The session card is not a prompt trick
A bounded AI build session works because the substance and failure-capable checks are on disk, not because the instruction sounds strict.

The card is a handle on the work
We use a short session card when a model works on a larger build. It tells the agent what bounded unit it owns, where the unit’s specification and manifest live, which red check must pass, and where to record facts the next session will need. It also says when to stop: if the work is too large to complete without losing context, split it openly rather than improvise a smaller product.
The card alone is not magic. An agent can obey a beautifully stern prompt and still build against the wrong source if the real specification is in a private conversation it cannot read. A test command that always returns green can make an agent confidently “complete” nonsense. The practical system is the combination of accessible context, small units, an executable check capable of failing and a fresh reviewer who reads the result without inheriting the builder’s assumptions.
Our site rebuild keeps its plan and unit ledger in git. That made a useful failure visible: a page could pass link checks while its story was thin, or an illustration could be technically valid while its meaning lived only in the caption. Automated gates kept links and routes honest; human criticism changed the question being asked. Neither substitute alone was enough.
The more general lesson is that an AI role needs a boundary and a handoff. The builder proposes, a gate challenges, a person owns consequential decisions, and the next session starts from written evidence. That makes model choice less fragile. A frontier model can define the difficult unit; a workhorse can execute bounded work; a cheaper model can grind mechanical tasks—but none receives an exemption from the same acceptance evidence.
The full Five Laws paper explains the operating doctrine and its limitations. If you want to install it in a team, AI-Native Delivery Enablement is the service; the card is only one component, not a sticker for an AI slide deck.