. I build them because I want them to exist. They live on their own pages.
What is not on this page. Client work and anything from my day job. I do real work for real organizations, and none of their names, numbers, or screens belong on my personal site. If I cannot show it to you openly, I do not count it as proof.