x runs product demos side by side, and each one carries the research that argues for it: entities on a 0–10 depth ladder, primary sources, a typed graph of how they connect. The order is research, then the demo, then the launch page, because a claim gets on a page only once something a reader can open has earned it.
The 5 agents below are markdown files in this repo. A model reads one as its system prompt before it does any work, and the 13 skills are the craft primitives those files equip, one capability per file. What an agent claims to own is one column. What it puts on a page you can open is the other, and 1 of the 5 put nothing there.
They fill the corpus and hold the citation bar.
Research lead for a product's research site. Owns the product's entities and moves them up the 0–10 depth ladder with primary-source citations, mechanisms, and cross-links. Runs inside a product scope (ACTIVE_PRODUCT=<slug>). Use when adding an entity, refreshing one with new data, or wiring the edge graph.
Writes the entity pages and holds the depth ladder. Nothing is promoted past skeleton without a primary citation, and every URL it leans on is counted in the sources ledger.
They own the substrate and the numbers that grade it.
Build agent for a single product demo. Scope is exactly app/(products)/<slug>/ (including that product's own .claude/). Set ACTIVE_PRODUCT=<slug> when claiming work. Never touches other products or the platform spine.
New to the roster, and nothing on the platform renders from it yet.
Platform-spine build agent — the inverse of a product agent. Builds shared platform code (app/_platform, app/lib, app/api, root config, the research engine, global .claude skills), never a product's private folder. Use for cross-cutting changes that inherit to all products.
Builds the spine every product inherits: the research engine, the product shell, the registry. This page is its work.
They hold the voice bar and the visual bar, and they can block a merge.
Adversarial design reviewer. Runs on a diff of any UI-shipping change and returns block / ship-with-fix / approve against the Peel design system. Use as the pre-commit gate alongside critic-voice. Never invoked by the author to grade their own work.
The same three verdicts on the Peel axis: tokens instead of hard-coded colour, one focal point per viewport, hierarchy that survives a squint.
Adversarial voice reviewer. Runs on a diff of reader-visible text before it ships and returns block / ship-with-fix / approve. Use as the pre-commit gate on any research prose, product copy, roadmap, or changelog change. Never invoked by the author to grade their own work — the point is that it is a different agent.
Reads the diff before it ships and returns block, ship-with-fix, or approve on the voice axis. Never the author of what it grades — that is the whole point of it being a second agent.
They belong to x, not to a product. The same researcher fills Atlas Learn’s corpus and the next product’s; the same two critics read the diff either way. What each product keeps is its own build agent and its own research skill, naming the sources that field trusts and the frontier it has already covered — and its folder is the only place either can write. The do_not_impact_other_product skill states that rule; .githooks/check-product-scope.sh enforces it. The pre-commit hook reads ACTIVE_PRODUCT and rejects a commit that reaches into another product’s files.
No product currently keeps agents of its own.
13 skills, one capability each. A skill is how a lesson stops being one agent’s habit and becomes something the next agent inherits. When two roles keep contradicting each other in review, the org writes the coordination down as a skill instead of relitigating it every run. They are grouped here by who carries them.
No agent's doctrine names these. Some are new and no role has picked them up; some outlived the surface they were written for.