Designing with Jax

This guide shows you how to drive Jax to produce a real design — a network, a route, a layout — by working in small, reviewable steps rather than one giant prompt. You will need a project open with an ontology and at least one Skill loaded, and ideally any base data (roads, parcels, existing assets) already in the Nexma DataStore.

The core idea: treat Jax like a fast junior engineer who never gets tired. Give it a clear objective, let it act, inspect the result on the Globe, then refine. You stay in the loop at every step, and you can roll back at any time.

Steps

1. State the objective, the scope, and what matters. A strong opening prompt has three parts — what you want produced, where it should apply, and any constraint beyond what the Skill already enforces. Do not restate the Skill; Jax already knows it.

Plan cabinet locations for the polygon I just drew. Keep each cabinet within 250 m of a road. Optimize for total trench length.

1. Let Jax act, then inspect before you continue. After Jax finishes, look at the result on the Globe and in the side panels. Resist the urge to fire the next prompt immediately — catching a wrong assumption now is far cheaper than unwinding three steps of work built on it.

1. Refine in small bites. Once the first pass looks right, adjust one thing at a time. Each prompt builds on the last, and every step is a separate point you can return to.

Move cabinet 7 closer to the existing handhole, then re-balance the assignments.

1. Build up the design layer by layer. Sequence the work the way you would by hand — structure first, then connections, then validation. Reference what already exists instead of describing it in prose.

Now route distribution from those cabinets. Snap to the road network already loaded and avoid existing duct.

1. Ask Jax why. Jax is built to explain itself, and asking is the fastest way to both learn the system and catch silent mistakes. After any non-trivial action, interrogate the reasoning.

Why did you put cabinet 4 there? Which constraints were binding, and what drove the cost?

1. Validate against the Skill. When the design feels complete, ask Jax to check it against the constraints and list any violations. Fix them, then re-validate until the list is empty.

Validate the whole design against the constraints in the Skill and list every violation with its location.

Tips

• Hand Jax a skeleton instead of a blank page. If you have prior work or a partial design, import it, point at it, and say "refine, do not replace." This is far faster than describing what you already have.
• Take pushback seriously. When Jax says a request is infeasible — no road access, an over-specified constraint set, math that would not converge — loosen a constraint, narrow the scope, or ask for a counter-proposal.
• Name the objective explicitly when it matters. "Minimize trench length" and "minimize number of cabinets" produce different designs; say which one you want.
• Use branches to explore two approaches in parallel without one overwriting the other.

Where to go next

• Sharpen your requests with the Prompting Jax guide.
• When a step needs optimization, see Run an optimization.
• To explore alternatives safely, read Branch, review and merge.