Field Log // Testing Claude Fable 5
I'm Brandon Smith, and I run a personal setup I call Claude OS: a workspace plus a roster of specialist AI agents, conducted by my voice agent Eugene, that carry a job from a rough idea to a reviewed, tested, committed result. Here is what we built while I put Fable through its paces, with me holding the wheel at every gate.
The setup
The idea is simple. Instead of one general helper, there is a crew of named specialists, each with a single job and its own standards. A spec writer scopes the work. A designer mocks it up. An engineer builds it. A reviewer reads the code for bugs. A QA agent walks the finished thing like a real user would.
The part that makes it trustworthy is the track they run on. Work moves through the same sequence every time, and I approve each gate before it advances. Nothing ships on an agent's own say-so.
The crew
Every agent has a lane and stays in it. When a task lands, it routes to whoever owns that kind of work, and the ones who do not fit stay out of the way. The one doing the routing is Eugene.
Eugene
My voice agent, and the one who runs the crew. I talk, and he routes each job to the specialist who owns it, then keeps the long‑running ones in sync while they work. Sixteen specialists, and Eugene makes them feel like one hire. Real credit where it is due: the orchestration is his.
The build log
Four things, from a quick wiring job to a full app rebuild. Each one went through the whole track.
Added on‑demand image and vector generation that routes to the right model for the job. Connected the code and database stack (read‑only, least‑privilege). Added office‑document‑to‑markdown conversion and a live status readout. Each one was vetted against what already existed, so only the genuinely additive tools stayed. The rest were named and skipped, on purpose.
Eugene runs the whole crew, so his recall has to be sharp. We built him a re‑ranker so he surfaces the right note and pushes the near‑misses out of the way. The first model was accurate but slow, about 10 seconds a lookup. We swapped it for a lighter one at roughly 0.4 seconds, running fully on‑device. Code review caught a real bug before it shipped. QA reproduced the fix under pressure. Then it was committed. The orchestrator got faster at the one thing he does most.
Four inspectors reviewed the setup in parallel: its structure, its memory, its wiring, and its notes. The pass scaffolded missing project spaces, set a cadence so the memory stops going stale, retired eight redundant tools, and reclaimed 134 MB of dead weight. Housekeeping, but the kind that keeps the whole thing sharp.
The flagship. A study app I use to prep for a certification felt clunky. The design agent rendered every screen and named the real problem: no mobile‑first hierarchy, and headers physically colliding on my actual iPhone.
Then it caught something the eye would miss. The correct answer was the longest option in 96% of questions, so the app was quietly training me to pick the longest answer instead of the right one. We rewrote the question bank. That is down to 28% now, where pure chance is 25%.
The interface was rebuilt mobile‑first with proper handling for the iPhone's notch. Code review caught a timer that could freeze the screen mid‑test, and it was fixed. QA walked every flow and proved it. The last gate is the real phone, because headless testing cannot see a physical notch. Being honest about that is the point, not a footnote.