My AI Harness
Last updated September 2026
I have not opened Figma in months. I design and ship product at Allo, across web and mobile, straight from code. Every project still starts with research and a sharp problem, and it still ends on the numbers it moves. What changed is the middle: how the work gets built.
What makes it work is not the model but the small system I built around it: three skills that run together where I work. I call it the trinôme. The repos are private, but how they fit together is not, so here it is. This is the method. The product it makes lives in the case studies.
The problem
Generic AI drifts. It forgets what we settled last week and re-asks what it should already know. It over-verifies what I can judge in a glance and needs re-briefing every session. For a one-off answer, that is fine. For shipping real product, day after day, it is not. I did not want a smarter assistant, but one that inherits my context instead of rebuilding it every time.
The trinôme
So I built it as three skills, one home per job. Together they carry my memory, my standards and my taste.
- cami-brain is my second brain: who I am, the people I work with and projects I work on, and every decision and lesson worth not re-litigating. It never forgets, so I never repeat myself.
- cami-spec is my kickoff and thinking partner, part PM. I bring a two-sentence intent, it works the problem through with me and turns it into a launch-ready brief. On a big feature it becomes a hub that runs a whole fleet of sessions at once.
- cami-design is my design craft made reusable: my taste, my best practices and my patterns, brought to every layout, interaction, copy and engineering pass.
They share one memory, so a decision made once is known everywhere. And before the work reaches me, cami-design reviews it the way I would, catching what I would catch.
A day in the loop
It starts with intent. cami-spec works out the job and opens a fresh session with real autonomy on its slice. That session builds and checks its own work, then hands me the call that is mine: the visual judgment. When it ships, the brain records what was decided, and the next kickoff starts from there. The loop closes on itself.
The hub, or when one PR becomes ten
Most work is one loop. Some features are ten PRs, and that is where the harness stops being a helper and becomes a team. I promote a planning session to a hub. It holds the roadmap, and instead of building anything itself, it spawns spokes: fresh sessions, one per PR, each with its brief, its identity, and the knowledge that it has a hub.
From there they run in parallel. Each spoke builds its slice and reports back when it is ready, but nothing merges without the hub and me. The hub never guesses: it reads what actually merged, reconciles the roadmap, and hands out the next slice. I sit above it all, one person conducting a fleet.
It feels like magic the first time a hub hands off its third session while the first two are still building. It is not magic. It is a small set of rules that lets a dozen sessions work like one, so I stay on the product, not the machinery.
What it changes
The biggest change is not speed, it is that I build what I used to only design. I used to craft in Figma, climb into an engineer's head to hand the idea over, review what came back, send feedback, and go round again. Now I build it myself, front to back, across web and mobile.
The speed comes with it, a fix that took hours takes minutes, a feature that took weeks I ship in days. I never trade control for the pace: I scope the work, judge the craft, approve the review, and press merge.
I did not build this to use generic AI. I built it so AI would work like me. And now I design the thing that builds it too.