Scope as strategy

Build the foundation before the features that need it.

The MVP did one thing: make the JD the single source of truth. Information consistency had the highest impact and the least competition — and everything we deferred would run on that data.

IA as permission model

One hierarchy made permissions listable.

Company → Job → Candidate is the spine for every feature. That turned access into a matrix instead of ad-hoc rules — scope words like "assigned firm" or "team members" compose most of it, and the layer stays open for group-based permissions later.

▏Keywords cover most rules, not all. The few real exceptions were redesigned to fit the system rather than special-cased.

One status, two views

One status engine. Two role-specific views.

Both parties feel they own the workflow; what they see is a controlled view. The same logic runs down to field level, and handoffs become tasks instead of emails.

▏Every new stage has to make sense in both views at once — adding a status means designing it twice.

The clean version came second. The first spec ran two status tracks — one for the talent pool, one for the job pipeline — and keeping them mapped started producing exactly the ad-hoc rules the system was meant to remove. Both boards still sit side by side in Miro: the one that taught us, and the one that shipped.

Later evolution

Once the data foundation held, AI layered on top: four agents across the flow — screening and ranking, data insights, interview preparation, and end-to-end coordination. (Detailed AI workflows are omitted for confidentiality.)